Afretec Proceedings Book

Proceedings Book

Afretec Network Principal Investigators Meeting June 2-4, 2025 ∙ Kigali, Rwanda

Table of Contents

*

An AI-Driven Environmental Monitoring Platform for Low-Resource Settings* Edwin Mugume (Carnegie Mellon Africa)

1

An Investigation of a Monitoring and Predicting Algorithm for Climate Change Related Diseases in African Urban Cities* Mokolade Johnson (University of Lagos)

4

Application of AI Techniques for Extracting Carbon from Landfill Waste for Renewable Energy*1 Abdulganiyu Adelopo (University of Lagos)

7

Birds’ Detector and Repellent System for Smart Farming Kiruki Mutugi (University of Nairobi)

10

Continental Digitized African Sensing Platform Mohamed Swillam (American University in Cairo)

15

Culturally Sensitive Social Robotics for Africa Pravesh Ranchod (University of the Witwatersrand)

19

Deep Learning Driven Multiplexed Prospectivity Modeling of Rare Earths in Radiothermic Carbonatites Hudson Kalambuka Angeyo (University of Nairobi)

24

Digital Malaria Control for the Developing World Carine Pierrette Mukamakuza (Carnegie Mellon Africa)

29

Empowering African Communities: Culturally Relevant Cybersecurity Education through Comic Books and AI Models for Children's Online Protection* Jema David Ndibwile (Carnegie Mellon Africa)

34

Enhanced Cardiovascular Diseases Discovery in Medically Underserved Communities via AI-Assisted Stethoscopy Khalil Elkhodary (American University in Cairo)

38

Evaluating Digital Transformation and Maturity in Youth-Led Micro, Small, and Medium Enterprises across Sub-Saharan Africa: A Comparative Study in the Health, Energy, Environment and Sustainability Sectors in Nigeria, Kenya, and South Africa* Stephen Odock (University of Nairobi)

43

Finia: Financial Inclusion via Novel Intelligence and Alternative Data Chimwemwe Chipeta (University of the Witwatersrand)

46

Seed grants

*

Fully Integrated Label-Free Virus/Parasite Detection System from Blood-Serum Samples based on Electrical Features from Microscale Electronic Sensors and an AI Software Ismail Yehea (American University in Cairo)

51

Improving Digital Education and Learning Innovation in the South African and Rwandan Teacher Education Systems: Towards Bridging a Digital Knowledge Divide in Africa Juliet Perumal (University of the Witwatersrand)

55

Leveraging Additive Manufacturing to Improve Access to Quality Prosthetic and Orthotic Services in Developing Countries Moise Busogi (Carnegie Mellon Africa)

60

Leveraging IoT and Edge-AI for Detecting and Managing Schistosomiasis*2 Hamada Rizk (American University in Cairo)

65

Low-Cost, Accessible Biotechnologies for African Hair and Dark Skin Colours Assane Gueye (Carnegie Mellon Africa)

69

Optimizing a Mobile Application for Enhancing Parents’ Reporting and Prediction of Adverse Effects following Maternal and Child Immunization in Rwanda* Aimable Musafiri (University of Rwanda)

74

Real-Time Noise Level Web Mapping through Crowdsourcing: Toward Creating Sustainable Urban Environments* David Siriba (University of Nairobi)

77

Revolutionizing Oxygen Supply in Healthcare: A Sustainable Approach through Water-Based Oxygen Harvesting Davies Segera (University of Nairobi)

80

Strengthening Water Access and Quality in Selected African Countries Hakeem Amuda (University of Lagos)

84

Supporting Inclusive Air Quality Management in Africa with a Novel System of Digital Tools Telesphore Kabera (University of Rwanda)

89

Towards IoT-Enabled Privacy-Preserving Large-Scale Healthcare Analytics in Africa: A Use Case on Monitoring a Cardiovascular Disease Tamer ElBatt (American University in Cairo)

93

Wakanda AI: Engineering Smart Solutions for Mother Tongues by Advancing Language Technologies in the African Context Bhiksha Raj (Carnegie Mellon University)

98

Seed grants

An AI-Driven Environmental Monitoring Platform for Low-Resource Settings Edwin Mugume1, Andrew Katumba2, Pierre Bakunzibake3; 1Carnegie Mellon University Africa, 2Makerere University, 3University of Rwanda

1. Project Objectives The world is rapidly urbanizing, with over half trhe world’s poplation already living in urban areas. This trend will continue, with most of this urbanization happening in SubSaharan Africa (SSA). This rapid urbanization in SSA has resulted in unplanned and informal settlements that have replaced green spaces and led to poor urban livability.

Climate change disproportionally affects the poor who lack the means to guard against it yet do the least to cause it. Climate change has led to poor thermal comfort, increasing urban heat islands, flooding and other weather effects, reduced agricultural yields, etc.

Most countries in SSA lack sufficient environmental and weather monitoring stations, mainly due to their high cost, and this has hindered the accurate characterization and tracking of the effects of climate change in this region. The same is true for air quality monitoring, where the cost of air quality reference measurement nodes is prohibitive.

It is therefore necessary to build low-cost nodes that can be deployed at scale to maintain large open datasets and aid the tracking of air quality and weather effects.

In this project, we will design, develop, and test automatic weather stations and air quality sensor nodes using low-cost sensors and accessories. The project has the following objectives: 1. Design, develop, and calibrate a low-cost automatic weather station that measures ambient temperature, relative humidity, wind speed and direction, and rainfall.

2. Design, develop and calibrate a low-cost air quality/particulate matter measurement node.

3. Design, develop and test low-cost custom acoustics-based sensor nodes for tp measure rainfall, wind speed, and wind direction.

4. Build and test a platform for the data storage and access of weather and air quality data.

2. Recent Progress The design of the different nodes has already been completed. Custom-designed printed circuit boards (PCBs) have been designed for the nodes and are currently being printed. In addition, we have completed the design of the node enclosures and radiation shields, and samples have already been 3D-printed. All nodes are designed to run sustainably on solar power.

1

The automatic weather station acts as an evaluation node for various low-cost temperature and humidty sensors i.e. SHT31-D, HTU21D, BME280, and AHT20, all purchased from Adafruit [1]. The most acurate and consistent sensor will then be chosen for the final production weather nodes. We have also considered two different rainfall and wind (speed and direction) sensors i.e. from Davis Instruments (part of the Vantage Pro2) [2] and the SparkFun Weather Meter Kit by SparkFun Electronics [3].

The air quality measurement node uses the PMSA5003 air quality sensor from Adafruit [1]. This sensor has been favoured by other researchers who have developed similar low-cost nodes. In addition, the SHT31-D is also included in this node based on initial lab-based evaluation which shows that the SHT31-D is superior. The 3D-enclosure and custom PCB designed for the air quality node are shown in Figure 1.

Figure 1: The 3D-enclosure and custom PCB designed for the air quality node

The project is also developing rainfall and wind speed sensors using acoustic-based and machine learning techniques. The rainfall sensor uses a MEMS microphone and Raspberry Pi Zero to record and process rain audio, which is then used to estimate rain rate and intensity. A wakeup mechanism is implemented using a raindrop sensor and MOSFET on an ESP32 MCU to ensure that audio is recorded only when rainfall is detected. The acoustic-based wind sensor uses ultrasonic transducers to measure wind speed based on the time-of-flight (ToF) principle. The sensor emits sound pulses and measures the time they take to travel a known distance. This ToF is influenced by wind speed and can be used to calculate it. To improve accuracy and account for environmental variations, machine learning regression models will be used for sensor calibration, enhancing the sensor’s robustness and reliability in real-world conditions.

These sensors are being considered as low-cost options to those from Davis Instruments and SparkFun Electronics.

2

Figure 2: Block diagram of the custom acoustic-based rainfall sensor.

3. Next Targets In the next steps, the development of the weather and air quality sensor nodes will be completed by installing all components on the respective custom PCBs and fixing them within their respective enclosures. In addition, the alternative acoustic-based rainfall and wind sensors will also be installed on their PCBs. All the four prototypes will then be tested in the lab to ensure that they perform as expected. Finally, the weather and air quality sensor nodes will be deployed and collocated with reference sensor nodes in the field to gather weather and air quality data respectively. Two automatic weather stations and two air quality sensor nodes will be deployed in Kigali, Rwanda, where collocation will be achieved through a partnership with the Rwanda Environmental Management Authory (REMA) and Rwanda Meteorological Agency (Meteo Rwanda).

In Uganda, three automatic weather stations and three air quality sensor nodes will be deployed, and collocation will be achieved in partnership with Uganda National Meteorological Authority (UNMA). The reference nodes will provide the accurate data needed to calibrate the sensor nodes using machine learning techniques. Another potential source of data is the Trans-African Hydro-Meteorological Observatory (TAHMO), which provides free access to its weather data for research purposes. The acoustic-based rainfall and wind sensors will be deployed in Kampala, Uganda, and calibration data will be obtained through the same partners (UNMA and TAHMO).

4. Peer-Reviewed Publications Resulting from this Project No publications to report for the moment.

References [1] [2]

[3]

Adafruit. https://www.adafruit.com/.

Davis Instruments, “Anemometer for Vantage Pro2 & EnviroMonitor”, https://www.davisinstruments.com/products/anemometer-for-vantage-pro2vantage-pro (accessed: April 28, 2025).

SparkFun Electronics, “Weather Meter Kit”, https://www.sparkfun.com. https://www.sparkfun.com/weather-meter-kit.html (accessed: April 28, 2025).

3

An Investigation of a Monitoring and Predicting Algorithm for Climate Change Related Diseases in African Urban Cities Immaculata Nwokoro1, David Rene Seagra2. Mokọ́ládé Johnson1, Abdulganiyu Adelopo1, Tinuola Odugbemi1, Oluyemi Akindeju1, Ife Albert1, Mina Ogbanga3 1

University of Lagos, Nigeria, 2 University of Nairobi, Kenya, 3 University of Port Harcourt, Nigeria

1. Project Objectives Climate change presents a growing risk to public health, particularly in urban African settings where rapid urbanization, poor waste management, and infrastructural deficits intersect with worsening environmental conditions. The increasing frequency of climate-sensitive health issues such as malaria, cholera, and heat-related morbidities demands innovative public health surveillance tools. This study seeks to address these emerging challenges by developing a community-centered, artificial intelligence-supported Interactive Voice Response (IVR) system that functions as an early warning and symptom surveillance mechanism. The main objective is to identify the information needs, user perceptions, and systemic gaps necessary to co-create and deploy a scalable digital health intervention adapted for resource-constrained and vulnerable communities. It aims to develop a comprehensive data-gathering system using Interactive Voice Response (IVR) connected to the Internet of Things (IoT), ensuring data inclusivity and confidentiality.

The project explores how digital health tools, particularly voice-based technologies accessible through basic mobile phones and multimedia, can bridge the gap in health surveillance and disease prediction. With a focus on Ìlàjẹ-Bàrígà, a climate-vulnerable urban slum community in Lagos, Nigeria, the study aims to: assess the feasibility and willingness of local users to engage with an AI-IVR system; identify the enablers and barriers to adoption; and generate community-informed content that can guide the system’s design and implementation. Ultimately, the project aspires to create a scalable model for early detection and reporting of climate-induced health symptoms, supporting both proactive public health responses and long-term resilience strategies.

2. Recent Progress To ground the proposed framework in local realities, the project adopted a mixed-method approach incorporating semi-structured interviews, a structured survey, and stakeholder engagement sessions. A total of 139 stakeholders were surveyed, comprising patients, caregivers, health workers, and informal community health actors. The study population was predominantly female (77.2%), below the age of 35 (58.5%), and mainly individuals with at least a secondary school education (95.1%). Most health worker respondents were employed for less than three years and were active residents of the Ìlàjẹ-Bàrígà area. The prevalence of younger and middle-aged participants with a minimum of secondary education highlights a potentially technology-literate demographic capable of adapting to IVR solutions.

The use of the Unified Theory of Acceptance and Use of Technology (UTAUT) for respondents’ perception and readiness to use IVR indicated participants’ critical needs for an IVR system— specifically for saving time, improving access to health information, and enhancing the ease of symptom reporting—which correlated positively with high communalities for awareness (0.750) and benefits of IVR (0.580). This highlights the importance of designing a system that emphasizes these aspects. Crucial enablers identified included social influence and peer recommendations, 4

community ownership and trust, and the development of culturally appropriate content that removes linguistic barriers, with a rating of 0.94 for such inclusion. Respondents emphasized the importance of using local languages and features which would make the system more intuitive and acceptable.

The development of the Interactive Voice Response (IVR) system for this project successfully integrated several core technologies tailored to the needs of a low-resource and linguistically diverse community. The initial phase focused on implementing voice recognition technology, which enabled the system to accurately capture and process spoken responses from users during interaction. Natural Language Processing (NLP) algorithms were deployed to enhance comprehension and interpretation of the collected audio data. A cloud computing platform was used to support the high volume of data generated during user interactions. This provided scalable storage and reliable processing power necessary for real-time data management, analysis, and model updates. The cloud infrastructure also enhanced the system’s resilience and allowed for remote monitoring and maintenance.

Furthermore, full telephony integration was achieved, enabling the IVR system to handle incoming messages, route them appropriately, and provide automated responses as required. This integration ensured seamless interaction between users and the system across multiple phone networks, regardless of device sophistication.

Finally, a user-friendly interface was designed, prioritizing ease of use for individuals with varying levels of technological literacy. The menu navigation was simplified and structured logically, with audio prompts in simple, user-friendly English. This ensured that users with minimal or no experience with digital systems could still interact effectively with the IVR system. The overall design approach emphasized inclusivity, cultural relevance, and trust-building through familiar and intuitive interaction features.

In addition to the deployment of the IVR system, a comprehensive dashboard analytics platform was developed to support real-time monitoring, feedback dissemination, and data-driven decisionmaking. The dashboard offered layered statistical analyses tailored to the needs of different stakeholder groups—ranging from health officials and policymakers to community leaders and system administrators. These analytics visualized patterns in symptom reporting, user engagement, and message volumes, helping to interpret the broader implications of collected data within the context of climate-sensitive disease surveillance. By translating complex datasets into actionable insights, the dashboard empowered stakeholders to track the progression of reported illnesses, assess geographic disease hotspots, and evaluate the timeliness and effectiveness of public health responses.

Crucially, the analytics platform was designed to facilitate adaptive responses by allowing system administrators to customize IVR messaging and interventions based on real-time trends and user feedback. For example, if the system detected a spike in heat-related symptoms in a particular locality, automated heat safety tips or health alerts could be disseminated through follow-up IVR messages in the local language. This feedback loop made the system dynamic and responsive to emerging needs, strengthening its relevance and effectiveness. The dashboard’s user interface was intuitive and accessible, allowing non-technical users to interact with data meaningfully. Overall, the integration of dashboard analytics enhanced the responsiveness, inclusivity, and accountability of the IVR system, making it a truly user-adaptive digital health platform for vulnerable urban populations.

From this pilot project, a total of 367 data entries had been processed and data interpreted for researchers’ evaluation. Figure 1 below provided pictorial representation of the IVR data system.

5

IVR Input

Cloud Data Collection Graphics

Data Visualization

Figure 1. Pictorial representation of the IVR data system

3. Next Targets Building on the foundational insights gathered from the needs assessment and stakeholder engagement, the next phase of this project will focus on the prototyping and operationalization of the AI-IVR system. This involves expanding the culturally and linguistically appropriate message scripts with local community input to develop multi-dialectal response recognition, and integrating IoT-enabled climate data into the predictive model. To optimize system performance and scalability, the project will expand its data collection scope and enhance analytics capabilities through cloud-based infrastructure and strengthened telephony integration.

With further funding sources, efforts will be directed toward improving end-to-end functionality— from data capture to dashboard visualization—while training workshops and digital literacy programs will be rolled out for both users and frontline health actors. This human-centered approach will address the digital skills gap and increase system uptake. Furthermore, policy dialogues with government agencies, health ministries, and regulatory bodies will be initiated to ensure alignment with national health strategies and secure the necessary institutional support for sustained deployment. Partnerships with telecom providers, local tech hubs, and community-based organizations will also be explored to reduce operational costs and bolster backend reliability.

Ultimately, the project is poised to offer a replicable, inclusive model for symptom-based disease forecasting in Africa’s rapidly growing cities. With AI and IVR technologies, the proposed solution holds transformative potential—not only in disease surveillance but also in creating a proactive, people-centered public health framework that anticipates the health impacts of climate change and empowers communities to respond early and effectively.

4. Peer-Reviewed Publications No publications to report yet.

6

Application of AI Techniques for Extracting Carbon from Landfill Waste for Renewable Energy Adelopo AbdulGaniyu¹, Siphila Wanjiku Mumenya², Babatunde Sawyerr³, Sarath Tennakoon⁴ ¹Waste and Water Quality Unit, Works and Physical Planning Department, University of Lagos, Nigeria ²Department of Civil & Construction Engineering, University of Nairobi, Kenya ³Department of Computer Sciences, University of Lagos, Nigeria ⁴Carnegie Mellon University Africa, Kigali, Rwanda

1. Introduction and Objective African cities are grappling with increasing waste generation due to rapid urbanization, leading to overburdened landfills and growing environmental concerns. These landfills, often unmanaged or poorly regulated, consume valuable land, emit greenhouse gases, and lack structured data systems to support material recovery and resource reuse. Despite global advancements in digital waste management, African municipalities remain constrained by limited capacity for technology adoption, especially in leveraging data for waste to wealth.

This project explores the application of Artificial Intelligence (AI) to address these gaps by using Convolutional Neural Networks (CNNs) to predict the carbon recovery potential of landfill waste for renewable energy storage, with a specific focus on carbon conversion for use in supercapacitors. AI models have demonstrated global success in improving waste sorting accuracy and reducing operational costs; however, their integration into African landfill systems remains underexplored. Recent studies show that AI-enhanced material classification can increase recycling efficiency by over 35%, especially when supported by sensor-enabled data systems (Kalantar-Zadeh et al., 2024). Furthermore, carbonized waste materials, particularly biochar from municipal waste, are increasingly being researched as viable electrode materials for renewable energy devices (Tetteh et al., 2023).

The pilot study targets municipal landfills in Lagos (Nigeria) and Nairobi (Kenya), aiming to map the composition of waste, evaluate operational practices, and identify environmental factors that influence carbon yield. It will also assess stakeholder readiness and capacity for digital transitions, particularly among informal waste workers and municipal authorities.

The objectives of the project are to: 1. Identify the data types and parameters essential for training AI models to predict carbon yield from landfill waste.

2. Assess stakeholder needs and existing knowledge gaps in AI-based landfill monitoring.

3. Perform systematic field sampling and laboratory analysis to extract and characterize carbon from landfill materials.

4. Develop, train, and validate machine learning models (e.g., CNNs) to predict carbon extraction potential based on waste profiles.

5. Establish a standardized and scalable landfill dataset that can serve as a benchmark for future AI-driven landfill management systems.

7

Figure 1: Diagrammatic presentation of the project concept

2. Recent Progress and Key Findings This research project has successfully demonstrated the feasibility of deploying Artificial Intelligence (AI) in landfill carbon recovery efforts in African urban contexts. Strategic progress was made across both technical and stakeholder engagement fronts in the two selected study sites: Olusosun Landfill in Lagos, Nigeria, and Dandora Landfill in Nairobi, Kenya.

Field access and data collection began with the formal securing of sampling sites at both landfills.

Collaborative arrangements were established with local waste management authorities, allowing for the integration of informal and formal waste actors into the study. These partnerships facilitated the smooth execution of field activities and strengthened stakeholder buy-in.

A total of 20 key informant interviews and 879 surveys were conducted across the two sites to evaluate stakeholder readiness for digital transformation. These efforts helped map the current capacities, perceived challenges, and willingness of landfill workers and managers to adopt AIbased systems. As part of capacity building, 100 landfill workers were trained through two structured workshops held in Lagos and Nairobi, focusing on basic digital literacy, data handling, and environmental monitoring techniques.

The technical component of the research introduced a dual-data sampling procedure that integrates material sampling and optical recognition. A total of 200 representative waste samples were collected and categorized, while more than 3,500 images and 2,500 numerical entries were compiled for model training. These datasets were used to train three (3) Convolutional Neural Networks (CNNs), resulting in 95% data training and validation rates which underscore the feasibility of AI-assisted carbon estimation strategies.

Initial findings were presented at the AI Action Summit in France, providing a global platform for visibility and peer feedback. One of the standout findings was the successful conversion of landfill waste images into machine-readable formats, with a 96% accuracy rate for classification of waste into the relevant categories An analysis of the workforce demographics revealed that 64.8% of landfill workers were under the age of 36, suggesting a youth-dominated workforce with a high potential for technology adoption. Additionally, 62% of respondents had over four years of operational experience, offering valuable insights for the development of user-centric digital tools. Notably, 69.7% of the 8

workforce occupied key operational roles such as waste pickers and truck operators, underscoring the importance of targeting digital tools at the core of landfill logistics.

Material characterization also revealed that three waste categories—polythene, inert, and organic waste—constitute approximately 70% of the total municipal solid waste at the landfills. These categories were found to possess relatively higher carbon content and thus represent a very promising feedstock for conversion into materials suitable for supercapacitor development.

3. Next Targets Building on the momentum from the pilot phase, the project will now focus on optimizing AI models for greater prediction precision by further validating them against laboratory-analyzed carbon content data. This validation process is critical for improving the robustness of the models and expanding their applicability across different landfill scenarios.

In the coming phase, a short course will be delivered to landfill workers, focusing on advanced digital data collection techniques. This training will aim to deepen understanding of the link between waste categorization, environmental metrics, and AI integration among the stakeholders To enhance real-time monitoring, Internet of Things (IoT) devices will be deployed across selected landfill zones. These devices will collect live data on temperature, pH and waste images enriching the AI model inputs and enabling dynamic predictions.

Laboratory trials will also begin on converting landfill-derived materials into carbon, which will then be tested for applicability in supercapacitor technologies. These trials will help establish proof-of-concept for the waste-to-energy storage pathway being proposed.

Through bigger funding opportunity, the project will develop the ALANTA Tool—a comprehensive AI-IoT dashboard designed for landfill data management, carbon potential prediction, and decision support. ALANTA will serve as a practical interface for city authorities, researchers, and waste operators, supporting smarter, circular, and data-driven landfill management practices in African cities.

It is expected that the sampling of a greater number of landfills across Africa will be carried out to strengthen the ALANTA Tool’s coverage and applicability to diverse waste streams, and to develop a scalable business model for African landfill mining while structured policy engagement will be undertaken to support enabling frameworks for AI integration into landfill circular economy waste practices.

4. Publications Peer-Reviewed Publications Resulting from this Project No publications to report for the moment.

References •

•

Kalantar-Zadeh, K., Tiemann, G., & Morris, D. (2024). Smart waste systems: AI and sensor-enabled waste classification in urban environments. Journal of Sustainable Smart Cities, 12(1), 45–61. https://doi.org/10.1016/j.jssc.2024.01.004 Tetteh, E. K., Kader, S. H., & Muzenda, E. (2023). Carbonized municipal solid waste as an energy storage material: A path to sustainable urban energy. Waste Management & Research, 41(8), 938–950. https://doi.org/10.1177/0734242X231161052

9

Birds’ Detector and Repellent System for Large-Scale Smart Farming Emmanuel Ndashimye, PI (CMU-Africa); Evariste Twahirwa, Co-PI (University of Rwanda); Mutugi Kiruki Co-PI (University of Nairobi); Christine Niyizamwiyitira, Co-PI (CMU-Africa); Peace Bamurigire, Co-PI (University of Rwanda); Moise Busogi, Co-PI (CMU-Africa) ●

Project Objectives:

The overall objective of this project is to develop a low-cost sensor device that can effectively detect the presence of birds and deploy appropriate deterrent measures to protect crops in agricultural environments. This initiative aims to address the significant challenge of bird damage to agriculture, which causes economic losses, disrupts social dynamics, and even affects educational pursuits as children are often kept out of school to guard crops. When it comes to the context of large agricultural farms, current bird detection and repellent methods are often ineffective, highlighting the need for innovative solutions.

To achieve this overall objective, the specific project objectives are four-fold: 1. Vision-Based Bird Detection: We will implement computer vision algorithms for bird detection based on visual cues.

2. Building an Acoustic Sensor-Based Bird Detection: We will develop a cost-effective bird detection system using acoustic sensors by leveraging robust signal processing algorithms and machine learning.

3. Repellent System Development: We will design and implement an environmentally friendly bird repellent system, exploring various deterrent measures. We will assess the effectiveness of repellent strategies against different bird species, optimising for cost-effectiveness and minimal impact on non-target species.

4. Integration and Testing: We will integrate acoustic and vision-based systems with the chosen repellent mechanism into a functional unit. We will also develop a user-friendly interface and conduct extensive real-world testing to assess and validate the performance, accuracy, and efficiency of the system. ●

Project Progress:

In this section, the recent progress is presented based on the main areas being addressed. The research areas are: Data Acquisition, Bird Detection, Detection System Deployment, and the Repellent System.

Figure 1: The detection algorithm’s flowchart

10

●

Data Acquisition:

Research focuses on determining the minimum image quality and camera settings required for effective bird information retrieval. It also includes developing a camera planning algorithm to optimise field coverage given the number of cameras and their specifications.

To determine the requirements for effective retrieval of bird information, a survey is being conducted to gather statistical data on bird characteristics, such as body sizes, shapes, and flying speeds, by utilising existing published ecological studies. The necessary amount of information, in terms of minimum pixels (as the main parameter) and frames per second is to be decided on, based on the prediction accuracy of the algorithms we will develop against existing bird detection datasets.

For camera coverage optimisation, a camera planning program with a GUI has been developed, incorporating a mean-shift algorithm that decides the right positioning of 360° camera rigs over the map of the agricultural field, given their depth of field. Upon the user’s request, the program can recommend a well-suited camera depth of field from an input list for better coverage.

(a)

(b)

(c)

Figure 2: From left to right: (a) Visualising the Horizontal Field of View of a camera as a slice of a circle. (b) Putting together 5 cameras of the same HFoV and Depth of Field, d, to make a full 360° HFoV. (c) An Illustration of a 360° camera rig.

(a)

(b)

Figure 3: Example outputs of the mean-shift algorithm, showing the recommended positions of camera rigs, given their depths of field, defined as the radii of the shown circles. In (a), the radii were fixed: one having 50 m, two having 75 m, and other two having 100 m; while in (b) the radii were initialized, as in (a) and the algorithm was allowed to improve them, increasing the coverage from 83% to 98%. (Image source: Google Maps, 2025)

11

●

Bird Detection:

This area investigates the most efficient change detection techniques for real-time bird movement detection and the optimal image object detection algorithm for identifying birds.

The project has been evaluating state-of-the-art deep learning models, with initial assessments indicating that the YOLO family offers lightweight options with comparable performance. Testing of YOLOv8n with confidence threshold and Intersection over Union hyperparameters revealed trade-offs in Precision and Recall, as some of the birds wouldn’t be detected due to their small sizes, while other background structures would be confused with birds. With subsequent efforts involving constraints based on considering birds intersecting the sky, and the adoption of YOLOv8s-world with text prompts improved the Precision from 0.76 to 0.92 and the Recall from 0.55 to 0.93. However, the challenge of detecting small birds due to image downscaling in YOLO models remains an area of focus.

Figure 4: An example of detecting birds using change detection (optical flow) and the YOLO-world object detection model. ●

Detection System Deployment:

Objectives in this area include identifying the minimum computational requirements for effective realtime deployment of the bird detection algorithms and exploring model compression methods to ensure computational efficiency on embedded systems with minimal impact on model accuracy. The 26MB YOLOv8s-world model performs inference in under half a second on CPUs of devices like the Raspberry Pi and NVIDIA Jetson Orin NX, making compression currently unnecessary. However, ongoing research is exploring optimal compression methods for deployment on lower-powered devices. ●

The Repellent System:

This area focuses on developing the repellent system that communicates with the bird detection module.

Once birds are detected, this system will receive the bird’s location coordinates, in terms of vertical and horizontal angles. An integrated dual servo motor setup will then adjust the orientation of the repellent module to target the birds' location, enabling precise aiming of the repellent actuators. The repellant system currently comprises 3 main components: actuator part, power supply and the communication part. The three components are briefly described in this section. a) Repellent Actuator Part Currently, the repellant actuator consists of a microcontroller and audio repellant actuators. This is depicted in Figure 6, showing the PCB and the prototype accomplished.

12

(b)

(a)

(c)

Figure 6: (a) Bird repellent actuator PCB Layout; (b) PCB with all components soldered; (c) The prototype of the actuator The prototype integrates essential features for an automated bird repellent system, combining audio playback, physical actuation, visual deterrence, and manual control capabilities. At its core is the ESP32-WROOM-32 microcontroller. Using acoustic deterrence as a placeholder, a DFPlayer Mini MP3 module is included. Physical actuation is supported via a dedicated "SERVO OUT" header that outputs PWM signals from the ESP32. System status is visually indicated by four LEDs, each paired with a currentlimiting resistor. The entire board and its peripherals are powered through a 2-pin power input terminal. b) Power Supply

Figure 7: A block diagram for the bird repellent system power supply.

The power supply system has been designed to be a solar-powered, microcontroller-based architecture, as it is shown in Figure 7. It comprises a solar panel that charges a lead-acid battery through a charge

13

controller (XH-M604), ensuring regulated energy storage. The battery supplies power to the system, with current and voltage monitored by a power sensor (INA219). This energy is then regulated to 5V using an LM2596 voltage regulator, which distributes power to both the communication gateway and the actuator system. The ESP32 microcontroller serves as the central unit for power monitoring and system control. It receives power from the regulator and gathers environmental data via a DHT22 temperature and humidity sensor. Additionally, it interfaces with an OLED display to provide real-time visual feedback. c) Communication Link The communication link serves as the bridge between the detection system and the repellent system, enabling real-time activation or deactivation of repellent nodes based on detection data. It utilises XBee modules operating on the Zigbee wireless protocol, with a UART interface at 9600 baud for efficient data exchange. Designed with minimal processing requirements, the system ensures fast transmission of angle readings from the detection part to the repellent system. ●

Next Targets:

Next to the progress on the key research areas presented above, this section briefly presents the remaining gaps to be addressed and the methods to be used.

●

●

Data Acquisition: Following an on-field survey focusing on specific agricultural fields in the Eastern Province of Rwanda, the collected data—including terrain characteristics, bird species and movement behavior, presence of water bodies or swamps, optical barriers, and field sizes and shapes—is to be used to update and refine the current camera planning methodology.

●

Bird Detection: A labelled dataset will be created for both training and testing purposes. It will consist of images captured from a selected agricultural field used as a case study and will serve to train and assess custom bird detection models, plus model compression approaches. Unlike existing pre-trained models, these will be specifically fine-tuned to detect the targeted bird species.

●

Repellent Actuator: Currently, research is underway to determine the most suitable repellent actuator to be integrated with the system. Options under evaluation include a laser diode, ultrasonic speaker, magnetic actuator, or other mechanisms capable of safely and effectively deterring birds.

Each option is being assessed based on several factors: effectiveness in repelling different bird species, safety for humans and animals, and the coverage.

Peer-Reviewed Publications Resulting from this Project:

No publications to report for the moment.

14

Continental Digitized African Pathogens and Climate Pollutants Sensing Platform Mohamed Swillam (The American University in Cairo), Albert A. Presto (Carnegie Mellon University), Rose Alani (University of Lagos)

1- Abstract The Continental Digitized African Sensing Platform (CDASP) project represents a transformative initiative aimed at addressing the critical gap in air quality monitoring across Africa through the development, testing, and deployment of low-cost optical sensors. Air pollution remains a leading public health crisis in the Global South, with Africa suffering from a severe lack of infrastructure to monitor pollutants like CO2. This project leverages the complementary expertise of three institutions—AUC (sensor design and nanophotonics), CMU (sensor calibration and deployment), and the University of Lagos (local policy integration and health impacts)—to create a sensor platform that is both innovative and locally sustainable.

Significant progress has been made during the reporting period (June–December 2024). The team designed a high-sensitivity plasmonic grating for CO2 detection, achieving a sensitivity of 12–14 μm in the infrared range. Collaborative efforts with Egyptian research institutions accelerated material innovation, including the integration of zinc oxide layers to enhance selectivity.

Computational simulations using Lumerical FDTD validated the sensor’s theoretical performance, while rapid prototyping enabled iterative testing. Field testing preparations are underway, with pilot deployments planned for Cairo, Kigali, and Lagos in early 2025. Cross-cutting objectives, such as capacity building and stakeholder engagement, have advanced through workshops and the inclusion of underrepresented groups in the research team (35% female representation).

The project’s broader impact includes fostering STEM workforce development in Africa, strengthening university-industry partnerships, and elevating the global visibility of African research. Challenges such as funding delays and technical hurdles in sensor calibration were mitigated through adaptive strategies, including external expert consultations and software training for students. Future work will focus on prototype fabrication, large-scale field testing, and policy advocacy to translate data into actionable air quality management strategies.

15

Explanatory Statement: Portions of this abstract were refined using generative AI tools to enhance clarity and coherence. All technical content, data, and conclusions remain solely the work of the research team and are grounded in empirical results.

2- Project Objectives The CDASP project seeks to revolutionize air quality monitoring in Africa by achieving three primary goals. First, it aims to develop low-cost, high-performance optical sensors capable of detecting key pollutants like CO2. These sensors will be designed for durability and scalability, leveraging AUC’s expertise in nanophotonics and CMU’s experience in real-world deployment.

Second, the project focuses on building local capacity through knowledge exchange, ensuring African researchers can independently maintain and expand the sensor network. Third, it targets policy impact by collaborating with governments and NGOs to translate data into clean-air initiatives.

The initiative is rooted in inclusivity, with 30% of the research team comprising underrepresented groups and partnerships spanning academia, industry, and community stakeholders. By integrating machine learning calibrations and IoT connectivity, the project aspires to create a replicable model for environmental monitoring across the Global South.

3- Recent Progress Sensor Design and Simulation: The team successfully designed a plasmonic CO~2~ sensor using a silicon-based grating structure with metal-dielectric-metal (MDM) layers. Simulations confirmed a sensitivity of 1,200 nm/RIU in the 12–14 μm wavelength range, outperforming conventional electrochemical sensors. A novel zinc oxide coating was introduced to reduce cross-sensitivity to humidity, addressing a key challenge in low-cost sensor reliability.

Collaborative Milestones: •

Held two workshops with Egyptian research partners to refine fabrication techniques.

•

Trained 32 students (60% from disadvantaged backgrounds) in sensor simulation software.

16

•

Submitted one journal paper (under review) and two conference papers to SPIE on grating design innovations.

Field Testing Preparations: •

Secured co-location sites in Lagos and Kigali with reference monitors.

•

Developed a machine learning pipeline for real-time data calibration using Python and TensorFlow.

Challenges and Adaptations: •

Funding delays necessitated reallocating resources to computational work, delaying prototype fabrication by three months.

•

Regulatory hurdles in Rwanda were resolved through partnerships with CMUAfrica’s local network.

4- Next Targets a) Prototype Fabrication: Complete fabrication of 10 sensor units by Q2 2025, incorporating feedback from simulations. b) Field Deployments: Initiate six-month pilot deployments in Cairo, Kigali, and Lagos, with real-time data integration into a public dashboard. c) Policy Engagement: Host stakeholder forums in each city to present findings and advocate for data-driven air quality policies. d) Publications: Publish three peer-reviewed papers on sensor performance, calibration models, and health impact assessments.

17

5- Peer-Reviewed Publications Resulting from this Project •

1 journal paper is currently under review.

•

2 conference papers accepted at SPIE conference (in preparation for final submission).

[1] M. Swillam et al., "Design of High-Sensitive Plasmonic Grating Sensor for CO₂ Detection," Under Review, 2025.

[2] M. Swillam et al., “Compact Plasmonic Sensor for On-Chip Air Quality Monitoring,” Proc. SPIE Photonics West, 2025, piblished.

[3] A. Hamed, M. Swillam, “Infrared Metasurface Sensor for Selective Gas Sensing Applications,” Proc. SPIE Photonics West, 2025, published.

18

Culturally SensitiveSocial Social Robotics for for Africa Culturally Sensitive Robotics Africa David Vernon1 , Pravesh Ranchod2 , and Benjamin Rosman2 1

2

1

Carnegie Mellon University Africa, Kigali, Rwanda University of the Witwatersrand, Johannesburg, South Africa

Project Objectives

Culturally Sensitive Social Robotics for Africa (CSSR4Africa) is a three-year research project to equip social robots with the ability to interact sensitively and politely with people in Africa; see Fig. 1. The objectives of the project are (i) to identify the verbal and non-verbal social and cultural norms of human interaction that are prevalent in countries in Africa, (ii) to encapsulate them in the behavioral patterns of social robots so that they can engage with African people in a manner that is consistent with their expectations of acceptable social interaction, and (iii) to demonstrate these culturally-sensitive social robot behaviors in two use cases: one for giving a tour of a university laboratory, and one for assisting and giving directions to visitors at the reception of a university. In order to ensure that the project objectives can be achieved in the time available, we restrict the scope of the project to the cultures and social practices that are prevalent in Rwanda and South Africa.

2

Recent Progress

2.1

Overview

As we approach the end of the second year, the project is progressing well, assiduously following the work plan and updating it when necessary. This has resulted in the publication of 32 deliverables on the CSSR4Africa website www.cssr4africa.org, four ROS packages and 11 ROS nodes (not including unit tests) on the CSSR4Africa software respository on GitHub site, and publications in the form of three abstracts, five posters, five presentations, and five peerreviewed papers: (Zantou & Vernon, 2023; Akinade et al., 2023; Vernon, 2024; Akinade, Barros, & Vernon, 2025; Akinade, Barros, Danso, et al., 2025); see details in Section 4.

Figure 1: Based on surveys to acquire cultural knowledge about acceptable modes of communication, the CSSR4Africa project is equipping social robots, e.g., Pepper, with the ability to interact sensitively and politely with people in Africa.

1 19

2.2

Work Plan Updates

In the CSSR4Africa work plan, 23 of the 51 tasks are assigned to the University of the Witswatersrand. However, due to extensive delays in the delivery of the Pepper robot to the University of the Witswatersrand, no significant progress had been made on six tasks that were crucial for the integration of a complete system and the ability to execute a robot mission. The six deliverables are as follows.

D5.4.1 Cultural Knowledge Ontology & Culture Knowledge Base.

D5.4.2 Robot Mission Language.

D5.4.3 Robot Mission Interpreter.

D5.5.2.1 English Text to Speech Conversion.

D5.5.2.4 Integrated Text to Speech Conversion.

D6.1 Use Case Implementation.

D6.2 Use Case Evaluation.

Consequently, Carnegie Mellon University Africa agreed to take joint responsibility for these deliverables. Since this involved a significant amount of additional, unplanned e↵ort, only one use case, the laboratory tour, has been implemented to date, leaving the second use case, the receptionist, to be implemented next year, along with the use case evaluation.

2.3

Submitted Deliverables

Table 1 lists the 32 deliverables that have been submitted to date. For more details, please refer to the deliverable reports on the CSSR4Africa website, also linked from the deliverable number in Table 1. The full list of deliverables due over the lifetime of the project, their due dates, their submission dates, and their revision dates, are available on the CSSR4Africa website. In the following, we summarize the work done to date, organized by work package topic.

2.4

African Cultural Knowledge

Deliverable D1.2 comprises a compendium of population-based cultural knowledge regarding behaviors, activities, actions, and movements that are either culturally sensitive or insensitive.

This knowledge is formalized in the cultural knowledge ontology and knowledge base in Deliverable D5.4.1. The cultural knowledge has been gathered by developing a detailed questionnaire and using it to survey a cross-section of Rwandan citizens. The survey is available online in Kinyarwanda and English.

2.5

Use Case Scenario Specification

Deliverables D2.1, D2.2, and D2.3 together present a detailed scenario definition for the two use cases — lab tour guide and receptionist — along with the associated robot behavior specification, and the visitor behavior specification. The layout of the environment in which the scenarios are set is specified by the environment knowledge base detailed in Deliverable D5.4.2.

2.6

Systems Engineering

Deliverable D3.1 specifies the CSSR4Africa system architecture, identifying the component subsystems, the modules (i.e., ROS nodes) comprising each subsystem, and the information exchanged between subsystems and modules. The architecture itself is implemented in ROS using ROS nodes developed and documented in deliverables in work packages 4 and 5, i.e., Deliverables 4.2.1 – D4.2.4 and Deliverables D5.2 – D5.5.4.

Deliverables D3.2, D3.3, D3.4, and D3.5 provide the software engineering foundation for the implementation of the CSSR4Africa system: a reference manual of software engineering standards, a comprehensive installation manual, the procedures used to validate and test software 2 20

D1.2 D2.1 D2.2 D2.3 D3.1 D3.2 D3.3 D3.4 D3.5 D4.1 D4.2.1 D4.2.2 D4.2.3 D4.2.4 D4.3.2 D5.1 D5.2 D5.3 D5.4.1 D5.4.2 D5.4.3 D5.5.1.1 D5.5.1.2 D5.5.2.1 D5.5.2.3 D5.5.2.4 D5.5.3 D5.5.4 D6.1 D7.1 D7.3 D8.1.3

Rwandan Cultural Knowledge, version 1 Use Case Scenario Definition Robot Behavior Specification, version 1 Visitor Behavior Specification, version 1 System Architecture, version 1 Software Engineering Standards Manual Software Installation Manual System Integration and Quality Assurance Manual System Integration and Quality Assurance Sensor Tests Person Detection and Localization Face & Mutual Gaze Detection and Localization Sound Detection and Localization Robot Localization Speech Event Actuator Tests Animate Behavior Subsystem Attention Subsystem Cultural Knowledge Ontology & Culture Knowledge Base Robot Mission Language Robot Mission Interpreter Gesture Execution Programming by Demonstration English Text to Speech Conversion Kinyarwanda Text to Speech Conversion Integrated Text to Speech Conversion Environment Map Generation Robot Navigation Use Case Implementation Online Presence Open-Source Software Repository Progress Report

Table 1: CSSR4Africa deliverables submitted to date; click on the deliverable number to see the full report. developed by the partners prior to integration into the CSSR4Africa software repository, and current progress in integrating the software modules and adding them to the software repository on GitHub.

2.7

Robot Sensing

Deliverable D4.1 comprises a suite of unit tests to verify that sensor data is successfully acquired on each sensor topic. The functionality encapsulated in this deliverable forms the basis for all the sensory ROS nodes in the CSSR4Africa system architecture, i.e., Deliverables D4.2.1– D4.3.2.

Deliverable D4.2.1 documents a ROS node, personDetection, which detects and localizes people under various conditions. Deliverable D4.2.2 documents a ROS node, faceDetection, that detects and localizes human faces under various conditions and determines whether mutual gaze is established between the Pepper robot and the human user through head pose estimation.

Deliverable D4.2.3 documents a ROS node, soundDetection, to detect and localize conspicuous sounds within a robot’s hearing range. It provides two outputs:(a) the direction of arrival (DoA) 3 21

of the sound, and (b) a filtered audio signal. It interfaces with the speechEvent ROS node and the overtAttention ROS node, facilitating automatic speech recognition (ASR) and enabling the robot to focus its attention on sound sources. Deliverable D4.2.4 documents a ROS node, robotLocalization, that estimates the pose of the Pepper robot in a Cartesian world frame of reference. The module achieves this through a combination of relative and absolute position estimation techniques, including odometry, IMU data, and triangulation using visually detected ArUco marker visual landmarks. Deliverable D4.3.2 documents a ROS node, speechEvent, which deploys a speech-to-text model using deep neural networks that enables speech utterances in Kinyarwanda and English languages captured by Pepper’s microphones to be transcribed into written text.

2.8

Robot Behaviors

Mirroring Deliverable D4.1, Deliverable D5.1 comprises a suite of unit tests to verify the accurate and reliable functioning of the actuators: head, arms, hands, legs, and wheels. The functionality encapsulated in this deliverable forms the basis for all the ROS nodes in the CSSR4Africa system architecture that control the actuators, i.e., Deliverables D5.2, D5.3, D5.5.1.1, D5.5.4, in addition to D5.5.1.2 which operates separately in a stand-alone mode.

Deliverable D5.2 documents a ROS node, animateBehaviour, which enhances the robot’s lifelike appearance by generating subtle body movements to flex the robot’s hands and slightly rotate its base around the Z-axis. Deliverable D5.3 documents a ROS node, overtAttention, which implements of the attention subsystem, enabling the robot to dynamically direct its attention towards salient features in its environment, primarily during social interactions, and to scan its surroundings when not actively engaged. The attention subsystem integrates several processes, including face and sound detection, saliency map generation, and gaze control, allowing the robot to engage in more human-like behaviors. Deliverable D5.5.1.1 documents a ROS node, gestureExecution, which enables the Pepper robot to perform a range of body and hand gestures: deictic, symbolic, and iconic hand gestures, and bowing and nodding body movements. Hand gestures are executed so that they exhibit biological motion, since this has been shown to enhance people’s engagement when interacting with the robot (Akinade, Barros, & Vernon, 2025). The node interfaces with the robotLocalization node, so that deictic gestures accurately indicate the location of points of interest in the robot’s environment. Additionally, it interfaces with the overtAttention node to control the head during diectic gestures toward a location. Deliverable D5.5.1.2 provides the robot with the ability to learn gestures through manual teleoperation or human demonstration, employing RGB-D camera technology to map human skeletal movements onto the robot’s joint system, i.e., Programming by Demonstration. Deliverable D5.5.4 documents a ROS node, robotNavigation, enabling the Pepper robot to traverse its environment autonomously. It too interfaces with the robotLocalization node, continuously updating its pose with real-time data to ensure the navigation path is registered with the robot’s environment.

Deliverables D5.4.1, D5.4.2, and D5.4.3 together provide the core functionality for executing a robot mission. Deliverable D5.4.2 documents the formal methodology — behaviour trees — for specifying a robot mission based on the use case scenarios documented in Deliverables D2.1, D2.2, and D2.3, and enact them in a culturally sensitive manner using the culture knowledge base and an environment knowledge base. This deliverable also documents and implements the environment knowledge base with the information required to complete the robot mission. Deliverable D5.4.3 documents the Robot Mission Interpreter, implemented as a ROS node, behaviorController, a central component in the CSSR4Africa system architecture, which interprets and executes the robot mission specifications defined in Deliverable D6.1. The behaviorController node interfaces with eight ROS nodes: animateBehavior, overtAttention, gestureExecution, textToSpeech, robotNavigation, tabletEvent, speechEvent, and faceDetection. Finally, Deliverable D5.4.1 formalizes the Rwandan modes of social inter4 22

action documented in Deliverable D1.2. It presents a cultural knowledge ontology and, based on this ontology, a simple representation of the cultural knowledge documented in Deliverable D1.2, in the form of the cultural parameter values that can be used by the robot to emulate these polite and respectful behaviours, activities, actions, and motions.

Deliverables D5.5.2.1, D5.5.2.3, and D5.5.2.4 together provide the functionality for converting English and Kinyarwanda text to speech, all encapsulated in the textToSpeech ROS node.

2.9

Use Case Demonstration and Evaluation

Deliverable 6.1 formalizes the use case scenarios in Deliverables D2.1, D2.2, and D2.3 using the robot mission specification methodology documented in Deliverable D5.4.2, i.e., using behavior trees. The resultant behavior tree provides the input to the behaviorController ROS node documented in Deliverable D5.4.3.

2.10

Dissemination and Impact

Deliverable D7.1 takes the form of a website, a wiki, and a report. The website is primarily a forum for disseminating information about the project to a broader audience, while the wiki is primarily a forum for exchanging information among the partners in the project consortium.

The website domain name is www.cssr4afrcia.org and is hosted on GitHub. The wiki is also hosted on GitHub and it is linked from the website. The ROS nodes described above are available on the CSSR4Africa software repository, documented in Deliverable D7.3.

2.11

Project Management

Six-monthly progress reports are provided in Deliverable D8.1. To date, there are three versions D8.1.1, D8.1.2, and D8.1.3. The agendas and minutes of the weekly CMU-Africa team meetings are available on the CSSR4Africa wiki.

3

Next Targets

With the departure of the PI, David Vernon, on 30 June 2025, the project will be concluded by the co-PIs at the University of the Witwatersrand. During the final year, and as anticipated in the work plan, most of the work will focus on assessment and enhancement of the ROS software to improve the performance of the Pepper robot in culturally-sensitive social interaction, in addition to completing the outstanding deliverables for which the University of the Witwatersrand is responsible. Finally, the behavior tree specification of the robot mission for the receptionist scenario use case has yet to be written and tested using the ROS behaviorTree node and the cssr africa ROS package as a whole.

4

Peer-Reviewed Publications Resulting from this Project

Akinade, A., Barros, D., Danso, M., Haile, Y., Birhan, E., Shimelis Girma, B., ... Vernon, D. (2025). Culturally sensitive social robotics for Africa. In Proceedings of the 2nd international workshop on cultural robotics: Diversified sustainable practices. Springer LNAI, in press.

Akinade, A., Barros, D., & Vernon, D. (2025). Biological motion aids gestural communication by humanoid social robots. International Journal of Humanoid Robotics, Vol. 22, No. 02.

Akinade, A., Haile, Y., Mutangana, N., Tucker, C., & Vernon, D. (2023). Culturally competent social robots target inclusion in Africa. Science Robotics, 8(85).

Vernon, D. (2024). An African perspective on culturally competent social robotics: Why DEI matters in HRI. IEEE Robotics and Automation Magazine, 31(4), 170–200.

Zantou, P. & Vernon, D. (2023). Culturally-sensitive human-robot interaction: A case study with the pepper humanoid robot. In Proceeding of IEEE Africon, Nairobi, Kenya.

5 23

Deep Learning Driven Multiplexed Prospectivity Modeling of Rare Earths in Radiothermic Carbonatites Angeyo1 H.K., Kaniu1 I.M., Usman2 I., Rwabuhungu3 D.E. 1 Department of Physics, University of Nairobi, Nairobi, Kenya.

2 School of Physics, University of the Witwatersrand, Johannesburg, South Africa.

3 School of Mining and Geology, University of Rwanda, Kigali, Rwanda.

1.

PROJECT OBJECTIVES There is increased demand for innovative mineral prospecting to identify and define strategic deposits as well as delimit the geologic structures that host them. Rare earths occur mostly in alkaline carbonatite complexes of which numerous are found in East Africa. Due to their unique properties rare earth elements (REE) are considered a strategic resource for applications in advanced technologies and carbon-free fuels.

Alkaline carbonatites (which are the primary source of the world’s rare REE) have been a natural part of the magmatic history of Eastern Africa [1-3]. The region’s carbonatites contain particularly high (³ 40%) total rare earth oxides (TREO). The deposits occur in complex association with heavy minerals such as monazite, bastnaesite, synchesite and parisite [4]. The carbonatite complexes of East Africa include Mrima [5], Homa and Ruri [6], Kerio [7], Tinderet [1], Buru [8], Oldoinyo Nyegi, Shompole [6], Napak, Oldoinyo-Dili, Tundulu, Mbeya, Kerimasi, Hanang, Kwaraha, Lashaine, Kazekere, Mahoma, Fort Portal, Natron-Engaruka, Chilwa, Rufunza, Chilwa Island, Matopon, Rufunza, Rangwe, Sadiman, Burko, Esimingor, Katwe, Bunyaruguru, Kikorongo, Tororo, Sukulu, Teno, Lokupoi, Panda [9], Wigu, Ngualla [10], etc.

In the high background radiation area (HBRA) carbonatites the unusually high uranium content in some heavy minerals such as zircon and ilmenite, can point to igneous syngenetic radioactive deposits that are also the radiogenic sources driving the remarked geothermal activity in the deposits. The alignment of most of East Africa’s carbonatites along the crustal locations of the Rift Valley indicates fractures of continental dimensions reaching into the mantle. Mantle-derived carbonatite melts are known to be carriers of REE. Thus REE are typically associated with uranic mineralogy and could be used as proxies for uranium prospecting and for gaining insight into U-Th geochemistry [11] to delineate REE-U-Th co-dependency.

Little is known about the chemistry of fluids that immobilize and concentrate REE in the radiothermic carbonatites; or the favourable environments for their exploration, although preliminary (mostly radiometric) studies have been done in some of the carbonatite complexes [6, 12-15]. Carbonatites may be absent in some REE deposits; however their formation is closely related to carbonatite magma as the rare earth minerals of hydrothermal type are formed by fluids that evolved from magmas [16]. Consequently there is no reliable evidence to constrain the genesis of radiothermic carbonatites. As a result current prospective methods for radiothermic carbonatite REEs cannot accurately aggregate their complex attributes and model them in relation to the multivariate processes associated with their complex geochemistry and mineralogy.

There is a growing consensus that solutions to problems where it is infeasible to run mechanistic models at desired resolutions in space and time require methods that integrate physics-based modeling with deep learning. This study, being at the realm where physics, geosciences and computing intersect is about turning complex geo-scientific challenges into simple deep learning driven critical mineral resources prospectivity solutions. This study investigates how and under what conditions deep learning driven spectral and imaging analytical models can be used to detect and characterize the rare earths mineral prospect in the radiothermic carbonatites of East Africa in relation to the associated uranium and thorium potential.

Carbonatite geochemistry is very diagnostic of both rare earth element (REE) and radioactive anomalies.

1

24

1.1 The Goal The goal of the study is to develop in articfical intellgence (AI) computational domain, deep learning-driven multiplexed spectral and spectral imaging analytical models for direct analysis and characterization of the rare earths mineral prospect in the radiothermically stressed carbonatite complexes of Eastern Africa.

Specifically to; (i) perform space-borne hyperspectral imaging to compute ‘hot spots’ corresponding to uranium and thorium associated REE in selected radiothermic carbonatite complexes of Eastern Africa. (ii) spectroscopically determine the concentrations of REE and associated minerals in the REE ore matrices (soil, rock, sediment) of selected radiothermic carbonatite complexes of Eastern Africa. (iii) perform spectral imaging and multivariate image analysis (MIA) of REE mineralogy and microstructure in the ore matrices of the studied carbonatites. (iv) distinguish and resolve the geochemical ‘signatures’ of radiogenically and hydrothermaly stressed carbonatite REE minerals using data from (i)-(iii) for the studied carbonatites. (v) model via AI (chemometrics, machine learning, deep learning) the multivariate relationships between occurrence, levels and composition of REE ore mineralogy of the studied carbonatites.

2.

RECENT PROGRESS 2.1 Space-Borne Hyperspectral Imaging We used hyperspectral imaging (HIS) to infer the occurrence of radioactive minerals in the Mrima Hill and Oldoinyo Lengai using ENVI and ArcGIS software. Hyperspectral imaging (HSI) integrates conventional imaging and spectroscopy, to obtain both spatial and spectral information from a specimen thus offering potential for detection, identification and visualization. Theoretically, space-borne spectral reflectance can identify minerals and consequently rocks (which are but mineral aggregations containing a wide range of molecules) using remotely sensed data [17-19]. K-means clustering and the maximum likelihood classifier (MLC) were applied to Landsat-8 Imager data [20] to yield ‘hot spot’ images which were then radiometrically calibrated to change the pixel values from DN (digital number) to top of atmospheric (ToA), followed by atmospheric correction using DOS (dark object subtraction) [21]. The diagnostic absorption features of the REE were generated in the visible to shortwave infrared (VNIR-SWIR) spectral region.

2

25

Fig 1.

Satellite spectrometry of REE mineralized zones for Mrima Hill carbonatite.

2.2 In situ Gamma-ray Spectrometry Field air-absorbed gamma dose-rates measurements are now being used to ground-truth the HIS results by first using a personal radiation detector to identify the ‘hot spots’, followed by in situ gamma-ray spectrometry to assess radioactivity levels at the ‘hot spots’ to guide the sampling of selected carbonatite matrices (soil, rocks, ores, sediment) for laboratory-based spectroscopic and spectral imaging analyses.

Fig 2.

Interpolated g- dose-rates by Kriging technique in Mrima Hill HBRA overlaid by a sampling grid.

2.3 XRF Spectrometry Although X-ray fluorescence (XRF) spectrometry is a sensitive method for elemental analysis, it is still plagued by challenges when directly analyzing complex matrix materials. Field-based trace analysis using portable, hand-held XRF requires a considerable push in analytical strategy. We have developed a method that may be scaled to applications using hand-held XRF spectrometry for the work described in this project, 3

26

enabled by machine learning, called Energy Dispersive X-Ray Fluorescence and Scattering (EDXRFS) spectrometry [22-24] which exploits, in addition to fluorescence, the scatter peaks obtained from complex matrix materials to realize rapid trace element analysis and modelling as well as prediction of the chemical properties of materials. Hand-held XRF [25, 26] is an under-utilized tool for the type of tasks in the current study as it is relatively low cost, time-efficient, simple and rapid; moreover, the deep learning protocols we are developing are programmable and thus embeddable in portable systems for AI-based workflows.

2.4 Data Fusion Combining and exploring in multivariate space, the data from the multiplexed platforms increases the accuracy and robustness of the models due to the complementarity of the information furnished by the multimodal approach. Data fusion is used arrive at robust rare earth prospectivity interpretations.

3.

NEXT TARGETS 3.1 Laser Induced Breakdown Spectrometry (LIBS) Situations where measurements need to be made rapidly and at stand-off makes LIBS suitable for this work.

The remarkable attributes of LIBS include (i) real-time response, (ii) in-situ analysis with little or no sample preparation, and (iii) a high sensitivity to the low-Z elements (does not face the analytical limitations of XRF for Z < 14 elements). Towards hand-held (HH)-LIBS [27-29] higher laser energy and shorter time delay will minimize measurement uncertainty by reducing the continuum emission due to the limited avalanche ionization. Higher sensitivity will however be more cost-effectively realizable in nanoparticle enhanced (NE)LIBS where the laser electromagnetic field is locally enhanced and a high current due to electron field emission is induced even at low irradiances to achieve ‘spectroscopic enhancement’ [30, 31].

3.2 Laser Ablation Molecular Isotopic Spectrometry (LAMIS) Molecular spectra are the basis for isotopic analysis at atmospheric pressure via a recent extension of LIBS namely laser ablation molecular isotopic spectrometry (LAMIS) [32] in which because the plasma exhibits large isotopic splitting due to contributions of the molecular rotational and vibrational states [33], it is possible to study different isotopes in a sample [34]. During our implementation of LAMIS, multiple emission lines will be used to improve the isotopic measurements [35].

3.3 Deep Learning The utility of the above techniques for direct and therefore rapid analysis and imaging is limited by sample complexity and data interpretation. The challenges in the detectability, processing, analysis, modelling and interpretation of the resulting spectral and image data may be overcome by deep learning [36] which via multivariate modeling can better delineate prospective rare earths [37].

Although the applications of deep learning have been reported in diverse fields [38, 39] until now only little has been applied in geosciences. Deep learning allows computational models that are built from sample inputs and upon multiple processing layers to learn data representations with multiple levels of abstraction and make predictions based on subsequent data. Machine learning finds it challenging to work with increased spectral data containing changing artefacts from multiple sources. Deep learning will be used to reduce the data dimensionality and extract rare earths signatures from the complex spectra and images, develop multivariate calibration strategies for trace analysis, and perform exploratory modeling analysis and attribution relative to diverse phenomena. Deep learning also enables mining (management, analysis and visualization) of large databases. Random forest, support vector machines, elastic net, decision tree, neural network and independent components analysis will be used in the Scikirt-Learn, Python, and R software environments.

4

27

4.

[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] 5.

Citations Caswell PV, Baker BH (1953). Geology of Mombasa-Kwale Area, Rept. No. 24, Mines and Geology Department, Ministry of Natural Resources, Kenya.

Riaroh D, Okot W (1994). Tectonophysics, 236, 117–130.

Smith CW (1956). Quarterly Journal of the Geological Society, 112,189-219.

Deans T, Roberts B (1984). Journal of the Geological Society, London 141, 563-580.

Mustapha AO (1999). Rad Prot Dosimetry 82, 285-292.

Achola SO, Patel JP, Mustapha AO, Angeyo KH (2012). Rad Prot Dosimetry 152 (4), 423–428.

Mangala JM (1987). MSc thesis, University of Nairobi.

Geological Survey of Kenya (1962). “Kenya Geological Map”. 2nd edn, Nairobi, Kenya.

Yu L, Kapustin L, Polyakov AI (1985). Int Geology Review 27 (4), 434-448.

Harmer RE, Nex PAM. (2016). Episodes 39 (2), 381-406.

Verplanck PL, Gosen BS (2011). Carbonatite and alkaline intrusion-related rare earth element deposits – a deposit model, Technical Report, US Geological Survey.

Kaniu MI, Darby IG, Angeyo KH (2019). J. Afr. Earth Sci 160 (8), 103606.

Kaniu MI, Angeyo HK (2018). Radiat. Phys. Chem. 146, 34-41.

Kaniu MI, Angeyo HK, Darby IG, Muia L (2018). J. Environ. Radioact. 188, 47-57.

Mukhono PM, Angeyo KH, Dahayem-Massop A, Kaduki AK (2013). Spectrochim Acta B 87, 81–85.

Anenburg M, Burnham AD, Mavrogenes JA (2018). Can. Mineral. 56, 331–354.

Boesche NK, Rogass C, Lubitz C, Brell M, Herrmann S, et al. (2015). Remote Sens 7, 5160-5186.

Neave DA, Black M, Riley TR, Gibson SA, Ferrier G, et al. (2016). Econ Geol 111, 641–665.

Turner DJ, Rivard B, Groat LA (2014). Am Mineral 99, 1335-1346.

Dev Acharya, T., Yang, I., Student, G. Int. J. IT, Eng. Appl. Sci. Res. 4 (2015), 4-10.

Ondieki JO, Mito CO, Kaniu MI (2022). Int. J. Appl Earth Obs Geoinf 107, 102700.

Angeyo HK, Gari S, Mangala JM, Mustapha AO. (2012). X-Ray Spectrom 41, 321-327.

Kaniu MI, Angeyo HK, Mwala AK, Mangala MJ. (2012). Anal Chim Acta 729, 21-25.

Angeyo HK, Gari S, Mangala JM, Mustapha AO. (2012). Appl Radiat Isot 70, 2596-2601.

Schramm R (2016). Phy Sci Rev 1 (9) DO: 1515/psr-2016-0061.

Lemière B (2015). Paper presented at the 27th International Applied Geochemistry Symposium, Tucson, Arizona, USA, 20–24 April 2015.

Harmon RS, Remus J, McMillan NJ, et al. (2015). Appl Geoch 24, 1125-1141.

Wainner RT, Harmon RS, Miziolek KL, et al. (2001). Spectrochim. Acta B 56, 777–793.

Crocombe RA (2013). Handheld spectrometers: the state of the art”, Proc. SPIE 8726.

Nie S, Emory SR (1997). Science 275 (5303), 1102-1106.

Adams FC, Barbante C (2013). Spectrochim Acta Part B 86, 3-13.

Russo RE, Bol'shakov AA, Mao X, et al. (2011). Spectrochim Acta B 66 (2), 99-104.

Mao X, Bol'shakov AA, Perry DL, Sorkhabi O, et al. (2011). Spectrochim Acta B 66 (8), 604-609.

Mao X, Bol’shakov AA, Choi I, et al. (2011). Spectrochim. Acta Part B 66 (11–12), 767-775.

Chan GC-Y, Mao X, Choi I, Sarkar A, Lam OP, et al. (2013). Spectrochim Acta B 89, 40-49.

Sun T, Li, H, Wu, K, Chen, F, Zhu, Z, Hu, Z-J (2020). China. Min 10, 102.

Carranza EJM, Laborte AG (2015). Ore. Geol. Rev.71, 777–787.

Fu X, Zhong L, Cao Y, Chen H, Lu, F (2021). Anal. Meth 13, 64-68.

Padarian J, Minasny B, McBratney AB (2018). Geoderma Reg. 15, e00198.

PEER-REVIEWED PUBLICATIONS No publications to report for the moment.

5

28

Digital Malaria Control for the Developing World Carine Pierrette Mukamakuza cmukamak@andrew.cmu.edu Carnegie Mellon University Africa Jean Paul Niyoyita jniyoyita@ur.ac.rw University of Rwanda, College of Medicine and Health Sciences Project Objectives The persistent burden of malaria in Rwanda has created a diagnostic bottleneck that undermines treatment efficacy and surveillance efforts, especially in resource-constrained areas. Traditional microscopy-based diagnosis, still considered the gold standard, suffers from limitations in field conditions. First, it requires specialized laboratory infrastructure often unavailable in rural healthcare facilities. Second, there is a shortage of skilled technicians with expertise in parasite identification, particularly in remote communities where malaria is most prevalent. Third, the process introduces substantial delays between sample collection and results generation, often extending to days rather than hours, directly impacting treatment decisions. Fourth, traditional methods mostly struggle with diagnostic accuracy, especially for low-parasitemia cases or mixed infections that require expert interpretation. Finally, manual record-keeping of sensitive healthcare data raises privacy and security concerns. These diagnostic inefficiencies translate directly into delayed treatments, ineffective resource allocation, and compromised epidemiological monitoring.

To address these critical gaps, this project proposes the development and deployment of an end-toend AI-powered malaria diagnosis and geospatial surveillance system, tailored specifically for resourceconstrained environments in Rwanda. As outlined in Figure 1, the project is guided by five interrelated objectives that form the foundation of the system’s design and implementation.

First, the project aims to build a fully functional diagnostic application that integrates AI-based detection models into routine clinical laboratory workflows. This application is being co-designed with stakeholders from the Rwanda Biomedical Center (RBC), who provide domain expertise, validate system requirements, and facilitate adoption within healthcare settings. Second, we are constructing a high-quality, annotated dataset comprising 2,500 microscopic images representing the four human-infecting Plasmodium species (P. falciparum,P. vivax,P. malariae , and P. ovale). This dataset will serve both as a foundation for training deep learning models and as an open-access resource for the broader scientific community focused on malaria research.

Third, the project will evaluate and optimize state-of-the-art object detection models to ensure accurate identification and classification of malaria parasites. By comparing multiple deep learning architectures, we aim to select the most reliable models for automated diagnosis and quantification, with performance validated against expert annotations. Fourth, we will incorporate geospatial analysis tools to enable finegrained malaria surveillance at the village and sector administrative levels. These tools will support the identification of high-risk microregions and inform precisely targeted intervention strategies, contributing to more effective resource allocation and improved disease control.

Finally, the system will be developed with a strong emphasis on integration within existing healthcare infrastructures to ensure its long-term sustainability and real-world impact. By aligning technological innovation with local health system needs and constraints, the project seeks to deliver a scalable, data-driven approach to malaria control and surveillance.

29

Figure 1: Visual representation of malaria diagnosis digitization goals: from parasite identification to surveillance integration.

Recent Progress Our implementation has successfully delivered several key components of the malaria diagnosis and surveillance system. We have developed a user-friendly web-based Clinical Decision Support System integrated with a parasite detection model. This platform accepts digitized microscopy images from both thin and thick blood smears, identifies parasite species and white blood cells in batch processing mode, computes parasite densities using WHO-aligned metrics, and generates secure clinical reports, all within a browser-accessible framework Designed specifically for low-resource environments. The platform is hardware-independent and intuitive enough for non-specialist users while maintaining secure data management with robust authentication controls to protect sensitive patient information.

Over 2,000 images representing the four Plasmodium species have been meticulously validated and annotated to support dataset creation, comprising 602 P. falciparum, 290 P. vivax, 824 P. ovale, and 727 P. malariae. Figure 4 displays sample annotations with species predictions, bounding boxes, and confidence scores.

We evaluated several state-of-the-art deep learning architectures for parasite detection, including earlier models like YOLOv5. Our tests show that the newer YOLOv10 architecture outperforms the models listed in Table 1, offering an ideal balance of accuracy, speed, and efficiency for use in resource-limited healthcare settings. Previously, YOLOv8-MobileViT was the top performer, with strong detection rates for P. ovale (0.912) and P. vivax (0.930), and solid results for P. falciparum (0.767) and P. malariae (0.738). However, YOLOv10 delivers even better accuracy and reliability. Despite these advances, detecting P. falciparum remains difficult due to its small size and ring-like shape. Table 1 provides a detailed comparison of specieswise detection performance across models.

For geo-spatial analysis of malaria risk factors, two key components for malaria surveillance have been completed. First, microgeographic mapping has finalized the analysis of malaria risk factors in Bugesera

30

district, Rwanda, at the village level using data from 2021–2023. This study revealed strong correlations between malaria transmission and outdoor occupations, migrant populations, and poor housing structures.

Second, an automated data pipeline has been developed as a streamlined system for malaria surveillance data collection, processing, and visualization, providing health officials with interactive maps and dashboards to support evidence-based decision making as illustrated in Figure 2 and Figure 3.

(a) Geospatial dashboard showing malaria positivity trends by gender, region, and temperature (2021–2023).

(b) Web-based system overview with diagnosis summaries, turnaround time, and simulated satisfaction ratings.

(c) Image upload and diagnostic progress view for patient cases.

Figure 2: Geospatial surveillance and MalariaAI platform components for clinical decision support

31

Figure 3: Detailed patient diagnosis report including parasite count, density, and severity.

(a) P. vivax detections (PV)

(b) P. malariae detections

(c) P. ovale detections (PO)

Figure 4: Sample image-level parasite annotations from the validated dataset showing species-level predictions with bounding boxes and confidence scores.

Preliminary Results of the Models Parasite PF PM PO PV

Faster R-CNN

YOLOv5

Mask R-CNN

YOLOv8-MobileVit

0.658 0.843 0.857 0.947

0.660 0.853 0.883 0.938

0.696 0.743 0.860 0.918

0.767 0.738 0.912 0.930

Table 1: Model performance on malaria parasite detection

32

Next Target The next phase of our project will focus on several key priorities. First, we will complete rigorous testing and preparation of the web-based system for piloting in Rwandan healthcare facilities. A vital component of this effort will be integrating the system with existing hospital information systems to create a seamless workflow that fits naturally into clinical practice. Simultaneously, we will develop a companion mobile application that allows field healthcare workers to capture and submit blood smear images even in areas with limited connectivity.

For dataset creation, we will expand our collection to reach 2,500 balanced images across all parasite species, conduct external validation to confirm dataset quality, and prepare the dataset with annotations for publication as a resource for the global malaria research community. To improve accuracy, we will retrain our models using this expanded dataset, with special attention to improving detection of difficult-to-identify parasites like P. falciparum. We will also conduct thorough evaluation against traditional microscopy to validate our system’s performance in real-world conditions.

Our malaria surveillance capabilities will be extended beyond Bugesera to additional districts in Rwanda, with field testing to ensure the system meets the needs of local health officials. We will also strengthen the integration between the surveillance dashboard and the main diagnosis system, creating a platform that not only supports individual patient care but also enables an effective public health response to malaria outbreaks.

Peer-Reviewed Publications [1] C. P. Mukamakuza, A. D. Nishimwe Karasira, E. M. Akpo, Y. A. Bogale, P. Fasouli and M. Salem, “A Comparative Analysis of Deep Learning Models for Malaria Plasmodium Classification,” 2024 31st IEEE International Conference on Electronics, Circuits and Systems (ICECS), Nancy, France, 2024, pp. 1–4, doi: 10.1109/ICECS61496.2024.10848723.

[2] H. R. Mary, C. P. Mukamakuza and E. Tuyishimire, "A Data Management Model for Malaria Control: A Case of Rwanda," 2023 IEEE AFRICON, Nairobi, Kenya, 2023, pp. 1–6, doi: 10.1109/AFRICON55910.2023.10293671.

[3] E. M. Akpo, C. P. Mukamakuza and E. Tuyishimire, “Binary Segmentation of Malaria Parasites Using U-Net Segmentation Approach: A Case of Rwanda,” 2024 Ninth International Congress on Information and Communication Technology (ICICT), Singapore, 2024, pp. 1–12, doi: 10.1007/978-981-97-4581-4_12.

[4] A. D. Nishimwe Karasira, C. P. Mukamakuza and E. Tuyishimire, “The Use of YOLOv5 as a Malaria Detection Model for the Developing World,” Proceedings of Ninth International Congress on Information and Communication Technology (ICICT 2024), London, UK, 2024, pp. 615–626, doi: 10.1007/978-981-97-3299-9_50.

[5] Y. Bogale, C. P. Mukamakuza and E. Tuyishimire, “Intelligent Malaria Detection and Species Classification: A Case of Rwanda,” Proceedings of Ninth International Congress on Information and Communication Technology (ICICT 2024), London, UK, 2024, doi: 10.1007/978-981-97-3299-9_41.

[6] I. Umuhoza and C. Mukamakuza, “SVM Model-Based Digital System for Malaria Screening and Parasite Monitoring,” 2023 IEEE Smart Cities Conference (SCC), 2023, pp. 1–6, doi: 10.1109/SCC59637.2023.10527507.

33

EmpoweringAfrican African Communities: Communities: Culturally Empowering CulturallyRelevant Relevant Cybersecurity Education through Comic Books Cybersecurity Education through Comic Books and AI and AI Models for Children’s Online Protection Models for Children’s Online Protection Jema David Ndibwile, Carnegie Mellon University Africa Jean Francois Maniraho, University of Rwanda

1

Project Objectives

This project aims to design, develop, and evaluate culturally appropriate, AI-powered cybersecurity education tools to protect African children online—particularly those in underserved settings such as refugee camps and rural communities. We explore multiple EdTech approaches, including machine learning models for detecting abusive content in low-resource languages, comic-based storytelling for cybersecurity awareness, and gamified digital platforms for interactive learning. The overarching goal is to create accessible, engaging, and locally resonant cybersecurity education that empowers young African learners with the skills and awareness to navigate online environments safely.

2

Recent Progress

2.1

Machine Learning for Abusive Language Detection (IJCAI 2025 - Accepted):

Developed and tested models such as Decision Trees, SVMs, and Logistic Regression on a Swahili dataset. Decision Trees achieved 99% accuracy. IJCAI is top-tier (top 4) conferences in artificial intelligence. We will present this work in Monreal Canada in August 6.

2.1.1 Comparison of Models Performance

Figure 1: Mean cross-validation accuracy scores for different models

1 34

Model

Precision

Recall

F1-Score

Logistic Regression Decision Tree Random Forest SVM

0.88 0.98 0.79 0.86

0.88 1.00 0.86 0.88

0.88 0.99 0.83 0.87

Table 1: Model evaluations

2.2

Comic-Based Cybersecurity Education (Springer Nature - under review)

Deployed culturally tailored comics in schools across Rwanda. One rural school improved post-test scores by 75%, and 90% of students reported strong engagement.

2.2.1 Statistical Analysis It demonstrates that the comic book intervention was effective in diverse learning environments. In well-resourced schools like Mother Mary International School and Kigali Parents School, the intervention built upon existing educational structures to enhance learning. G.S. Paysannat, a refugee camp, and rural schools such as Rugeshi Primary School, had the biggest improvements, reinforcing the comic’s adaptiveness and possibility for mitigating educational inequities. Rugeshi had seen the largest improvement based on the group only, pre-test mean of 4.0 to post-test mean of 7.0, demonstrating the power of culturally contextualized educational resources.

Statistical Analysis of Pre- and Post-Test Scores School

Pre

Post

Pre SD

Post SD

t

p

Sig.

Mother Mary Intl. School G.S. Paysannat (Mahama) Remera Catholique Rugeshi Primary Kigali Parents School

8.1 8.4 6.4 4.0 4.2

9.0 8.7 8.0 7.0 6.7

0.9 0.7 1.0 1.2 1.1

0.8 0.6 0.9 1.0 1.0

3.87 2.14 4.82 7.74 5.65

0.001 0.042 0.000 0.000 0.000

Y es Y es Y es Y es Y es

2.2.2 Student Engagement and Feedback Likert-scale surveys further demonstrate the positive reception of the comic book. Students across all schools overwhelmingly responded positively, reinforcing how well regarded the comic was on multiple fronts, and its educational efficacy.

Student Feedback Highlights Across Schools School Mother Mary Intl. School G.S. Paysannat Remera Catholique Rugeshi Primary School Kigali Parents School

New Concepts (%)

Relatable (%)

Confidence (%)

Prefer Comic (%)

Want Vol. 2 (%)

100 100 100 100 100

95 98 100 92.3 96

90 92 89.7 94.9 91

95 100 100 100 98

97 100 100 100 99

These results highlight the consistent effectiveness of the comic book intervention across diverse educational contexts. The positive feedback underscores the value of culturally tailored, engaging educational materials in improving students’ cybersecurity awareness and knowledge.

2 35

2.3

Our Findings

Both statistical and qualitative data show that comics are a superior educational tool, particularly in cross-cultural and resource-limited settings. They foster engagement, reduce cognitive load, and enhance retention. Improvement rates ranged from 20–60%, with younger learners benefiting most. The cultural relevance of the comics further strengthened student interest and comprehension, reinforcing our assertion that tailored educational tools can bridge digital literacy gaps and provide scalable, effective cybersecurity education.

2.4

Gamified Interactive Digital Storytelling (GIDS) - submission pending

Designed a GIDS solution targeting refugee learners. Deployment was constrained by infrastructure limitations, but pilot responses showed potential.

3

Next Targets

3.1

Language Expansion for Inclusivity

We intend to enhance our existing dataset to include other low-resourced languages like Kinyarwanda, Luganda, and other native languages as we seek to have wider applicability in African regions. This will allow for the development of AI models that are more linguistically inclusive for communities.

3.2

Real-Time System Integration

The next important step is to integrate our machine learning classifiers into real-time learning and content-filtering platforms, which will enable real-time identification and removal of harmful or unsafe content and provide safety and knowledge development within the same context for a young audience.

3.3

Long-Term Impact Evaluation

We aim to assess the sustainability of learning by evaluating long-term retention of cybersecurity concepts and observing any behavioral changes among students. These follow-up evaluations will provide insight into the lasting effects of comic and GIDS-based interventions.

3.4

Cross-Country Piloting

To test the scalability and cultural adaptability of our tools, we plan to pilot the materials and technologies in other African countries. This will help determine localization needs and validate the relevance of our approach in diverse socio-cultural settings.

3.5

Offline Access through Local Infrastructure

Recognizing infrastructure limitations in remote areas, we will further develop the GIDS platform for offline use by deploying it on Raspberry Pi-powered local servers. This approach is particularly wellsuited to rural schools and refugee camps with limited internet access.

3.6

For Multimodal Abuse Detection

As a final step, we want to extend the abuse detection framework to multimodal AI, which means we are also able to process text, audio and image data. This means we can build more effective safeguards into multimedia platforms.

3 36

4

Peer-Reviewed Publications 1. P. Nabangi, A.-J. Zakaria, and J. D. Ndibwile, “Using Machine Learning to Enhance the Detection of Obfuscated Abusive Words in Swahili: A Focus on Child Safety,” Proceedings of IJCAI-25, acceppted, 2025.

2. J.D. Ndibwile, F.E. Ayite, A.I. Musonera, R. Mpinganima, S. Gisagara, and Y. Kadobayashi, “Enhancing Cybersecurity Resilience in African Children Through Culturally Tailored Comics,” Springer Nature- Journal of Computers in Education, under review, 2025.

3. J.D. Ndibwile, G. Rukundo, N.U. Ndzerem, J.F. Maniraho, “Enhancing Cybersecurity Awareness Through Interactive Digital Storytelling with Gamification,” [TBD],pending experiments, 2025.

4 37

EnhancedCardiovascular CardiovascularDiseases Diseases Discovery Discovery in in Medically Medically Underserved Enhanced Communities Communities via via AI-Assisted AI-Assisted Stethoscopy Stethoscopy Khalil Elkhodary khalile@aucegypt.edu, The American University in Cairo, Egypt Vijayakumar Bhagavatula kumar@ece.cmu.edu, Carnegie Mellon University, USA Carine Pierrette Mukamakuza cmukamak@andrew.cmu.edu, Carnegie Mellon University, Africa Francesco Renna francesco.renna@fc.up.pt University of Porto, Portugal Miguel Coimbra mcoimbra@fc.up.pt University of Porto, Portugal

Introduction Cardiovascular diseases (CVDs) remain the leading cause of death globally, accounting for an estimated 17.9 million deaths annually. In sub-Saharan Africa alone, over 1 million deaths were attributed to CVDs in 2019, representing 13% of all deaths in the region and 5.4% of global CVD-related mortality. Early diagnosis plays a crucial role in reducing CVD-related morbidity and mortality through timely clinical interventions and rehabilitation [1]. However, limited access to diagnostic infrastructure and routine screening in underserved communities continues to hinder early detection and effective management. To address this gap, there is a growing need for portable, cost-effective, and intelligent point-of-care (PoC) diagnostic systems tailored for low-resource settings [2].

As such, our research proposes an integrated diagnostic framework that combines computational modeling, multimodal signal acquisition, and machine learning to enable real-time, in-field detection of CVDs. Specifically, we aim to leverage compact hardware to capture physiological signals and fuse them with a physics-informed, machine learning-based diagnostic algorithm. By incorporating multi-physics constraints through physically informed neural networks (PINNs), we aim to provide a reliable, explainable, and scalable diagnostic tool that bridges the current gap in cardiovascular care across resource-limited environments.

38

Project objectives

Figure 1: Project objective of different Institutions, Universidade Do Porto, The American University in Cairo, and Carnegie Mellon University, Africa

Recent Progress - Machine learning and Feature Engineering Preliminary results utilizing a pilot dataset indicate that frequency band-specific energies and ratiometric measurements can be utilized for discrimination of the AS conditions. The severity of AS is often marked with the inclusion of specific morphology of murmur in the systole contents of PCG and can be detected utilizing appropriate segmentation of the cardiac phases of PCG and focusing towards the frequency bands of interest. The general observation is about the increased spectral energy in the high-frequency regime as the AS severity increases. The same has been demonstrated to be a reliable measure with a highly preprocessed publicly available dataset. Further analysis will provide a reliable estimate of the importance of the spectral measures in AS severity class distinction and development of an expert system with the engineered features. The advanced denoising methodology for data recorded in a hospital environment will be utilized, and the biomarkers will be validated for the real-time data obtained with CardioSleev for the developed in-house dataset.

39

Figure 2. Representative PCG signals for (A) Normal cases and similarly for (B) Severe Aortic Stenosis with pronounced systolic murmurs. The corresponding scalogram representations (C and D) reveal spectral continents in the systolic period as distinctive patterns. The segment-wise systolic power content (E) for Aortic Stenosis demonstrate spectral power concentration in specific high frequency (HF) range (80-120 Hz), and (F) HF band-specific power to total spectral power ratio is a stable measure (Median: 0.71) for relatively cleaner variety of PCG signals.

- Computational Analysis and Simulation A transient finite element analysis (FEA) is being executed independently as a first step before coupling with computational fluid dynamics (CFD). The first FEA run is required in order to check mesh quality, suitability of boundary conditions, and most importantly, ensure solver convergence for the structural subproblem. This decoupled validation approach is a common practice for FSI workflows, as it allows for the identification and removal of geometric or numeric instabilities that may undermine fully coupled simulations. Stable convergence of the FEA solver at this stage facilitates two-way coupling in multiphysics FSI analyses with assurance necessary for simulation of aortic valve biomechanics and fluid dynamics comprehensively. Figure 3 A and B show the opened and closed aortic valve respectively.

(A)

(B)

Figure 3. Transient FEA of Aortic Valve and Leaflets (A) during Systole (B) during Diastole

- Physical set up: At AUC, the full MCL, including a pulsatile pump, valves, a tubing network, and a filling tank, has been assembled.

Various measurement systems have been developed and employed: Pressure readings and measurements 1. Pressure gauges are connected through custom adaptors at key access points highlighted in Figure 4.

2. Data acquisition is done through an Arduino setup with a conversion from PSI to Pascals implemented in the code for simplicity purposes.

3. Pressure readings are taken to allow us to compare our results and findings to the concurrently developing CFD model. Secondly, we compare real life pressure readings to existing literature, including known in vivo pressure conditions to aid in MCL model development, precision and accuracy.

40

Figure 4. Access points throughout the MCL

Area Measurement Systems 1. An endoscope camera is placed and fixed orthogonally to the valve using a 3D printed ring 2. Area measurements are taken to allow us to quantify valve stenosis by identifying the maximum valve opening area during systole in the cardiac cycle, this allows us to quantitatively and directly conduct geometry-based assessment of stenosis severity based on the effective orifice area. The relevance of this quantitative assessment is to aid in the manufacturing of an array of different valves to the degree of stenosis of our liking.

Valve Fabrication: 1. First iteration of non-stenosed valve completed using 3D-printed molds.

2. Second iteration currently in progress to refine geometry and behavior.

Next Steps: Machine learning: U Porto - Aortic Stenosis: Target for the next phase The next targets compatible with the originally specified with the WP3 are as follows: 1. Development of a hybrid, model-based/data-driven system using PCG measurements for AS severity grading.

2. Design of multimodal ECG-PCG deep learning approaches for aortic stenosis grading: using opportunistic ECG data to detect a range of CVDs such as left ventricular hypertrophy, left atrial enlargement, atrial fibrillation, and conduction defects.

Computational Analysis and Simulation:

41

Physical setup: From AUC’s end: 1. Validate and categorise area readings and manufacture different valves based on the given data 2. Refining pressure data consistency across cardiac cycles 3. Once various valves are manufactured, begin systematic collection of PCG recordings 4. Preprocessing and organizing audio data for machine learning model training.

CMU Africa’s Next Steps Are: 1. Conduct a field study to gather validation data and test the performance of existing models on real-world signals.

2. Explore recent foundation models for audio analysis and apply them to auscultation data related to rheumatic heart disease (RHD).

3. Combine audio-based models with large language models (LLMs) to enhance explainability and classification performance.

Peer-reviewed publications: [1] A. M. Ali, A. H. Hafez, K. I. Elkhodary, and M. El-Morsi, “A CFD-FFT approach to hemoacoustics that enables degree of stenosis prediction from stethoscopic signals,” Heliyon, vol.

9, no. 7, Jul. 2023. doi:10.1016/j.heliyon.2023.e17643 [2] A. M. Ali, A. A. Ghobashy, A. A. Sultan, K. I. Elkhodary, and M. El-Morsi, “A 3D scaling law for supravalvular aortic stenosis suited for stethoscopic auscultations,” Heliyon, vol. 10, no. 4, Feb. 2024. doi:10.1016/j.heliyon.2024.e26190 References: [1] N. W. Minja, D. Nakagaayi, T. Aliku, W. Zhang, I. Ssinabulya, J. Nabaale, W. Amutuhaire, S. R. de Loizaga, E. Ndagire, J. Rwebembera, and E. Okello, "Cardiovascular diseases in Africa in the twenty-first century: gaps and priorities going forward," Frontiers in Cardiovascular Medicine, vol. 9, p. 1008335, 2022.

[2] J. Rajendran and G. Slaughter, "Transforming Cardiovascular Care—Biosensors and Their Potential: A Review," IEEE Sensors Journal, 2025.

42

Evaluating Digital Transformation and Maturity in Youth-Led Micro, Small, and Medium Enterprises across Sub-Saharan Africa: A Comparative Study in the Health, Energy, Environment and Sustainability Sectors in Nigeria, Kenya, and South Africa Ochieng’ Duncan Elly (University of Nairobi), Saruchera Fanny (University of the Witwatersrand), Obigbemi Imoleayo Foyeke (University of Lagos), Murimbika McEdward (University of the Witwatersrand), Omoro Nixon (University of Nairobi), Onsomu Zipporah (University of Nairobi) and Odock Stephen (University of Nairobi) Project Objectives This mixed-methods study investigates the digital transformation and maturity of Micro, Small, and Medium Enterprises (MSMEs) across the health, energy, environment, and sustainability sectors in Kenya, South Africa, and Nigeria. Using surveys and case studies, the research explores levels of digital adoption, identifies key challenges and enabling factors, and offers policy-relevant insights to support inclusive and sustainable digital growth.

The study aims to: (a) Assess the current state of digital transformation among MSMEs in selected sectors (b) Identify key barriers and enablers of digital maturity (c) Identify and document illustrative case studies of youth-led MSMEs that demonstrate varying levels of digital maturity and innovation (d) Generate data to support policy and practical interventions for enhancing digital capabilities (e) Develop a region-specific Digital Maturity Framework tailored to the Sub-Saharan African context (f) Disseminate findings and engage policy and ecosystem stakeholders Recent Progress A pilot test of the digital maturity survey was successfully conducted in Kenya, targeting youth-led Micro, Small, and Medium Enterprises (MSMEs) across the health, energy, environment, and sustainability sectors. The pilot phase aimed to validate the survey tool and generate preliminary insights into the digital capabilities and challenges facing these enterprises.

The findings from the pilot revealed several systemic barriers to digital adoption among MSMEs. Key among these were limited financial resources, lack of information technology expertise, inadequate digital infrastructure, low levels of digital literacy, and cultural resistance to technological change. These challenges are consistent with existing literature. For instance, Boateng, Heeks and Molla (2011) highlight

43

that in Sub-Saharan Africa, resource constraints and inadequate infrastructure significantly hinder the integration of digital technologies in small enterprises. Similarly, UNCTAD (2019) notes that low digital skills and insufficient access to affordable internet and digital tools remain major impediments to digital transformation for SMEs in developing countries. Furthermore, Gachara and Munjuri (2019) emphasize that many youth-led enterprises in Kenya operate in informal settings with minimal exposure to technology, leading to slow adoption and reluctance to shift from traditional business practices.

In response to the foregoing barriers, MSMEs participating in the pilot expressed a strong need for targeted support mechanisms. These include access to digital skills training, digital marketing education, affordable and reliable ICT infrastructure, and government-backed financing and incentives to support technological uptake. This aligns with recommendations by World Bank (2020), which underscores the importance of integrated support systems, comprising skills development, financing, and enabling infrastructure, to accelerate digital transformation among SMEs in emerging economies. The pilot thus affirmed the critical role of a multifaceted support strategy in enhancing digital maturity and driving sustainable innovation among youth-led MSMEs in Kenya and the broader Sub-Saharan African context.

Regulatory approval for the study has been successfully secured in Kenya through the National Commission for Science, Technology and Innovation (NACOSTI), enabling full-scale data collection to proceed in the country. Ethical clearance applications for South Africa and Nigeria have been submitted and are currently under review by the respective Universities research and ethics authorities. Media engagement on the project initiatives has been ongoing.

Next Targets The next phase of the study involves data collection across Kenya, South Africa, and Nigeria to deepen the understanding of digital transformation among youth-led MSMEs. This will include the identification and analysis of case studies in each country, aimed at uncovering best practices, sector-specific innovations, and the persistent challenges that enterprises face in adopting digital technologies. Drawing on these insights, the study will develop a region specific Digital Maturity Framework tailored to the unique contexts of Sub-Saharan Africa. The research findings will be disseminated through policy briefs, academic conferences and book chapters. In addition, a position paper will be finalized and published to support ongoing policy dialogue, offering evidence-based recommendations for enhancing digital transformation strategies targeting MSMEs in the region. The targeted milestones are presented in the diagram below.

44

Peer-Reviewed Publications Resulting from this Project ●

A conference Paper on Digital Maturity Frameworks for MSMEs has been accepted for presentation during the fourth International Adaptive and Sustainable Science, Engineering and Technology (ASSET) conference, 2025 (8th to 10th July, 2025).

●

A book proposal on digital transformation and MSME development in Sub-Saharan Africa is currently under review by Palgrave McMillan publishers.

References Boateng, R., Heeks, R., Molla, A., & Hinson, R. 2013. Advancing E-Commerce Beyond Readiness in a Developing Country. E-Commerce for Organizational Development and Competitive Advantage, 1.

Gachara, J. and Munjuri, M. 2019. Resource limitations and market access in youth entrepreneurship.

Journal of African Entrepreneurship, 8(2), pp.56–74.

UNCTAD. (2019). Digital Economy Report 2019: Value creation and capture – Implications for developing countries. United Nations Conference on Trade and Development.

World Bank. 2020. The Future of Work in Africa: Harnessing the Potential of Digital Technologies for All. Washington, DC: World Bank.

45

Project FINIA: Financial Inclusion via NovelIntelligence Intelligence andand Alternative data FINIA:Title: Financial Inclusion via Novel Alternative Data Principal Investigator: Prof. Chimwemwe Chipeta, University of the Witwatersrand (PI); CoPrincipal Investigators: Prof. Yudhvir Seetharam, University of the Witwatersrand; Prof. Ganesh Mani Carnegie Mellon University, Africa & Pittsburgh; Prof Patrick McSharry, Carnegie Mellon University, Africa; Prof. Edith Luhanga, Carnegie Mellon University, Africa (Senior Investigator).

Graduate Student: Pierre Ntakirutimana.

Financial Inclusion via Novel Intelligence and Alternative data (FINIA) is a pan-African research initiative that leverages AI, behavioural insights, and alternative data to address persistent gaps in financial inclusion. With a focus on digital financial literacy, crisis resilience, and start-up sustainability, the project has already delivered new evidence from cross-country surveys, machine learning models, and chatbot-based interventions. FINIA is shaping inclusive fintech ecosystems by translating data-driven insights into policy and practice.

1. Project Objectives The FINIA (Financial Inclusion via Novel Intelligence and Alternative data) project is a bold and timely initiative aimed at addressing the persistent challenges of financial exclusion in Sub-Saharan Africa. As digital technologies continue to reshape the financial landscape, FINIA seeks to harness the power of data, artificial intelligence, and behavioural insights to build inclusive, resilient, and forward-looking financial ecosystems. Financial inclusion entails not just access but the ability to use financial services effectively and sustainably. To this end, FINIA is structured around three core objectives: 1) developing robust digital financial literacy strategies (informed by pilot surveys), 2) identifying some of the policy and infrastructural enablers of digital financial services during crises, and 3) leveraging advanced data analytics to predict and mitigate financial distress among African start-ups.

Objective 1: Digital Financial Literacy FINIA focuses on crafting effective, comprehensive digital financial literacy strategies tailored to the African context. Building on the original proposal, this objective has been expanded to include the development of the first continent-wide fintech index and the use of generative AI to support financial education. The objective is motivated by the current employment trends observed on the continent. According to a Mastercard Foundation survey, the gig economy - an economic system where people engage in short-term, often digitally mediated work (gigs) - is growing 20\% every year in Africa and will employ 80 million workers by 2030.

Other surveys also reveal that the majority of Africans (an estimated 86% in 2020) are engaged in informal employment without social protections. In this context, it is essential that the population have the knowledge and skills to research and use the increasingly digital financial products and services available to ensure their long-term financial security.

Objective 2: Crisis-Resilient Digital Inclusion FINIA investigates how the COVID-19 pandemic, while accelerating digital onboarding and FinTech adoption, also introduced significant financial hardships that negatively impacted digital financial inclusion. This component of the project hypothesizes that pandemic-induced financial stress is associated with reduced access to digital financial services, but that this effect can be mitigated by both individual-level characteristics—such as education, income, age, and employment status—and country-level factors like economic growth, institutional quality, and inflation stability. Drawing on data from 31 African countries, the study employs econometric modeling to examine how these variables interact with pandemic-related 1

46

financial concerns. The findings reveal that youth, urban residents, and individuals in countries with stable inflation are more resilient to the negative effects of financial distress on digital financial inclusion. These insights underscore the need for targeted policy interventions that enhance financial resilience and ensure inclusive access to digital financial services during times of crisis.

Objective 3: Predicting Financial Distress in Start-ups FINIA focuses on identifying the drivers of success for African start-ups and developing predictive models to assess financial distress, with a strong emphasis on explainability and practical application. This work involves a structured process beginning with a literature review to ground the research in existing theory, followed by data collection and preparation, including identifying relevant features and cleaning datasets. The project will then proceed to model development, where various machine learning approaches will be explored and refined. These models will be evaluated not only for their predictive accuracy but also for their interpretability, ensuring that stakeholders can understand and trust the insights generated. The final stages will involve deploying the models, analysing results to identify robust predictors of start-up success or failure, and formulating policy recommendations and risk mitigation strategies based on these findings. This approach aims to support more resilient entrepreneurship ecosystems and inform evidence-based decision-making across the continent.

Figure1. The summary of objectives.

Figure2. The Summary of the project Key Findings.

2

47

2. Recent Progress The first objective of the study revolved around developing recommendations on digital financial literacy strategies and programs based on findings from a preliminary survey. The research team conducted a survey of 300 urban and rural Rwandans aged 18–32 between August and November 2024 to assess digital financial literacy (DFL) levels and opportunities for DFL interventions. Three key findings emerged. Firstly, notable gender disparities exist in financial knowledge and informationseeking behaviours, with male participants scoring higher in financial concepts and using more information sources—a trend also observed globally and linked to factors such as lower confidence and educational access among women. Secondly, while awareness of financial products is high, particularly insurance (reported by 93% of respondents likely due to Rwanda’s widespread Community Health Insurance Scheme), this does not necessarily lead to broader financial adoption. For example, although 83% of participants were familiar with bank cards, only 36.33% actually used them. Thirdly, a gap exists between knowledge of cybersecurity and related actions. A follow-up study involving 10 participants was conducted to assess perceptions of chatbot-delivered mini-DFL interventions. A low-fidelity prototype of the MTN Mobile Money (MoMo) app was developed, featuring a ChatGPT-based bot that offered guidance on selecting an appropriate loan amount. This feature addressed loan literacy, an area where survey participants had scored particularly low on. We found that while chatbots have potential for DFL delivery, their acceptability is limited by the effort required to interact with them and users’ low perceived value of the service. These findings, along with those from the fintech index, will be used to generate recommendations for further research and for implementation of DFL initiatives.

The second objective yields important empirical and methodological insights into the relationship between crisis-induced financial anxiety and digital financial behaviours in Africa. Drawing on data from over 27,000 individuals across 31 countries, the analysis reveals a statistically significant and robust negative association between COVID-19-related financial worry and the adoption of digital financial services. This effect operates through both economic constraints, such as reduced disposable income, and behavioural barriers, including heightened risk aversion and distrust in technology. The study thus contributes to financial inclusion scholarship by elucidating the intertwined economic and psychological pathways through which crises can inhibit digital engagement.

Employing a probit regression framework and rigorous robustness checks, the research further identifies individual-level factors, such as youth status, urban residency, and remittance participation, that moderate the adverse effects of financial stress. Additionally, macroeconomic variables yield unexpected results: while low institutional quality and high inflation are associated with increased digital financial inclusion, economic growth appears to have a negative correlation. These findings suggest that inclusion outcomes are shaped less by macroeconomic advancement per se and more by context-specific behavioural adaptations and policy environments. Collectively, the results underscore the need for tailored policy interventions that address both structural inequalities and behavioural frictions in digital finance uptake.

The third subproject advances our understanding of financial distress by exploring the determinants of financial distress among small and medium-sized enterprises (SMEs) in Africa, employing both traditional econometric methods and advanced machine learning techniques. Using tools such as logistic regression, generalised least squares, mixed-effects logit, and instrumental variable methods alongside models like XGBoost, SVM, and LightGBM, the research uncovers uniquely African dynamics in firm vulnerability. Notably, the study identifies a gender-related paradox: female CEOs 3

48

are associated with increased financial distress, whereas female ownership appears to reduce it, suggesting differing implications of leadership versus investment roles. It also lends empirical support to the controversial "grease-the-wheels" hypothesis, where informal payments (‘gifting’) correlate with reduced distress, implying that, in weak institutional environments, corruption may function as an unofficial mechanism for navigating bureaucratic obstacles. Furthermore, the presence of informal competition and the use of audited financial statements are shown to significantly reduce the likelihood of financial distress.

On the technical front, the study constructs a composite index of financial distress based on self-reported liquidity challenges and revenue losses, drawing on data from the World Bank Enterprise Survey.

Methodologically, the analysis incorporates advanced data handling techniques, including Multiple Imputation by Chained Equations (MICE) to address missing data and SMOTE to correct class imbalances. Diagnostic tools such as multicollinearity checks ensure model robustness. A comparative regional analysis reveals significant differences between Africa and other parts of the world, most notably, informal competition is associated with reduced distress in Africa but increased distress elsewhere, while the effects of internal financing and unskilled labour vary regionally. The use of instrumental variables further refines the analysis, confirming the significance of macro-level risk indicators such as the World Uncertainty Index and geopolitical risk. These findings not only deepen the understanding of firm-level vulnerability in Africa but also provide a strong empirical basis for policy tailored to local institutional and economic realities.

3. Peer Reviewed Publications/Presentations Peer Reviewed Publications ● No finalized publications (the first working paper below has been submitted to the Investment Analysts Journal and is under review) Working Papers: ● Chipeta, C., Mani, G, McSharry, P, Seetharam, Y, & Luhanga, E. (2024). Covid-19 Financial Worry and Digital Financial Inclusion in Africa. In progress. ● Chipeta, C., Mani., G, McSharry, P, Seetharam, Y, & Luhanga, E. (2024). Predictors of Financial Distress for SMEs in Africa. In progress. ● Nkuruntimana, P., Chipeta, C., Mani, G., McSharry, P, Seetharam, Y., Sowon, K, & Luhanga, E. (2024). Assessing Digital Financial Literacy and Acceptability of Opportunistic Chatbot-Based Interventions Among Rwandan Youth, In progress.

Conference Presentations: ● Chipeta, C., Mani., G, McSharry, P, Seetharam, Y, & Luhanga, E. (2024). Predictors of Financial Distress for SMEs in Africa. Wits Global Fintech Conference, 14 November 2024. ● Nkuruntimana, P., Chipeta, C., Mani, G., McSharry, P, Seetharam, Y, & Luhanga, E. (2024). Measuring Digital Financial Inclusion in Africa: A New Fintech Index. Wits Global Fintech Conference, 15 November 2024. ● Chipeta, C., Mani., G, McSharry, P, Seetharam, Y, & Luhanga, E. (2024). Covid-19 Financial Worry and Digital Financial Inclusion in Africa. International Academy of Business and Economics Conference, Las Vegas Nevada, 14 December 2024.

Invited Talks: ● Chipeta, C., Mani., G, McSharry, P, Seetharam, Y, & Luhanga, E. (2024). Fintech 4

49

●

Revolution in Africa. AIxHeart Conference, 30 September 2024.

Chipeta, C., Mani., G, McSharry, P, Seetharam, Y, & Luhanga, E. (2024). Fintech Revolution in Africa, Challenges and Opportunities, Katz Business School, University of Pittsburgh, United States of America, 2 December 2024.

Events Organised and Attended: ● Organised: 1st Wits Global Fintech Conference. Wits Fintech Hub, University of the Witwatersrand. 14 to 15 November 2024. ● Attended: The Singapore Fintech Week, 25 to 28 February 2025.

4. Next Targets Significant progress has been made across several strategic initiatives aimed at advancing financial innovation and inclusive digital transformation across the African continent. One of the flagship activities is the finalisation of preparations for a high-level workshop titled "AI, Data and Africa – Charting a Course for Inclusive Growth," to be hosted in the second quarter of 2026. Parallel to these efforts, work is underway to roll out the Annual Africa Fintech Index, in conjunction with the Wits Fintech Hub, a continental benchmarking tool that captures the maturity, resilience, and policy readiness of fintech ecosystems across African economies. The Index will serve as a critical evidence base for governments, investors, researchers and innovators, enabling them to identify gaps, measure progress, and direct resources more strategically. Additionally, the initiative is making strides in identifying drivers of success for start-ups, with a focus on understanding the interplay between innovation, resilience, and the policy environment. This work includes mapping the challenges faced by early-stage startups including fintech firms, analysing risk mitigation strategies that enhance sustainability, and proposing evidence-based policy interventions that can catalyse scale and impact. Insights from this work will feed directly into capacity-building toolkits and future policy dialogues aimed at creating enabling environments for high-impact start-ups across Africa.

Finally, three academic manuscripts are in the pipeline for publication.The first manuscript, aligned with Objective 1, presents the digital financial literacy of Rwandans aged 18-32 years and presents a discussion on focal areas for digital financial literacy interventions. The second manuscript, aligned with Objective 2, explores the structural and behavioural determinants of digital financial inclusion in the context of pandemic-related financial stress. It provides a rigorous empirical analysis using data from 31 African countries, contributing novel insights into the economic and psychological channels that influence financial behaviour during crises. The third manuscript, aligned with Objective 3, examines firm-level predictors of financial distress in African SMEs, integrating machine learning and econometric approaches to derive actionable insights for both policy and practice. Together, these publications advance the evidence base on financial resilience and innovation in Africa and are expected to make meaningful contributions to the global development and fintech policy discourse.

5

50

Fully Integrated Label-Free Virus/Parasite Detection System from Blood-Serum Samples based on Electrical Features from Microscale Electronic Sensors and an AI Software Yehea Ismail (PI), Conrad Tucker (Co-PI), Reda Abdelbaset (Co-PI) Objectives: The proposed innovation focuses on developing a portable, label-free biosensing system using electronic microfluidic technology to detect key biomarkers (e.g., AFP for HCC), viruses (e.g., HCV), and parasites (e.g., Malaria) in blood serum. It employs impedance detection and machine learning, offering a cost-effective, rapid, and accurate alternative to traditional methods, particularly suited for remote and point-of-care settings. Aligned with the African Union Digital Transformation Strategy 2020–2030, the system aims to improve healthcare delivery and drive positive social and economic impact across Africa. The proposed EIS biochip features a developed spiral sensor with high sensitivity for the detection of very small analytes in the nanoscale. Figure 1 depicts the workflow of the proposed platform that begins with serum extraction, impedance sensing, and result processing, facilitated by a built-in GUI. Achievements: The impedance measurement is implemented using a novel spiral gold impedance sensor, reference resistor, an impedance analyzer, and a user interface illustrated in Fig. 1. The impedance analyzer (Analogue Discovery's signal generator and two oscilloscopes) is used to apply and measure electrical signals within a specific range of frequencies and amplitudes. The user interface manages the impedance analyzer, processes the measured data, and records it for the subsequent classification stage.

Finally, the exported database is classified utilizing machine learning (ML). Figure 1 illustrates the experimental setup, which comprises the fabricated impedance

51

sensor, the Analog Discovery 2 (an analog analyzer), and a personal computer used for feature extraction and machine learning classification.

Sample Release 2 μL

Impedance Analyzer 0.9 mm

Feature Extraction ML Classification

Spiral Gold Sensor 1mm Figure 1: The developed platform for the whole system

A total of 280 clinical serum samples were collected, comprising 140 from healthy individuals and 140 from patients with nonalcoholic fatty liver disease (NAFLD), including 29 with early-stage NAFLD and 111 with advanced-stage NAFLD, specifically nonalcoholic steatohepatitis (NASH). NAFLD is an emerging global health issue, ranging from simple hepatic steatosis to more severe forms such as NASH and liver fibrosis. To enhance the accuracy in distinguishing among the three groups, impedance measurements are conducted within a narrow frequency range of 10 to 100 Hz, as illustrated in Fig.2. Additionally, a Python-based graphical user interface (GUI) based on Windows for tablets and personal computers was developed to control the impedance analyzer and systematically store the results in a structured database for streamlined analysis in subsequent stages. A neural network model is employed to classify experimentally measured dielectric properties from a dataset comprising 140 control samples, 29 early-stage NAFLD samples, and 111 advanced-stage NAFLD samples. The results validate the system’s capability,

52

achieving 90.6% accuracy and an F1 score of 0.906 in distinguishing between healthy individuals and patients with early and advanced stages of NAFLD, as shown in Fig.3.

Figure 2: The impedance magnitude of all cases versus frequency around 10 Hz

Figure 3: Accuracy, F1 Score, Recall, and Precision over iterative relabeling and retraining

53

Next steps To achieve optimal performance for the proposed biochip, electronic components such as microcontrollers, waveform generators, and analog-to-digital converters will be carefully selected and customized. These elements are then integrated into an integrated system, specifically tailored to meet the functional and performance requirements of the biosensing application.

Peer-Reviewed Publications Resulting from this Project [1] “Detection of Nonalcoholic Fatty Liver Disease (NAFLD) using Impedance Spectroscopy and Machine Learning algorithm.” It is planned to be submitted before the PI meeting, Reda Abdelbaset, Ananyananda, Conrad Tucker, Yehea Ismail.

[2] Abdelbaset R, Ismail Y. Novel Adjustable 3D Electrokinetic Microelectrode for Blood-Formed Elements Separation. In2024 6th Novel Intelligent and Leading Emerging Sciences Conference (NILES) 2024 Oct 19 (pp. 341-344).

IEEE. [3] Abdelbaset R, Shawky SM, Abdullah MA, Morsy OE, Yahia YA, Ghallab YH, Matboli M, Ismail Y. A new label free spiral sensor using impedance spectroscopy to characterize hepatocellular carcinoma in tissue and serum samples. Scientific Reports. 2024 Jun 7;14(1):13155.

[4] Abdelbaset, R.; Morsy, O.E.; Hossam Eldin, M.; Shawky, S.M.; Ghallab, Y.H.; Ismail, Y. A Novel Microelectrode Based on Joule Heating and Impedance Spectroscopy for Inducing and Monitoring the Aggregation of HCV-specific Probes. Sensors 2025, 25, x.

[5] “Development of a CMOS-Based Microelectrode for DNA Heating and Characterization Using Joule Heating and Impedance Spectroscopy “ accepted conference paper at the 2025 IEEE International Conference on Electro Information Technology hosted by Valparaiso University, Valparaiso, USA, May 29-31, 2025.

[6] “On-chip Flow Cytometry Utilizing the Di-electrophoresis and Ring Oscillator-based Capacitive Sensor” submitted and under review at IEEE Access [7] “Novel Spiral Capacitive Sensor based on Ring Oscillator for Characterizing Biological Cells” submitted and under review at MethodsX

54

Improving Digital Education and Learning Innovation in the South African and Rwandan Teacher Education Systems: Towards Bridging a Digital Knowledge Divide in Africa Innocent Twagilimana (PI), Epimaque Niyibizi, Irenee Ndayambaje, Olivier Habimana University of Rwanda ∙ College of Education Juliet Perumal (PI), Emmanuel Ojo, Reuben Dlamini, Leketi Makalela University of the Witwatersrand

Introduction This research project aims to establish a digital knowledge creation ecosystem in South Africa and Rwanda through the Digital Education and Learning Lab (DELab). Participants include inservice and pre-service educators, as well as academic staff from the University of the Witwatersrand (Wits) and the University of Rwanda (UR). The DELab will function as an inclusive digital hub, designed to promote the adoption of innovative digital pedagogies and foster collaboration between the two partner institutions, with a focus on under-resourced rural schools. Employing a mixed-methods approach, the study is structured across three phases: Phase 1 (baseline and development), Phase 2 (application), and Phase 3 (post-application and evaluation). By leveraging information and communication technologies, the project endeavors to bridge the digital knowledge divide in both countries and provide a replicable model for broader implementation across the African continent.

Background The urgency to transform African education systems through digital innovation has become increasingly apparent in light of global technological advancement. While Western nations have rapidly adopted digital tools to enhance educational quality, Sub-Saharan Africa continues to experience a persistent digital divide that threatens the realization of key Sustainable Development Goals (SDGs) and the African Union’s Agenda 2063, which positions digital transformation as central to inclusive development. Research consistently underscores the potential of inclusive digital transformation in teacher education to expand access to contemporary educational tools and practices, thereby enhancing teaching quality and student outcomes. Digital technologies in teacher education empower educators to design interactive lesson plans, engage in collaborative learning, and offer personalized feedback, all of which improve student achievement and teacher retention. These benefits align closely with SDG Goals 4 (Quality Education), 10 (Reduced Inequality), and 17 (Partnerships for the Goals). Nevertheless, there remains a surprising lack of multi-institutional research collaborations in Africa that support scalable and transformative digital knowledge production.

This project addresses that gap by establishing a digital education and learning laboratory that serves both as a model for professional teacher development and a research-led platform for piloting best practices. Despite differing political histories, South Africa and Rwanda represent compelling case studies of post-conflict societies that have pursued inclusive and equitable education reforms. Both countries have launched initiatives to enhance teacher education, grappled with teacher shortages, and increasingly integrated technology into professional development. However, the absence of a shared digital knowledge platform continues to hinder broader systemic transformation. This initiative seeks to remedy that by fostering sustained collaboration and knowledge exchange.

55

Objectives The project’s key objectives are to: • Promote scalable use of innovative technologies in classrooms in South Africa and Rwanda to reduce disparities in digital access.

• Increase the participation of female students in Science, Technology, Engineering, Arts, and Mathematics (STEAM), with a particular focus on coding and robotics.

• Enhance the enrollment of technologically proficient pre-service and in-service teachers, especially from marginalized communities with limited digital access.

• Generate contextually relevant digital knowledge by triangulating data from both qualitative and quantitative research methodologies.

• Build the research capacity of early-career scholars in both countries in the field of educational technology.

• Establish a transnational network of education researchers to support collaborative knowledge sharing and innovation in digital pedagogy.

Recent Progress The Wits–UR partnership has yielded several notable achievements. Three draft capacitation modules, viz., Module 1: Interactive and Innovative Digital Learning Approaches, Module 2: Digital Support for Research and Innovation, and Module 3: AISupported Education have been uploaded to the DELab platform. These modules aim to strengthen the digital fluency of academic staff and postgraduate students.

The project has also provided significant support to postgraduate scholars across Ph.D., master’s, honors, and postdoctoral levels, helping to cultivate a new cohort of digitally engaged, research-oriented academics within the AFRETEC network.

AFRETEC's collaborative momentum has been reinforced through high-impact engagements. Two South African team members served as keynote speakers at the 3rd International Conference on Reshaping Education (14–16 May 2025), hosted by UR's College of Education. A three-day writing retreat on bibliometric analysis convened researchers from Wits and UR, resulting in draft manuscripts currently under review.

In addition, team members have contributed to international virtual conferences (ICERI, INTED, EDULEARN) and participated in in-person events in Ghana and Rwanda, further strengthening regional academic networks. To date, the project has yielded numerous peerreviewed conference proceedings and publications, many co-authored by postgraduate students. These outputs reflect the project's dedication to fostering academic excellence, mentorship, and collaborative research in digital education.

Next Targets • The next phase of the project focuses on the following priorities: • Completing data collection and analysis in both South Africa and Rwanda.

• Producing localized digital knowledge derived from empirical and desktop research.

• Disseminating findings to educators and stakeholders via publications and public platforms.

• Supporting the adoption of innovative, digital teaching practices.

• Informing educational policy and practice based on research outcomes.

• Expanding capacity-building initiatives through (training educators in the use of DELab resources and digital technologies, delivering workshops on interactive digital learning methods - including AI-supported education, offering targeted training for pre-service and in-service teachers to address instructional gaps, empowering emerging researchers to conduct independent studies in digital education, etc.).

56

Peer-Reviewed Publications Resulting from this Project Journal Articles Published: 1.

Hidayat, H., Anwar, M., Harmanto, D., Dewi, F. K., Orji, C. T. & Mohd Isa, M. R. (2024). Two decades of project-based learning in engineering education: A 21-year meta-analysis. TEM Journal, 13(4). [Scopus indexed] 2. Hidayat, H., Harmanto, D., Orji, C. T. & Anwar, M. (2024). The implementation and empirical analysis of Android learning application toward performance among students in electronics engineering education.

JOIV: International Journal on Informatics Visualization, 8(3), 1154–1161. [Scopus indexed] 3. Hidayat, H., Ardi, Z., Ahlunnazak, A. I., Harmanto, D., Orji, C. T. & Isa, M. R. M. (2024). Determining the influence of digital literacy on learning personal competence: The moderating role of fear of missing out.

European Journal of Educational Research, 13(4), 1775–1790. [Scopus indexed] 4. Obi, J. N. & Ojo, E. (2025). Enhancing employability through vocational and technical skill development among youths and adults in Nigeria. African Journal of Applied Research, 11(2), 478–495. https://doi.org/10.26437/ajar.v11i2.1048 5. Ojo, E. (2024). Education in the eye of the storm: A bibliometric review of research on global crises and their impact on Southern African education (2000–2024). Southern African Review of Education, 29(1), 15–40. [DHET accredited] 6. Orji, C. T. & Herachwati, N. (2024). Career transition and mentorship nexus: Unmasking the mediating role of career adaptability. Higher Education, Skills and Work-Based Learning, 15(7), 82–95. [Scopus indexed] 7. Orji, C. T., Perumal, J. & Ojo, E. (2024). From being a motivated to motivated: Evidence of the efficacy of problem-based learning in practical skills training. International Journal of Home Economics, Hospitality and Allied Research, 3(1), 162–172.

8. Orji, C. T. & Perumal, J. (2025). Enablers of student's satisfaction with work placement learning towards school-to-work transition: A case of Nigeria. African Journal of Applied Research, 11(2), 513–529.

9. Perumal, J. & Ojo, E. (2025). Bibliometric analysis of women's leadership in African universities during disruptive times. African Journal of Applied Research, 11(2), 496–512.

10. Makda, F. (2025). Digital education: Mapping the landscape of virtual teaching in higher education – a bibliometric review. Education and Information Technologies, 30(2), 2547–2575. [Scopus indexed] Book Chapters Published: 1.

2.

3.

Orji, C. T. (2025). Employability skills as a foundation for success: An exploration of youth experiences in skills acquisition programs. In Career Coaching and Employability Skills Acquisition (pp. 195–230).

IGI Global.

Obi, J. N. & Ojo, E. (2025). Climate justice, renewable energy, and food security in Sub-Saharan Africa: A bibliometric analysis of emerging research trends. In Onyeneke, R. U., Emenekwe, C. C. & Nwajiuba, C. U. (Eds.), Energy Transition, Climate Action and Sustainable Agriculture (pp. 303–328). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-83165-2_16 Obi, J. N. & Ojo, E. (2025). A scoping review on exploring job search strategies for unemployed youth: Implications for the agro sector in Sub-Saharan Africa. In Otu, M. S. & Sefotho, M. M. (Eds.), Career Coaching and Employability Skills Acquisition (pp. 1–30). IGI Global. https://doi.org/10.4018/979-83693-4014-1.ch001

Conference Proceedings Published: 1.

2.

3.

4.

5.

Dlamini, R., Ojo, E., Makalela, L. & Perumal, J., 2024. Comparative analysis of the status of ICT implementation and usage in education: The case of Rwanda and South Africa. In: EDULEARN24 Proceedings, 16th International Conference on Education and New Learning Technologies. Palma, Spain, 1–3 July 2024, pp. 9752–9760. IATED. https://doi.org/10.21125/edulearn.2024.2349 Flowers, B. & Tanner, M., 2024. Exploring the digital readiness of underprivileged secondary schools in South Africa. In: IFIP Advances in Information and Communication Technology, Vol. 708, pp. 328–341.

Presented at the 18th IFIP WG 9.4 ICT4D Conference, Cape Town, South Africa, 20–22 May 2024.

Springer. https://doi.org/10.1007/978-3-031-66982-8_23 Hassan, M.N., Marpuah, S. & Nomawati, S., 2023. The difference in ICT proficiency among secondary school Islamic education teachers based on demographic characteristics. In: AIP Conference Proceedings, Vol. 2827, Article 030034. AIP Publishing. https://doi.org/10.1063/5.0164660 Khunou, B. & Ojo, E., 2024. A bibliometric analysis of culturally responsive teaching in pre-university education: Trends and impact in the South African context. In: ICERI2024 Proceedings, pp. 8104–8113.

IATED. Kimuli, J. & Ojo, E., 2024. Navigating success in distance education: A mixed-methods study of highachieving students at Makerere University, Uganda. In: ICERI2024 Proceedings, pp. 9633–9639.

IATED.

57

6.

Kimuli, J. & Ojo, E., 2025. The role of technology in supporting high achievement in distance education: A uses and gratifications expectancy perspective. In: INTED2025 Proceedings, pp. 6756–6762. IATED. 7. Le Roux, J. & Perumal, J., 2024. Aspiring to lead for change: The career aspiration experiences of South African women education leaders. In: EDULEARN24 Proceedings, pp. 9871–9880. IATED. 8. Le Roux, J. & Perumal, J., 2024. Being of service in an educational leadership terrain, laden with challenges–A case of servant leadership. In: EDULEARN24 Proceedings, pp. 9804–9814. IATED. 9. Nkomo, S. & Matli, W., 2022. Adoption of mobile applications to improve the reading habits of rural readers in Southern Africa's secondary schools. In: 2022 IEEE 28th International Conference on Engineering, Technology and Innovation (ICE/ITMC) and 31st IAMOT Joint Conference. IEEE. https://doi.org/10.1109/ICE/ITMC-IAMOT55089.2022.10033311 10. Nyembe, B. & Ojo, E., 2024. Mapping the education-work-graduate employability nexus in South Africa: A bibliometric analysis of research trends (2000–2023). In: ICERI2024 Proceedings, pp. 8095–8103.

IATED. 11. Ojo, E., Dlamini, R., Makalela, L. & Perumal, J., 2024. Integrating technology in teacher education: A bibliometric insight into innovations in sub-Saharan Africa. In: EDULEARN24 Proceedings, p. 9681.

IATED. 12. Ojo, E., Perumal, J., Makalela, L. & Dlamini, R., 2024. Decade-long evolution of digital learning labs in Africa: A comprehensive bibliometric analysis. In: EDULEARN24 Proceedings, pp. 9668–9674. IATED. 13. Orji, C.T. & Perumal, J., 2024. Impact of technology-enhanced problem-based experience on students' intrinsic motivations, ability beliefs, engagement and practical skills outcomes. In: EDULEARN24 Proceedings, pp. 8496–8504. IATED. 14. Orji, C.T. & Perumal, J., 2024. Technology-enhanced learning usage and engagement in practical skills: Does students' e-learning self-efficacy matter? In: EDULEARN24 Proceedings, pp. 8487–8495. IATED. 15. Orji, C.T. & Perumal, J., 2024. Utilization and impact of technology-enhanced learning approaches in TVET during disruptive times: A systematic review. In: ICERI2024 Proceedings, pp. 10172–10181.

IATED. 16. Orji, C.T. & Perumal, J., 2025. Development needs analysis of digital technology in STEM education: Insights from rural and under-resourced schools. In: INTED2025 Proceedings, pp. 7494–7502. IATED. 17. Orji, C.T. & Perumal, J., 2025. Exploring the potential of synchronous e-tutoring for practical STEM education. In: INTED2025 Proceedings, pp. 7486–7493. IATED. 18. Perumal, J., Ojo, E., Makalela, L. & Dlamini, R., 2024. A scoping review of research on digital literacy skills and competencies among pre-service teachers, in-service teachers, and teacher educators in Africa. In: EDULEARN24 Proceedings, pp. 9769–9774. IATED. 19. Perumal, J. & Orji, C.T., 2024. Digital transformation of TVET: A bibliometric examination of innovations in sub-Saharan Africa. In: ICERI2024 Proceedings, pp. 10265–10272. IATED. 20. Tenywa, S. & Ojo, E., 2025. Enhancing educational leadership through technology: A scoping review of the impact and strategies for effective integration in South African schools. In: INTED2025 Proceedings, pp. 6731–6736. IATED. Manuscripts Under Review: 1. Twagilimana, I., Perumal, J., Habimana, O., Dlamini, R., Ndayambaje, I., Ojo, E., Niyibizi, E., Makalela, L. & Nsabayezu, E. Exploring ICT Integration Among Rwandan Pre-Service Teachers: Attitudes, Barriers, and Strategies (Corresponding author: Twagilimana) 2. Ndayambaje, I., Perumal, J., Habimana, O., Dlamini, R., Twagilimana, I., Ojo, E., Niyibizi, E., Makalela, L. & Nsabayezu, E. ICT Integration in Rwandan Schools: Addressing Gaps in Access, Proficiency, and Utilization among In-Service Teachers (Corresponding author: Ndayambaje) Postgraduate students funded by Afretec Project: Wits: Postdoctoral Fellow: Tobias Orji (A0083620): Improving digital education and learning, technological innovations in education, technology-enhanced learning, innovative pedagogy, problem-based learning, career transition, and employability UR: Doctoral Candidate: Ezechiel Nsabayezu: Design and Implementation of a Multimedia-Supported Flipped Classroom Model for Organic Chemistry Instruction in Selected Rwandan Secondary Schools.

Wits Master’s Candidate: Brenda Chirwa (passed with distinction)

58

Conclusion The AFRETEC Project has laid a strong foundation for digital innovation in teacher education through strategic collaboration between South African and Rwandan institutions. The establishment of the DELab and development of targeted training modules have enhanced digital competencies among educators and students. Notable scholarly contributions, including peer-reviewed articles, conference presentations, and book chapters, demonstrate the project's academic impact. Support for postgraduate and early career researchers has strengthened research capacity and fostered a new generation of digitally proficient scholars.

Cross-border knowledge exchange and engagement in global academic platforms have broadened the project’s influence. Looking ahead, the project is well-positioned to inform educational policy, promote inclusive digital practices, and serve as a model for similar initiatives across the continent.

59

Leveraging Additivemanufacturing Manufacturing totoImprove Access Leveraging additive improve access to to Quality Prosthetic, and Orthotic Services quality prosthetic, and orthotic services in developing in Developing Countries countries Moise Busogi Conrad Tucker Jean Bosco Byiringiro Nicholas Okumu Bernard Munyazikwiye

1

Carnegie Mellon University Carnegie Mellon University Dedan Kimathi University of Technology University of Nairobi/Kenyatta National Hospital University of Rwanda

Project Objectives

According to the World Health Organization (WHO), only one in ten people in need can access assistive products, including prostheses, due to their high cost, lack of awareness, unavailability of trained personnel, policy, and financing issues [1]. One significant aspect of prosthetics aggravating this limitation is the need for customization. For prosthetics to be e↵ective, they must be customized to each patient’s unique anatomy and functional needs [2, 3]. Demand for customization is especially highlighted by young patients undergoing physical development and early-stage patients whose residual limb shape is changing [4–6]. An analog method is widely adopted to produce customized prosthetics in resource-constrained communities. In this method, professionals manually shape thermoforming materials to fit the physical model of the residual limb cast from the plaster mold [7]. Although it is widely used, it faces several limitations. Therefore, digital technologies and additive manufacturing have been introduced to overcome these limitations [7]. However, the digital approaches have not been widely implemented in resource-constrained communities for multiple reasons; including access to the required medical equipment such as MRI and 3D scanners, which are inaccessible in some communities and require extensive training for their use [8–10]. To this end, this project seeks to leverage advances in emerging technologies to accelerate access to prosthetic and orthotic services. It integrates elements of artificial intelligence and additive manufacturing to ensure cost-e↵ective, speedy, and high-quality service delivery. The objectives of the project can be broken down as follows: i. 3D Limb Reconstruction from 2D Smartphone Image Data • Develop an AI-based model to generate 3D limb geometry from 2D smartphone images or video.

• Optimize capture protocols (e.g., lighting, angles, backgrounds) to improve reconstruction quality with minimal user training. ii. Data-Driven Prosthetic Socket Design • Build an automated pipeline that incorporates data from 3D limb models to inform prosthetic socket design. iii. Mechanical Simulation and Design Optimization • Develop an optimization framework that maximizes user comfort and maintains the biomechanical functionality of prosthetic devices.

• Optimize printing parameters to ensure that the physical model is aligned with the desired prosthetic design. iv. Clinical Assessment and Validation • Conduct comparative trials to assess fit, comfort, and functionality.

• Use standardized clinical metrics (e.g., Socket Comfort Score, 6MWT, pressure mapping) for validation.

• Establish feedback loops from patients and clinicians to iteratively refine the the pipeline.

• Establish strategic collaborations with qualified partners to ensure e↵ective deployment and longterm success.

60

2

Recent Progress

2.1

Digital (3D) Reconstruction

This work enables 3D reconstruction of a residual limb using only a smartphone video, making the process accessible with minimal equipment dependency, without sacrificing accuracy. The general process is shown in Figure 1.

2.1.1

Video Recording

A video of the residual limb is recorded by rotating a smartphone once around it for 20-30 seconds, capturing both top and side views. This duration provides an optimal balance between mesh quality and processing time. The video path ensures sufficient visual coverage while avoiding motion blur.

2.1.2

Reconstruction Pipeline

Frames are sampled from the video at three-frame intervals.

Backgrounds are automatically removed using the RemBG package [11], ensuring the mesh reconstruction focuses solely on the limb. Neuralangelo [12] then reconstructs the 3D mesh from these cleaned images. Post-processing, automated through CAD scripts, removes visual artifacts, reduces surface noise by iterative Laplacian smoothing, and reorients the mesh to a centered and upright position [8].

2.1.3

Mesh Evaluation

Mesh quality is evaluated by computing the average deviation between the original and reconstructed meshes. Equal numbers of points are randomly sampled on both and nearestpoint distances are calculated until convergence. This evaluation follows the principles of the Chamfer-L1 metric, but is applied directly to mesh surfaces rather than point clouds [12–16].

Figure 1: 3D reconstruction of a residual limb using our proposed pipeline

2.2

Customized Prosthetic Fabrication

This process automates the customization and fabrication of a 3D-printable prosthetic device using the virtual model of the patient’s residual limb. The workflow minimizes specialist involvement through scriptable CAD operations and 3D printing.

2.2.1

Socket Generation

A custom socket is generated by subtracting the residual limb model from an initial socket design using a Boolean di↵erence operation. This creates a cavity that matches the anatomy of the patient. A tolerance o↵set is applied to ensure comfort and wearability. An example of a residual arm and its corresponding socket is shown in Figure 2

61

2.2.2

Socket Scaling

To ensure the socket is both functional and efficient: • The initial socket is iteratively scaled to find the minimum size that fully encloses the residual limb.

• This minimum scale ensures that the socket is not oversized or damaged during subtraction.

• A small scale increment is added post-fit to provide adequate material thickness for 3D printing.

2.2.3

Device Adjustment and Fabrication

Once the socket scale is fixed, other prosthetic parts are resized to match the socket’s dimensions; followed by various optimization as explained below.

Figure 2: a) The generic model of the residual arm (b) The generated socket from the generic model of the residual arm

2.3

Mechanical Functionality Prediction and Optimization

As part of the proposed pipeline, mechanical functionality of the customized prosthetic is predicted using parametric modeling, finite element analysis (FEA), and machine learning. This component ensures that the 3D-printed prosthetic meets mechanical requirements before fabrication by integrating material behavior and print settings into the design process.

2.3.1

E↵ective Material Data Modeling

A dataset is generated to represent how filament material properties and 3D printing parameters a↵ect mechanical performance. The dataset includes 10,000 samples created via Latin Hypercube Sampling. Each sample includes: • Inputs: Young’s modulus, tensile yield strength, Poisson’s ratio (material), and layer height (printing).

• Outputs: Directional e↵ective properties—Young’s modulus, shear modulus, yield strength, and Poisson’s ratio.

Repetitive Volumetric Element (RVE) CAD models are generated parametrically based on print settings.

FEA is performed on each RVE to extract anisotropic material properties resulting from the extrusion shape and print layer geometry.

2.3.2

Machile Learning-Based Property Prediction

An XGBoost regression model is trained to predict the e↵ective material properties from the filament data and layer height. The model: • Inputs: Filament material properties and printing layer height.

• Outputs: Directional e↵ective properties for use in FEA. The model is trained with tuned hyperparameters (e.g., eta = 0.05, max depth = 5, 1000 estimators), and evaluated using RMSE and R2 , with 5-fold cross-validation for robustness.

62

2.3.3

FEA-Based Mechanical Evaluation

The predicted e↵ective properties are applied to the full prosthetic CAD model for FEA. The simulation follows ISO-10328 Condition 2 (P4), representing a peak load in the gait cycle for users under 80 kg. Structural failure is determined based on the minimum directional tensile yield strength.

Although FEA is a simulation of a physical test, it provides rapid and cost-e↵ective feedback. This enables iterative refinement of prosthetic designs, materials, and print settings without requiring full fabrication.

2.4

Next Steps

2.4.1

Enhancing Robustness and Generalization Across the Pipeline

We intend to strengthen the robustness and generalizability of the overall pipeline to ensure reliable performance across diverse settings. For reconstruction, we plan to enhance tolerance to variations in lighting, smartphone type, and capturing distance by refining preprocessing and feature extraction methods. We also aim to validate the reconstruction on a broader range of limb geometries and user conditions.

In mechanical prediction, we will expand the dataset to include more diverse printing setups and filament types, enabling the model to generalize across hardware and process variations. We also plan to incorporate features reflecting print quality and inter-layer bonding.

For prosthetic design, our aim is to explore a wider range of parameters, such as infill patterns, wall geometries, and connectors, and test under varied loading conditions. These improvements will make the method more adaptable to di↵erent patients, devices, and environments.

2.4.2

Clinical Trials and Real-world Validation

To transition this research from lab to practice, we plan to initiate clinical trials involving human participants using 3D-printed prosthetic generated with our method. These trials will allow us to evaluate not only mechanical performance but also comfort, fit, usability, and long-term durability in real-life scenarios. We also aim to validate our predictive models against physical tests to identify and address any performance gaps between simulated and actual outcomes. Importantly, we will collaborate with clinicians and patients to gather feedback that can inform the iterative refinement of the method, ensuring that the resulting prosthetics meet both functional and user-centered needs.

2.4.3

Automating the Pipeline and Building Accessible Community Tools

To support broader adoption and scalability, we intend to automate key stages of the pipeline, including 3D reconstruction, FEA setup, and mechanical prediction. Our goal is to reduce the reliance on expert input and enable faster and more consistent generation of prosthetic designs. In addition, we plan to build a suite of userfriendly tools tailored to di↵erent stakeholders: patients, healthcare workers, and 3D printing technicians. These tools will include mobile applications, web-based interfaces, and visual guides to facilitate use by people with varying levels of technical expertise. We are particularly focused on ensuring accessibility in resource-limited settings by optimizing for low-cost devices and minimizing computational requirements.

3

Peer-Reviewed Publications Resulting from this Project i. Lee, J., Nkama, C., Yusuf, H., Maina, J., Ikuzwe, J., Byiringiro, J., ... & Tucker, C. (2025). Accessible Digital Reconstruction and Mechanical Prediction of 3D-Printed Prosthetics. Journal of Mechanical Design, 147(6), 062001. ii. Lee, J., Nkama, C., Yusuf, H., Maina, J., Ikuzwe, J., Byiringiro, J., ... & Tucker, C. (2024, August).

Increasing Accessibility of 3D-Printed Customized Prosthetics in Resource-Constrained Communities. In International Design Engineering Technical Conferences and Computers and Information in Engineering Conference (Vol. 88377, p. V03BT03A024). American Society of Mechanical Engineers. iii. Lee, Junghun, Hadiza Yusuf, and Conrad Tucker. ”Inverse-Prediction of Material Property in Fused Filament Fabrication/Fused Deposition Modeling.” International Design Engineering Technical Conferences and Computers and Information in Engineering Conference. Vol. 88360. American Society of Mechanical Engineers, 2024. iv. Lee, J., Chesang, A., Gichane, M., Busogi, M., Byiringiro, J., & Tucker, C. (2023, August). Reducing the barriers to designing 3d-printable prosthetics in resource-constrained environments. In International

63

Design Engineering Technical Conferences and Computers and Information in Engineering Conference (Vol. 87318). American Society of Mechanical Engineers.

References [1] W. H. Organization, “Who standards for prosthetics and orthotics,” 2017.

[2] L. Amrutsagar, G. Parit, R. Ghyar, and R. Bhallamudi, “Parametric design and hybrid fabrication of above-knee prosthesis,” Indian journal of orthopaedics, vol. 54, pp. 381–390, 2020.

[3] I. Vujaklija and D. Farina, “3d printed upper limb prosthetics,” Expert review of medical devices, vol. 15, no. 7, pp. 505–512, 2018.

[4] M. B. Burn, A. Ta, and G. R. Gogola, “Three-dimensional printing of prosthetic hands for children,” The Journal of hand surgery, vol. 41, no. 5, pp. e103–e109, 2016.

[5] D. E. Krebs, J. E. Edelstein, and M. A. Thornby, “Prosthetic management of children with limb deficiencies,” Physical therapy, vol. 71, no. 12, pp. 920–934, 1991.

[6] A. T. Tantua, J. H. Geertzen, J. J. van den Dungen, J.-K. C. Breek, and P. U. Dijkstra, “Reduction of residual limb volume in people with transtibial amputation,” Journal of rehabilitation research and development, vol. 51, no. 7, pp. 1119–1126, 2014.

[7] R. K. Chen, Y. Jin, J. Wensman, and A. Shih, “Additive manufacturing of custom orthoses and prostheses—a review,” Additive manufacturing, vol. 12, pp. 77–89, 2016.

[8] J. Lee, A. Chesang, M. Gichane, M. Busogi, J. Byiringiro, and C. Tucker, “Reducing the barriers to designing 3d-printable prosthetics in resource-constrained environments,” in International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, vol. 87318. American Society of Mechanical Engineers, 2023, p. V03BT03A020.

[9] T. N. Van, T. Le Thanh, T. N. Van, N. Naprstkova et al., “Smartphone-based data acquisition method for modelling 3d printed arm casts,” Manufacturing Technology Journal, vol. 23, no. 2, pp. 260–267, 2023.

[10] S. Cullen, R. Mackay, A. Mohagheghi, and X. Du, “The use of smartphone photogrammetry to digitise transtibial sockets: Optimisation of method and quantitative evaluation of suitability,” Sensors, vol. 21, no. 24, p. 8405, 2021.

[11] D. Gatis. rembg-singhaman092. rembg-singhaman092/tree/master

[Online].

Available:

https://github.com/mhkim0720/

[12] Z. Li, T. Müller, A. Evans, R. H. Taylor, M.-Y. Liu, and C.-H. Lin, “Neuralangelo: High-fidelity neural surface reconstruction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 8456–8465.

[13] A. Hernandez and E. Lemaire, “A smartphone photogrammetry method for digitizing prosthetic socket interiors,” Prosthetics and orthotics international, vol. 41, no. 2, pp. 210–214, 2017.

[14] N. H. Mahmood and T. Tjahjadi, “3d reconstruction for prosthetic design,” in 2010 Second International Conference on Computer Engineering and Applications, vol. 1. IEEE, 2010, pp. 431–435.

[15] J. Lu, Z. Li, J. Bai, and Q. Yu, “Oriented and directional chamfer distance losses for 3d object reconstruction from a single image,” IEEE Access, vol. 10, pp. 61 631–61 638, 2022.

[16] L. Mescheder, M. Oechsle, M. Niemeyer, S. Nowozin, and A. Geiger, “Occupancy networks: Learning 3d reconstruction in function space,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 4460–4470.

64

Leveraging IoT and Edge-AI for Detecting and Managing Schistosomiasis Moustafa Youssef, Cheikh Ahmadou Bamba Gueye, Benyl Ondeto

Objective This project aims to develop an integrated IoT and Edge-AI framework to mitigate schistosomiasis risk across Africa. By continuously monitoring key physicochemical water parameters (pH, temperature, dissolved oxygen, and electrical conductivity), the system seeks to correlate these indicators with snail population dynamics, which are the primary vectors of schistosomiasis.

An AI model optimized for Edge-IoT platforms enables real-time inference and delivers actionable insights to support proactive and scalable disease management in resource-constrained environments.

Current Progress The project has made substantial progress in developing an advanced water quality prediction system tailored to support schistosomiasis management.

Initially, a BERT-based transformer architecture, adapted from natural language processing, was repurposed for the temporal forecasting of environmental parameters, as illustrated in Figure 1. Leveraging historical time-series measurements (e.g., pH and dissolved oxygen). This system demonstrated robust performance in predicting future trends of physicochemical water variables. Its accuracy significantly outperformed conventional approaches such as Long Short-Term Memory (LSTM) networks [1].

65

Building upon this foundation, the project subsequently addressed two critical challenges: (1) enhancing computational efficiency for deployment in resource-constrained environments and (2) extending predictive capabilities to estimate the population growth dynamics of intermediate host snails. To meet these objectives, the team implemented Fourier Network (FNet), a transformer variant that substitutes self-attention mechanisms with Fourier transforms (see Figure 2). This architectural change significantly reduced computational demands while preserving high forecasting acFigure 1 FNet forecasting model. curacy. Controlled experiments confirmed that FNet, even with equivalent parameter counts, achieved performance comparable to the BERT model across all target variables. Notably, FNet delivered an 80% reduction in training time and a 90% decrease in inference latency per sample, as verified in [2], thereby making it highly suitable for real-time, on-device deployment.

To further improve prediction accuracy and generalization, the system was augmented with comprehensive data augmentation techniques. These enhancements yielded substantial benefits, with normalized Mean Absolute Error (MAE) reductions of up to 72%, ensuring robustness and reliability even in challenging real-world conditions.

Through these efforts, the project has delivered a highly optimized and fieldready solution capable of providing real-time, accurate, and computationally efficient water quality forecasting to inform schistosomiasis risk management strategies in low-resource settings. Ultimately, by enabling timely interventions to control the snail population, this system contributes directly to reducing human exposure to schistosomiasis-infested water sources, thereby supporting healthier communities and mitigating the burden of disease among vulnerable populations.

66

Next Targets Building on our current framework, we are advancing the system through the implementation of a federated learning (FL) architecture, specifically designed for Edge-IoT deployments in resourceconstrained, schistosomiasis-endemic regions (Figure 3). Preliminary results are highly encouraging, demonstrating that the FL approach significantly reduces communication bandwidth compared to conventional centralized learning methods, while maintaining predictive accuracy within 1.21% MAE of centralized benchmarks.

Figure 2 Federated learning-based system architecture

To overcome the challenges posed by non-independent and identically distributed (non-IID) data across edge devices, the system integrates FedProx optimization, which has proven effective in achieving stable convergence where traditional algorithms like FedAvg are limited. In addition to reducing bandwidth requirements, the federated approach provides exceptional resilience in real-world deployments. It maintains robust performance even under noisy sensor conditions, ensuring reliable and uninterrupted operation in dynamic and unpredictable environments. Collectively, these advancements mark a major step forward toward a practical, scalable, and field-ready solution for distributed water quality monitoring. This system ultimately enables more effective and timely schistosomiasis risk assessment and intervention at scale.

Looking ahead, we aim to further enhance the system’s decision-making capabilities by incorporating advanced AI techniques that integrate heterogeneous sensor data, quantify predictive uncertainty, and improve overall robustness. These will be also integrated with the information from the domain expert’s knowledge. The developments will ensure even more reliable and scalable schistosomiasis risk assessment, advancing our goal of

67

safeguarding vulnerable populations through smarter, data-driven public health interventions.

On the hardware side, we plan to update our prototype for the IoT device to include better sensing and communication capabilities for the continuous and remote monitoring of water. These include designing and implementing efficient solutions that cover regions with inefficient or nonexistent network connectivity. Along the same lines, we will investigate the cost, availability, and power tradeoffs of using different communication technologies in the kits. This will allow a more practical and scalable deployment of the kits.

For the sensing part, the knowledge of the domain expert will allow us to select and introduce new sensors to our kits, such as rainfall and salinity, which may lead to better accuracies.

Publications [1] Mohsen, Mohamed, Hamada Rizk, and Moustafa Youssef. "Enhancing Water Quality Predictions with Transformers for Schistosomiasis Management." Proceedings of the 2nd ACM SIGSPATIAL International Workshop on Geo-Privacy and Data Utility for Smart Societies. 2024.

[2] Mohsen, Mohamed, Teegwende Zougmore, Bamba Gueye, Hamada Rizk, and Moustafa Youssef. " Efficient Real-Time Water Monitoring for Schistosomiasis Control with a Fourier Network." Proceedings of the 3rd IEEE MDM International Workshop on Mobile Crowdsensing for Smart Cities 2025.

68

Low-Cost, Accessible Biotechnologies AfricanHair Hair and Dark Colors Low-Cost, Accessible Biotechnologies for for African and Dark SkinSkin Colors Research Team (PIs underlined): CMU-Pittsburgh, USA: Pulkit Grover, PhD, Sossena Wood, PhD, Jana Kainerstorfer, PhD, Jasmine Kwasa, PhD, Marlene Behrmann, PhD CMU-Africa, Kigali, Rwanda: Assane Gueye, PhD The American University in Cairo, Cairo, Egypt: Seif Eldawlatly, PhD University of the Witwatersrand, Johannesburg, South Africa: Ziporah Katz, MD, Rachel Nossel, MD, Yair Katz, MD University of Ghana, Legon, and University of Health and Allied Sciences, Hohoe, Ghana: Elsie Kauffman, PhD, Solomon Ofori-Acquah, MD University of Rwanda, Kigali, Rwanda: Charles Mudenge, MD, MMED 1. Project Objectives: Our team is at the forefront of developing biotechnologies that work for all, and includes biomedical and clinical researchers working with different populations all across Africa. Over the three-year project term, we focused on establishing and sustaining collaborations and collecting data towards making two key and widely used non-invasive technologies more reliable and accurate for adoption in the African population. We worked to ensure initial easy access to healthcare in Rwanda, Ghana, Egypt, and South Africa.

Aim 1: EEG that works with all African hair-types. Electroencephalography (EEG) is a century old noninvasive technology for measuring brain activity by placing electrodes on the surface of the scalp. Our work [Etienne et al.,'20; Kwasa et al., ‘24] demonstrated that current EEG systems do not work well with some African hair-types. EEG is essential when diagnosing and treating many disorders, including epilepsy, stroke, and brain injuries. The proposed work will test and refine solutions, building on our recent solution, to work with all African hair-types; Aim 2: Pulse oximetry (PulseOx) that has no bias with any skin color. PulseOx uses light to measure blood oxygen saturation, but readings are inaccurate with dark skin colors. PulseOx is widely used in monitoring health, e.g., during surgery, diagnosing sleep apnea; and, of recent interest, in making critical decisions on the need for a ventilator in the COVID pandemic. Our team will characterize sources of bias, and develop techniques to remove it.

Aim 3: Learning, training, dissemination, and societal impact. The team aims to better understand these problems in the African context, co-developing, prototyping, and testing solutions. In the process, we will learn from each other as well as produce education modules, including videos, short-courses, brochures, and publications for techniques for addressing biases and best practices for a wide range of hair types and skin colors.

2. Recent Progress: We have made progress at each site in each aim, specifically regarding IRB approvals, data collection, analysis, and publications. Progress varies among each site. In addition, Dr. Grover and Dr. Wood were able to share the work of this project with the NIH BRAIN initiative director during his 2024 visit to CMU-Pittsburgh.

Progress on Aim 1: EEG that works with all African Hair.

Egypt Site: Accessible EEG: Developing a Brain-Computer Interface (BCI) platform to test Sevo electrodes.

In terms of the P300 recognition pipeline that we are developing to test the Sevo electrodes. We have enhanced the pipeline for P300 recognition as follows: ●

Tested in online and offline experiments with two team members, achieving: ○

88% in P300 component recognition.

69

○

95% in character recognition.

We developed a novel deep learning approach that takes both spatial and temporal relationships into consideration to recognize P300 patterns. A paper is now in preparation to be submitted to a top-tier conference summarizing the results obtained using this approach.

We also developed an offline Steady-State Visually Evoked Potentials (SSVEP) speller interface.

We developed the AfroBCI software to be used in testing the Sevo electrodes using P300 and SSVEP experimental paradigms. We have added new features to session management settings to allow better operation of the software and to be more user-friendly as follows: ●

●

Optimized the model by finding the best hyperparameters for feature extraction. The hyperparameters include the duration of the analysis window, the window beginning and ending, and the number of principal components to use in the analysis.

Allowed exporting session files in EDF format.

In addition, we have obtained the Institutional Review Board (IRB) AUC approval to start recording from subjects in order to test the developed system. Finally, in collaboration with CMU team, we have tested various Sevo electrode sizes to determine the best fit for the Unicorn headset. CMU team sent us the designs of the Sevo adaptor to be used with the Unicorn Hybrid Black EEG headset. We customized these designs to fit our needs.

Rwanda Site: To date, the Rwanda team has the printed Sevo electrodes. IRB approval is still pending but the team is ready to enroll participants once approval is received.

Pulse oximetry project: Progress on Aim 2: Pulse Oximetry Without Skin Color Bias. We are conducting a multi-site investigation to identify and mitigate racial bias in pulse oximetry by comparing device readings (SpO₂) against arterial blood gas measurements (SaO₂) across diverse skin tones in Pittsburgh, Ghana, and Rwanda. Our approach combines signal characterization, device prototyping, and field deployment.

Pittsburgh Site: Signal Characterization and Hardware Development.

We have made significant progress in identifying physiological and signal processing contributors to skin-tone-based bias in pulse oximetry. A comprehensive study identifying reduced signal-to-noise ratio

70

(SNR) as a major source of measurement bias in individuals with darker skin was submitted for peer-reviewed publication (Cao et al., 2024, in review), building on earlier work accepted at IEEE EMBC ‘23. In parallel, our recently published work in the Journal of Biomedical Optics (Roy et al., 2024) characterizes the impact of melanin on PPG signal morphology and quantifies mismatch against reference SaO₂ values.

Based on these insights, we designed a multi-wavelength pulse oximeter incorporating red, infrared, and green LEDs (Fig 1). This custom-built device allows raw data acquisition without embedded gain correction, enabling evaluation of intrinsic spectral dependencies on skin tone.

This prototype is being used in laboratory and field settings to validate signal quality and algorithmic correction methods across diverse populations.

Ghana Site: Clinical Data Collection and Bias Quantification The custom pulse oximeter is deployed to compare SpO₂ measurements against arterial SaO₂ collected via blood gas analyzers. Preliminary data (n=36), including those with varying medical backgrounds, reveal a modest (R = 0.08) and statistically non-significant correlation between SaO₂ and the difference in SaO₂–SpO₂ readings. Points are visualized by Monk Skin Tone scale, derived from DSM III melanin index measurements (see Fig 2). Early findings highlight variability potentially attributable to skin tone and signal quality, reinforcing the need for increased participant numbers to reach statistical power.

We collected a dataset (n=57, including both patient and control groups) in Ghana, and we identified a moderate correlation between melanin index and SaO₂–SpO₂ differences. This reinforces our working hypothesis that melanin-dependent light absorption plays a significant role in the bias detected in standard pulse oximeters, particularly in Ghana, where the melanin index exceeds the calibration standards of these oximeters.

Rwanda Site: Device Evaluation Without SaO₂ Reference Although blood gas measurements are not feasible at this site, preliminary testing is underway using a simplified two-wavelength oximeter to examine PPG signal distortion across skin tones.

Data from four participants with SpO₂ <95% has been collected, with a recruitment goal of 30 participants in each of the control and patient groups. These measurements will contribute to understanding pigmentation-induced distortions at the raw signal level in the absence of algorithmic corrections.

Progress on Aim 3: Learning, Dissemination, and Societal Impact A key component of Aim 3 involves co-development and testing of bias-mitigating strategies in local contexts. In collaboration with the Ghana site, we have developed a 3D-printed clip prototype for fNIRS measurements that addresses hair-type-related signal loss. The device

71

enables stable, long-duration recordings in participants with tightly coiled or braided hair—frequent barriers to optical scalp recordings in populations of African descent.

As demonstrated in Figure 2, the clip securely interfaces with braided hair, yielding high-quality, motion-resistant fNIRS signals including distinct cardiac and respiratory pulsations. This result not only supports dissemination of open-source hardware designs but also aligns with our broader mission to enable inclusive neurotechnology across diverse populations.

We are currently translating this work into publicly available educational modules and videos, as well as preparing conference presentations and journal articles that document design rationale, clinical usability, and cross-site reproducibility.

3. Next Targets: Collect more data, test the new hardware for the custom CW-NIRS system, and continue to recruit participants from the hospital at the Rwanda and Ghana sites.

Aim 1: EEG that works with all African Hair. Develop and test electrodes that work for sensing and stimulation (led by Kwasa, Grover).

Aim 2: Pulse Oximetry Without Skin-Color Bias 1. Complete Multi-site Data Collection ● ●

Ghana site: Enroll an additional 64 participants (to reach n≈100), balancing patient and control cohorts, to achieve ≥80% power for detecting a correlation of r ≥ 0.3 between melanin index and SaO₂–SpO₂ difference.

Rwanda site: Finish recruitment for 30 controls and 30 patients (SpO₂ < 95%), yielding a total n = 60 for raw PPG waveform analysis.

2. Advanced Signal & Algorithm Development ● ●

Prototype and test the novel CW-NIRS system: Prototype the input intensity adaptable CW-NIRS system and test the prototype at both the Pittsburgh and Rwanda sites.

Multi-wavelength calibration: Develop and implement an adaptive spectral calibration algorithm that dynamically weights red/IR/green channels to minimize SpO₂ error. Perform 5-fold cross-validation using the combined Ghana + Pittsburgh dataset.

Aim 3: Learning, Dissemination, and Societal Impact Open-Source Hardware & Documentation ● ●

Finalize CAD files, assembly instructions, and bill of materials for the 3D-printed fNIRS hair clip; publish on a public repository (e.g., GitHub) with versioned releases.

Produce a concise “Quick-Start Guide” PDF and an annotated video walkthrough (< 5 minutes) demonstrating clip fabrication and deployment.

Educational Modules & Training Develop two short online modules (each 15–20 minutes): ● ●

Optical Bias in Pulse Oximetry – principles, sources of error, and correction strategies.

Inclusive Neuroimaging Hardware – design considerations for hair-type and skin-tone diversity.

Workshops & Outreach ●

Host one in-person training session in Accra, Ghana (target: 10 engineers/clinicians), covering both pulse-oximeter calibration and fNIRS clip use.

local

72

●

Organize a follow-up webinar for participants in Rwanda, including live Q&A and hands-on troubleshooting of the two-wavelength oximeter prototype.

4. Peer-Reviewed Publications Resulting from this Project: Our team has produced several early publications and conference proceedings to support the work in Aims 1 and 2.

I. ●

● ●

Peer-Reviewed Journal Articles Kwasa, J., Roy, S., Opoku, E., Etienne, A. and Et al., 2024. Pilot Evaluation of Sevo Systems for Epilepsy: Equitable EEG for Coarse, Dense, and Curly Hair. In: 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). [online] IEEE. Roy, S., Cao, J., Berisha, V. and Et al., 2024. Exploring the impact and influence of melanin on frequency-domain near-infrared spectroscopy measurements. Journal of Biomedical Optics, 29(S3), pp.S33310–S33310.

Cao, J., Roy, S., Berisha, V. and Et al., under review. Signal-to-noise ratio as a potential mechanism for racial bias in pulse oximetry. IEEE Transactions on Biomedical Engineering. (Previously presented at IEEE EMBC 2023).

II. Conference Proceedings ●

●

Appiah Anokye, C., Cao, J., Roy, S., Mehta, N. and Grover, P., 2025. Advancing peripheral oxygen saturation (SpO₂) measurement accuracy using biophysics-informed machine learning techniques to mitigate racial disparities. In: CBAS Biennial Science and Development Conference. [conference presentation] CBAS, 2025.

Cao, J., Mehta, N.A., Wu, J., Wood, S., Kainerstorfer, J.M. and Grover, P., 2023. Scaling of algorithmic bias in pulse oximetry with signal-to-noise ratio. In: 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE, pp.1–4.

73

Optimizing a Mobile Application for Enhancing Parents’ Reporting and Prediction of Adverse Effects following Maternal and Child Immunization in Rwanda Aimable MUSAFILI1, MD, PhD, Christine NIYIZAMWIYITIRA2, PhD, Nicole IGIRANEZA ISHIMWE2, Michael MUGISHA1, PT, PhD, Lazare NTIRENGANYA3, Phn, MPH, Emmanuel NKURUNZIZA1, MD, MPH, Augustine MUHIRWA1, RN, MSc, Angelique UWAMAHORO1, OCO 1

College of Medicine and Health Sciences, University of Rwanda; 2Carnegie Mellon University, Africa Campus–Rwanda; 3Pharmacovigilance and Food Safety Monitoring, Rwanda Food and Drugs Authority

1. Project Objectives The overall objective of optimizing digital platform for enhanced parents’ reporting and prediction of Adverse Effects following Immunization in Rwanda is to design a web-based platform and its related mobile application that report adverse effects after immunization. Furthermore, this application is integrated with machine learning algorithm to help predict adverse effects. This platform will be piloted in health facilities to collect live data and provide live predictions.

The platform will be used by parents and guardians to report Adverse Events Following Immunization (AEFI) directly bridge the existing reporting gap.

Additionally, the integration of machine learning algorithms will facilitate the identification of factors influencing AEFI within the Rwandan population, enhancing understanding and response, while also addressing the underreporting of AEFI, this solution will strengthen vaccine safety monitoring and improving public health outcomes in Rwanda.

The specific objectives are (1) to evaluate machine learning models (XGBoost, LightGBM, Random Forest, and SVM) to identify the most effective approach for predicting BCG vaccine adverse effects severity using Rwanda FDA's Excel-format datasets. (2) To implement SMOTE (Synthetic Minority Over-Sampling Technique) to address the class imbalance inherent in adverse event severity data, where moderate and severe side effects are rare. (3) To analyze the performance trade-offs between accuracy and computational efficiency among the selected models, with particular attention to the balance achieved by LightGBM. Finally, (4) To develop a practical predictive tool that Rwanda health professionals can integrate into existing systems to identify patients at higher risk of moderate to severe complications.

2. Recent Progress Currently, were able to train four (4) machine learning models to be able to choose the best performer. The best performer model is integrated into a web-based application. We have used historical data of Bacillus Calmette-Guérin (BCG) a vaccine to prevent tuberculosis (TB) from Rwanda Food and Drugs Authority (RFDA). The available data was for the period of 4 years (2021 to 2025) covering urban and rural areas of the city of Kigali.

We trained four machine learning algorithms (Random Forest, Support Vector Machines, XGBoost, and LightGBM) with Synthetic Minority Over-Sampling Technique (SMOTE) to

74

address extreme class imbalance in BCG’s AEFI data, particularly for moderate cases which represented only 0.1% of the dataset.

Our models achieved remarkable performance, with Random Forest demonstrating 85.68% test accuracy and 94.33% cross-validation accuracy, while LightGBM offered comparable accuracy with significantly reduced computational demands (4.8 seconds vs. 15.3 seconds for Random Forest). SMOTE successfully transformed our training data by increasing moderate cases from just 6 to 3,022 balanced samples. Key determinant factors influencing AEFI severity included adverse event location, patient age, dose number, and specific reaction types.

Figure 1 shows the comparison of the results of the trained machine learning algorithm, and Figure 2 shows the application interface.

Figure 1.Results for CV accuracy, Test accuracy, and computation time/training time comparison

Figure 2. Mobile Application Interface for BCG Vaccine

This research demonstrates the potential of machine learning to enhance vaccine safety monitoring by predicting adverse event severity following BCG vaccination. The model integrated within application shows severity levels with high accuracy of 85.68% o. This application is useful for strengthening Rwanda's immunization program. Although the application is still at the pilot level,

75

the successful implementation of this severity prediction model represents an important step toward more proactive vaccine safety monitoring in resource-constrained settings. With the recently established health intelligence center1 for the Ministry of Health, immunization program will already have something to start with, and that is our application.

3. Next Targets We are currently working on incorporating genetics/immunology data, add more vaccines such as measles-rubella, pentavalent, and other vaccines, integrate our application with community health workers’ application, and holistically integrate with other existing active surveillance systems.

4. Peer-Reviewed Publications Resulting from this Project No publication to report for the moment.

This project has already produced one master’s research project, which will be published as paper, focusing on machine learning-based prediction of AEFI related to BCG vaccine. The manuscript related to this paper is being developed and will be submitted to the SAIEE Africa Research Journal. We are also planning to publish another paper on optimizing caregiver reporting of AEFI through mHealth platform in the Journal of Interventional Epidemiology and Public Health. Two more upcoming publications will look into the contextual factors influencing caregivers’ and healthcare providers’ reporting of AEFI, as well as the prevalence and associated factors of AEFI among children in Rwanda.

1

https://www.moh.gov.rw/news-detail/new-health-intelligence-center-to-drive-real-time-evidence-based-decisions

76

Real-Time Noise-Level Web Mapping through Crowdsourcing: Toward Creating Sustainable Urban Environments David N. Siriba1, Ernest Uwayezu2, Collins M. Mwange1, Christine Musanesa2 Department of Geospatial and Space Technology, University of Nairobi 2College of Science and Technology, University of Rwanda

1

Project Objectives Urban noise pollution is an ever-increasing challenge in African cities, with significant concerns and implications for quality of life and public health. This project explores innovative, citizencentric approaches to urban noise monitoring through the development and deployment of mobile and web-based tools in Nairobi, Kenya and Kigali, Rwanda. Leveraging both the existing Noise Capture app and a custom mobile application tailored to local contexts, the initiative empowers residents to contribute real-time geotagged noise data (Figure 1).

Recent Progress The custom mobile application dubbed Urban Ear integrates user-friendly features for sound recording, visualization, while ensuring compatibility with Android devices commonly used in the region. On the web front, an interactive, cloud-based noise web mapping platform has been developed, enabling visualization of collected data across space and time. The platform supports filtering by time, location and noise thresholds, enhancing its utility for urban planning and policy decision-making.

Figure 1: Citizen-Driven Approach to Urban Noise Reduction

Data collection campaigns, involving university students have been successfully conducted in selected high-traffic and residential zones in both cities. Preliminary analysis (Figure 2) reveals spatial paUerns in noise levels, with notable hotspots near transport corridors, markets, and construction sites. Comparatively, Nairobi records higher average noise levels (68.8dB) than

77

Kigali (63.7 dB); The noise levels also highlight differences in urban form, regulation, and socioenvironmental behaviours.

Figure 2: Data collection in Nairobi and Kigali

The challenges that have been encountered during the study include: firstly, data harmonization between the two apps - but this is being addressed through customized back-end processing.

Secondly, the limitations in smartphone microphone calibration; crowdsourced data from smartphones may be inconsistent due to variability in microphone quality and user behaviour.

This will require calibration of individual phones. In this regard, a statement about the accuracy of the provided noise level data is given to ensure cautious use of the data, even as mechanisms are explored on how to improve the accuracy and reliability of the data.

The third challenge is on the sustainability of participant engagement over time. This will be addressed by creating awareness creation of the impacts of noise pollution through campaigns to encourage communities to participate in data collection. This will ensure adequate spatialtemporal coverage. Finally, crowdsourcing may raise ethical and legal concerns, limiting adoption. To address privacy concerns, the identity of the contributors will be anonymised.

78

Next Steps The next steps (in Figure 3) will include scaling up data collection efforts, refining the web map for enhanced analytics, integrating AI-based cleaning of data and classification/characterisation of noise sources, and engaging city authorities in co-designing noise mitigation policies. The project lays the groundwork for a replicable framework for participatory urban noise monitoring in Sub-Saharan Africa.

Figure 3: Next steps

Peer Reviewed Publications/Presentation None to report for the moment.

References Othman, E.; Cibili´c, I.; Poslonˇcec-Petri´c, V.; Saadallah, D. (2024). Investigating Noise Mapping in Cities to Associate Noise Levels with Sources of Noise Using Crowdsourcing Applications. Urban Sci. 2024, 8, 13.

Picaut, J., Fortin, N., Bocher, E., Petit, G., Aumond, P., & Guillaume, G. (2019). An open-science crowdsourcing approach for producing community noise maps using smartphones. Building and Environment, 148, 20–33. hUps://doi.org/10.1016/j.buildenv.2018.10.049

79

Revolutionising Hospital Oxygen Supply through Revolutionising HospitalWater-Electrolysis: Oxygen Supply through Water-Electrolysis: A Modular, Solar-Powered Solution for Kenyatta National Hospital A Modular, Solar-Powered Solution for Kenyatta National Hospital Dr Davies Rene Segera1 *, Prof Grace Irimu1 , Prof Nageh Allam2 , Prof Rose Alani3 , Prof T. O. Mbuya1 1

University of Nairobi, Kenya

2

American University in Cairo, Egypt

3

University of Lagos, Nigeria

1. Project Objectives Water-electrolysis now stands as the most economically viable route to medical-oxygen supply for East African hospitals. While cryogenic plants require minimum capacities of ⇠ 100 t O2 d 1 and PSA systems plateau at 93–95% purity, PEM and alkaline electrolysers deliver 99.5% O2 with co-production of revenue-generating green H2 . Recent alkaline membrane technology achieved 95% efficiency using inexpensive nickel compounds.

Maggio et al. demonstrated that 1 MW PV-powered systems achieve positive returns when oxygen exceeds 3–4 C kg 1 —a threshold Kenya surpasses with current prices of $19–58 per 6.8m3 cylinder [1].

Kenyatta National Hospital (KNH) currently spends KES 24 million monthly on trucked liquid oxygen while consuming 4,000–8,000 litres daily—representing both financial burden and supply-chain vulnerability [3]. Our Afretec-funded initiative aims to develop a scalable PEM electrolyser technology platform through systematic research, targeting a multi-stack system capable of meeting 15–25% of KNH’s oxygen demand while generating saleable green hydrogen, aligning with Kenya’s National Green Hydrogen Strategy [4]. The program employs modular scaling: characterizing a 3-cell PEM fuel cell stack, developing a sophisticated IoT monitoring system, and integrating these into a larger multi-stack system comprising multiple 4 cm2 PEM units powered by Kenya’s exceptional solar resources (>5 kWh/m2 /day). A modular header-based IoT architecture features Arduino-based microcontroller with SIM800L GSM module, DS3231 RTC, and standardized sensor connectors supporting I2 C, UART, and analog protocols. This establishes a replicable foundation for oxygen production across Kenya’s 47 county hospitals, advancing SDGs 3, 7, 9, and 13.

2. Recent Progress Since January 2025, the consortium achieved critical milestones. KNH executive board formally endorsed the program in April, allocating laboratory space and seconding two biomedical engineers for technology transfer [2]. Technical risk mitigation includes ISO 10083:2006 compliance and KEBS pre-certification consultations.

We procured and began characterizing a 3-cell PEM fuel cell stack (4 cm2 active area) from CNL Energy Co., featuring platinum-plated titanium serpentine flow channels and gold-plated current collectors. Initial testing achieved stable operation demonstrating feasibility of high-purity oxygen generation, with ongoing experiments characterizing voltage-current relationships and efficiency curves. Performance data will inform scaled multistack design where 15–20 units achieve partial oxygen supply for critical care. The Egyptian team synthesised Ni–Fe–Co layered double hydroxide catalysts demonstrating 320 mV overpotential at 10 mA/cm2 , reducing precious metal requirements by 85%.

The modular IoT monitoring system was successfully developed using header-based PCB design, featuring Arduino Nano with SIM800L GSM and DS3231 RTC. The custom PCB incorporates dedicated power regulation with 3300 µF capacitance, level-shifting circuits, and separate analog/digital ground planes. Standardized sensor * Correspondence: <davies.segera@uonbi.ac.ke>

1

80

interfaces enable plug-and-play connectivity. This architecture currently streams operational parameters to cloudbased monitoring, with plans for expanded multi-stack capabilities.

Digital twin development using MATLAB/Simulink incorporates Kenya-specific solar profiles to model optimal PV-to-electrolyser ratios. Life-cycle assessment indicates potential 46% carbon footprint reduction versus cylinder delivery [5]. Community engagement included stakeholder workshop with 42 participants, securing buy-in for Environmental and Social Impact Assessment (July 2025).

3. Next Targets (June – December 2025) The next phase transitions to integrated system development and pilot-scale validation. By July, we complete performance mapping of the PEM stack across temperature (60–80°C), pressure (1–3 bar), and current density (0.1–2 A/cm2 ) ranges. Critical milestone: integrating IoT monitoring with PEM electrolyser, creating unified monitoring through Nextion HMI dashboard. Based on results, we construct a multi-stack system (15–20 individual 4 cm2 stacks) targeting 3–5 Nm3 /day oxygen output (15–25% of KNH critical care requirements).

The system integrates 10–15 kWp solar PV and 20 kWh battery storage for 24/7 operation in acoustically treated enclosure (<55 dB). IoT deployment includes field-hardening: conformal coating, mechanical reinforcement, IP65 enclosures. Sensor expansion leverages modular design for vibration, current, and leak detection without new PCB designs.

Laboratory validation focuses on ISO 10083 oxygen quality verification. Target metrics: oxygen output 3–5 3 Nm /day, hydrogen co-production 0.3–0.5 kg/day, system availability >90%. Capacity building: two workshops training 15 KNH technicians using WHO-UNICEF framework. Economic modeling projects 25–35% oxygen cost reduction for critical care units. Consortium will submit to Applied Energy (Q4 2025) and produce Afretec policy brief.

4. Envisioned Publications 1. Modular IoT monitoring for electrochemical applications, in preparation for submission to Wiley Journal of Computer Networks and Communications 2. HMI dashboard for electrolyser monitoring, in preparation for submission to Wiley Journal of Mobile Information Systems 3. ChatGPT-powered IoT electrolyser analytics, in preparation for submission to Wiley Journal of Software: Practice and Experience

5. Project Figures

81

(a) Laboratory PEM fuel cell stack (3-cell, 4 cm2 ).

(b) Modular IoT monitoring PCB with GSM module.

Figure 1: Core hardware components for oxygen generation and real-time monitoring.

(a) 7-inch Nextion HMI touch interface.

(b) Cloud-based web dashboard with ChatGPT analytics.

Figure 2: Monitoring interfaces: (a) Nextion HMI with real-time visualization; (b) Web analytics platform for predictive maintenance.

82

Figure 3: Research consortium meeting with KNH Management (April 2025).

References [1] G. Maggio, G. Squadrio, and A. Nicita, “Hydrogen and medical oxygen by renewable energy based electrolysis: A green and economically viable route,” Applied Energy, vol. 306, p. 117993, 2022.

[2] B. Otieno et al., “A comprehensive approach to medical oxygen ecosystem building: Implementation in Kenya, Rwanda, and Ethiopia,” Global Health: Science and Practice, vol. 10, no. 6, p. e2100781, 2022.

[3] Nation Media Group. (2023, June 19). The status of Kenya’s medical oxygen infrastructure. Daily Nation.

[Online]. Available: https://nation.africa/kenya/health/ [4] Green Hydrogen Organisation. (2023). Kenya country profile: Green hydrogen opportunities. [Online].

Available: https://gh2.org/countries/kenya [5] M. Abou Houran et al., “Design and optimization of COVID-19 hospital wards to produce oxygen and electricity through solar PV panels with hydrogen storage systems,” Energy, vol. 259, p. 125006, 2022.

In preparing this abstract, the authors used [Antropic Claude Sonnet 4] with the following prompt(s): [Please find attached my LaTeX code for my abstract. While maintaining the code’s structure, correct any grammatical errors and ensure coherence throughout the abstract.]. The authors reviewed and edited the abstract as needed.

They take full responsibility for its content.

83

Strengthening Water Access and Quality in Selected African Countries 1. Research Team: 1. Prof. M.O.H. Amuda, University of Lagos, Nigeria (Principal Investigator) 2. Assoc. Prof. F.O. Agunbiade, University of Lagos, Nigeria (Co-PI WP1) 3. Prof. T. Lawanson, University of Lagos, Nigeria (Co-PI WP1) 4. Prof. T.A. Fashanu, University of Lagos, Nigeria (WP2) 5. Prof. T.O. Mbuya, University of Nairobi, Kenya (Co-PI, Kenya) 6. Prof. U.G. Wali, University of Rwanda (Co-PI, Rwanda) 7. Dr. A.A. Yinusa, University of Lagos, Nigeria (WP2) 8. Engr. Micheal Ogundero, University of Lagos (Ph.D. Student) 2. Project Objectives The overarching aim of this project is to provide a platform for harnessing human capital through multiinstitutional partnerships to develop an Afrocentric knowledge ecosystem for providing solutions to some of Africa’s existential challenges via innovation and digital inclusion.

In the present project, the focus is on addressing water insecurity as an existential challenge. Specific objectives of the project are to: Deploy the concept of community participation in the design, implementation, and validation of a digital technology framework capable of real-time monitoring of contaminants which will continually advise on the choice of optimum filtration policy and procedure.

Use knowledge and understanding obtained through citizen science in conjunction with parametric optimisation of transport of contaminants through porous media to develop a scalable, replicable, fitfor-purpose smart water purification technology to strengthen water access and quality.

Provide a platform for multi-institutional collaboration in the creation of Africa’s knowledge ecosystem for building human capital for the transformation of the African economy, improving job creation and prosperity.

Develop a sustainable pathway for the consolidation of the multi-institutional partnership which will include the co-creation of graduate learning activities, staff exchange, joint graduate research supervision, joint publications, co-hosting of scientific meetings and exhibitions.

Implement a framework for technology transfer and policy declarations for consolidating and extending the pilot project to wider communities in the African continent as well as building a resilient strategy to improve health and economic well-being.

3. Recent Progress

84

3.1 Strengthening Urban Resilience through Community-Led Water Management: A Case Study of Slum Communities in Lagos, Nigeria Strengthening the capacity of slum communities to manage their own water resources, while simultaneously advocating for inclusive and equitable government policies, is key to achieving water security and building resilience in rapidly urbanizing cities like Lagos. Two community engagement sessions have been conducted across each of the three selected areas to educate them on the need for a low-cost solution to water insecurity. Surveys have been conducted as a follow up to the sessions to ascertain their water quality and economic activities to proffer an optimal solution. This study has explored the role of community-led water management initiatives in enhancing resilience within urban slum communities, focusing on Iwaya, Ago Egun, and Ilaje in Lagos, Nigeria. The findings show the critical interplay between socio-economic characteristics, water access patterns, and the availability of both community-driven and government-supported water infrastructure.

3.2 A Soft Sensor Based Inference Engine for Water Quality Assessment and Prediction Water potability prediction and quality parameter estimations are very important to improve water quality.

Machine learning algorithms have been used by several researchers in this area and have proven to be efficient. In this study, water quality sample data was acquired from Consortium of Universities for the Advancement of Hydrologic Science (CUAHSI) through https://www.hydroshare.org/resource/4ab43e1b507b496b9b42749701daed5c/. Machine learning algorithms were used to develop soft sensors that estimated pH in the absence of Dissolved Oxygen (DO) and DO in the presence of the estimated pH. An inference engine was built using logistic regression, decision tree, random forest, XGBoost and neural networks to correctly predict water potability. The developed soft sensors and inference engine were validated using various performance metrics like model accuracy, F1 score and Confusion matrix. The study used a data driven approach to develop soft sensors as opposed to the model driven (white box) approach. The user interface for water potability prediction is shown in Figure 1 whilst a typical interface for water portability prediction result is shown in Figure 2.

85

Figure 1. User interface for the soft sensors that estimate pH and DO.

Figure 2. User interface showing results of estimated pH and DO from the soft sensors water potability prediction from the inference engine 3.3 A Smart IoT cum Citizen Science Framework for Real-Time Remote Water Quality Monitoring and Assessment

86

The IoT is a network of interconnected devices that communicate with each other to exchange data and perform various functions. The architecture which is illustrated in Figure 3 uses sensors to collect data from the environment, and communication heads to send the data to other devices for processing and analysis.

In order to guard sensors and obtain data for a long period of time, people who stay in the vicinity of the study site are often needed. When this is done, they are referred to as Citizen Scientists. Citizen science refers to the participation of non-professional individuals, typically members of the public, in scientific research activities. Citizen scientists contribute to scientific research by collecting, processing, or analyzing data, or by engaging in other activities related to scientific research. In this study, citizens were trained at the study site on how to operate the sensor. Participants demonstrated a clear understanding of the water sensor kit and its components. They successfully collected and tested water samples, recorded data, and submitted it as per the guidelines. Water quality data was obtained from the DANOPLUS water parameters sensor deployed onsite at Iwaya, Lagos. Real time data was transmitted to a custom server developed with the Flask framework. Data is accessible on a custom dashboard which can be accessed on https://dev-waterprob.vercel.app. Soft sensors were integrated to serve as a form of validation for pH and temperature data.

The developed IoT system sends warnings to all stakeholders in the form of emails and slack messages whenever parameters are observed to be out of range. This is also displayed on the monitoring server dashboard.

Figure 3. Schematic diagram of the remote smart cloud-based real time water quality monitoring system 4. Next Targets ● Design a multi-stage filtration system using affordable, locally available materials such as clay,

87

charcoal, sandstone and ceramic elements. ● Construct and test the filtration system to evaluate its effectiveness in improving water quality in accordance with WHO and NIS standards. ● Analyze the system’s performance in terms of turbidity reduction, pH adjustment, total dissolved solids minimization, etc. ● Revisiting the communities to know the status of the IoT technologies 5. Peer-Reviewed Publications Resulting from this Project 1. Ogundero, M., Fashanu, T., Agunbiade, F., Orolu, K., Yinusa, A., Daudu, U., & Amuda, M. (2025). A Soft Sensor Based Inference Engine for Water Quality Assessment and Prediction.

Air, Soil and Water Research, https://doi.org/ 10.1177/11786221251315618.

88

Supporting Inclusive Air Quality Management in Africa with a Novel System of Digital Tools Principal Investigator: Peter J. Adams, Carnegie Mellon University (Pittsburgh) Co-Principal Investigators: Engineer Bainomugisha (Makerere University), Rebecca Garland (University of Pretoria), Telesphore Kabera (University of Rwanda), Anderson A. Kouassi (UJLoG), Mogesh Naidoo (CSIR, unfunded collaborator), Deo Okure (Makerere University), Siele Silue (UJLoG), N'Datchoh E. Toure (UJLoG), Raoul P. Toure (UJLoG)

1. Project Objectives Development decisions being made in Africa, including energy policies and transitions, will determine the economic livelihood of the continent’s growing population, have an impact on the world’s future greenhouse gas emissions, and the health and wellbeing of Africans affected by associated air pollution. However, there is limited technical capacity to assist decision-makers wishing to pursue sustainable development and protect public health. Air quality modeling and measurement tools used in wealthy countries are often prohibitively expensive in the African context.

The goal of this project is to build digitally enabled, low-cost, technical and policy analysis tools to aid African decision-makers manage air quality and protect public health. Here we have assembled a collaborative team spanning five African universities with expertise in air quality modeling, low-cost sensors, emissions inventory development, and policy assessment to re-think air quality management systems for Africa. There are three key ideas to our research plan. First, we will leverage low-cost sensors, which are enabled by machine learning calibrations and a cloud-based data management platform. We will leverage AirQo’s sensors, a premier African air quality measurement network in East Africa run out of Makerere University as a test bed for our approach. Second, for policy assessments, we will replace the resource-intensive chemical transport models (CTMs) used in the US and Europe with low-cost “reduced-complexity models” (RCMs). Third, we will explore synergies between low-cost sensors and RCMs, e.g. as ways to evaluate and refine emissions inventories.

Specific tasks to be pursued in this proposal are: 1) develop and compare two RCM approaches in East Africa and other African regions 2) evaluate RCMs against measurements and against each other and CTMs 3) enhance low-cost sensor calibrations with next-generation machine learning approaches 4) develop custom emissions inventories for specific African regions and assess the value of local emissions inventories versus generic global inventories

89

5) use sensors and RCMs together to improve emissions inventories via inverse modeling 6) demonstrate how these tools work together in a variety of policy assessments.

2. Recent Progress Task 1A: Development of the REACH model for major African regions A major tool used in this project is the Rapid Estimation of Air Concentrations for Health (REACH) model. This tool is designed to be highly accessible and portable to support air quality decision-making in the Global South. Work prior to the start of this project demonstrated that REACH performed well in the United States.

REACH is a Gaussian-Plume reduced-complexity model with chemistry that estimates ambient PM2.5 concentrations and marginal social costs over a designated region. Like similar tools, the main purpose of REACH is to predict marginal social costs, which are the health damages that result from air pollutant emissions of a given species at a given location.

In this work, a domain for Southern Africa was developed covering 10 countries in the region (see Figure 1). Pollutant emissions were taken from GFED4.1 (biomass burning), MEGAN2.1 (biogenic emissions), and either EDGAR6.1 or DACCIWA2 for other anthropogenic emissions.

Meteorological data is taken from the ERA5 global reanalysis data set. Resulting annual-average PM2.5 concentration predictions for 2018 are shown in Figure 1.

Additionally, the REACH model has been applied to similar regions in West and East Africa.

Preliminary results are promising but further evaluations are underway.

Task 1B: Chemical transport modeling and EASIUR-Southern Africa Although the REACH model is accessible and flexible, there may be applications where the simplicity of its chemistry and physics limits its utility in policymaking. Therefore, we are also developing a suite of air quality models for Southern Africa from most sophisticated to most simple: state-of-the-art chemical transport modeling (CTM) with the CAMx CTM, Estimating Air pollution Social Impacts Using Regression (EASIUR) which is a CTM emulator, and finally to REACH as the simplest approach. Comprehensive evaluations of the three approaches for Southern Africa will inform when REACH is sufficient versus when a more sophisticated approach may be necessary.

90

Fig 1. REACH 2018 estimates of annual-average PM2.5 concentrations in Southern and Western Africa.

Task 2: Evaluation of REACH predictions versus observations Evaluations of the PM2.5 predictions (Figure 1) against South Africa’s SAAQIS network show good model performance.

Task 3: Low-cost sensor calibrations with next-generation machine learning approaches Collocation of research-grade instruments with low-cost sensors remains limited across subSaharan Africa, creating challenges for site-specific calibration of low-cost sensors. Reliable calibration of low-cost PM2.5 sensors is essential for expanding air quality monitoring in the region, where reference-grade data are scarce. We are evaluating the transferability of a Random Forest calibration model across six sites in five African cities—Kampala, Nairobi (two sites), Accra, Antananarivo, and Addis Ababa.

Specifically, the AirQo team at Makerere University has assessed three calibration strategies: • • •

Direct transfer—applying a Kampala-trained model to other cities without retraining, (achieving: RMSE: 7.0–43 µg/m3, r: 0.65–0.89) Localised models—retraining with site-specific data (achieving: RMSE: 4.0–15.3 µg/m3, r: 0.86–0.96), and Leave-one-city-out—training on all cities’ data except one and testing on the held-out city (revealing high variability: RMSE: 5.9–42.5 µg/m3, r: 0.65–0.86).

Localised models consistently outperformed direct transfer, with particularly strong results in Nairobi and Addis Ababa. However, the leave-one-city-out approach highlighted challenges in generalisability, especially for cities with distinct pollution profiles like Accra (likely due to Harmattan dust). These findings demonstrate that localised calibration is critical for accuracy, but climate-based grouping (e.g., harmonising models for similar regions) could optimise scalability.

91

We recommend a hybrid strategy that prioritises localised training where feasible, while leveraging shared climatic or pollution patterns to extend coverage efficiently.

Task 6: Policy applications We have begun to use the REACH-East Africa model to assess the air quality co-benefits of Rwanda’s climate mitigation efforts reported in its Nationally Determined Contributions (NDC) reports. This work is a collaboration between Medinat Akindele and Peter Adams at CMUPittsburgh, Prof. Telesphore Kabera (U Rwanda), using data from REMA (Rwanda Environment Management Authority). This work is just beginning but will quantify lives saved in (and around) Rwanda due to some of the mitigation measures in Rwanda’s NDCs: clean cookstoves, hydropower, and cleaner vehicles.

3. Next Targets Plans for the next ~year are as follows: Task 1: Complete the development of the EASIUR-S Africa model (U Pretoria).

Task 2: Perform a comprehensive evaluation over South Africa that includes the CAMx CTM, EASIUR, REACH, and available observations. Evaluate REACH for East Africa with AirQo and Rwandan low-cost sensors networks.

Task 5: Use East African low-cost sensor data and the REACH model to do inverse modeling to assess likely emissions that lead to observations and compare those to “bottom up” emissions inventories.

Task 6: Complete the assessment of air quality health co-benefits from Rwandan NDCs. 4. Peer-Reviewed Publications Resulting from this Project Currently, a manuscript entitled, “Development and Evaluation of the REACH Air Quality Model for Policy Assessment in Southern Africa” by Akindele, Garland, and Adams is in advanced stages of draft and should be submitted in the next month.

Additionally, we anticipate submission in the coming year of manuscripts on the following work: • • •

CTM modeling over Southern Africa region (U Pretoria) Evaluation of REACH for East Africa with low-cost sensor data (U Rwanda) Health co-benefits of Rwandan NDCs (CMU-Pittsburgh)

92

Towards IoT-enabled Privacy-preserving Large-scale Healthcare Analytics in Africa: A Use Case on Monitoring a Cardiovascular Disease

1. Project Team PI: Tamer ElBatt, CSE Dept., The American University in Cairo, Egypt Co-PI: Turgay Celik, School of Electrical & Info. Eng., Univ. of Witwatersrand, South Africa Co-PI: Swarun Kumar, ECE Dept., Carnegie Mellon University, USA Co-PI: Dineo Mpanya, University of Witwatersrand, South Africa

2. Project Objectives The objective of this project is to research, design and develop digital technologies for healthcare, with focus on wireless, IoT and machine learning technologies for blood flow, heart monitoring and prediction of cardiovascular diseases (CVD). The project has the following main threads: • Explore novel wireless technologies, particularly mmWave radar, for monitoring blood flow. To the best of our knowledge, this is the first work to introduce this concept and build a prototype.

• Design and prototype a low-cost Ballistocradiography (BCG) acquisition circuit using a piezoelectric ceramic sensor (PCS) which captures ballistic forces, on the muscles, generated by the heart.

• Propose a Federated Learning framework across simulated hospitals for the prediction of a particular CVD, namely Coronary Heart Disease (CHD), due to its wide prevalence.

• Propose a novel BCG-to-ECG mapping using deep learning. This work is among the first around the world to introduce and explore this promising research direction in 2025.

• Along our team’s focus on ECG signals and motivated by the scarcity of large, diverse ECG datasets, our team is exploring the generation of synthetic ECG waveforms using conditional GANs (cGANs).

• Research and develop a novel ECG feature extraction machine learning model, towards automated CVD diagnosis. To the best of our knowledge, this is the first work to define the problem and introduce a sound, promising solution approach.

• Contribute to the broader societal impact in Africa, in line with Afretec, projected to be two-fold: o Train students and junior researchers at partner institutions, towards inclusive digital transformation capacity building with focus on the vibrant area of digital health. o Target, on the long-term, the health and well-being of African citizens, especially CVD patients.

3. Recent Progress 3.1 Wireless, Sensing and Signal Acquisition Innovations 3.1.1 Blood Flow Monitoring Using mmWave Radar A major contributor to CVD is high blood pressure, often called the “silent killer”, because it can damage arteries for years before any obvious symptoms. Under this thrust, we have developed PolyPulse, the first AI-driven radar system to measure pulse transit time and blood pressure at multiple key sites along the body

93

using a single device in a contact-free manner.

Our system transmits wireless signals towards a user and measures the tiny surface reflections from the pulse points. We then extract key features from the cardiac signal indicating the start of a heart beat and the arrival of a pulse at a body site. Our approach is privacypreserving, enables continuous monitoring without any contact-based sensors and fits within the form factor of smart home devices.

We have conducted a pilot clinical study on 35 subjects across a diverse population, including healthy participants and those with CVD, diabetes, or a high body mass index. In our study, we aimed to measure cardiac signals along the upper body and targeted mainly four sites: the heart, the wrist, the neck, and the head. The blood pressure derived by our system achieves a correlation of 0.92 at estimating diastolic blood pressure compared to a classic cuff-based blood pressure monitor on the arm. Our results exhibit performance that meets clinical guidelines established by the U.S. FDA for blood pressure monitors. We currently have a publication under review at a major journal. This work will push the envelope of what is possible with wireless health sensing as no such system exists today.

3.1.2 Ultra-low Cost Wrist-based BCG Acquisition Using a Piezoelectric Ceramic Sensor [4] Ballistocardiography (BCG), first recorded by Yendell Henderson (1905), offers a promising, cost-effective, user-friendly approach to heart monitoring, potentially expanding access to CVD detection in low-income, resource-constrained settings. It detects and measures ballistic forces that give rise to subtle muscle movements caused by heartbeats. Different approaches have been introduced for BCG acquisition in the literature, e.g., bed-based, wristband-attached, ear-worn and chair-embedded. In this task, we focus on userfriendly, wrist-based BCG acquisition using an ultra-low-cost piezoelectric ceramic sensor (PCS). The sensor converts mechanical stress, such as pressure or vibration, to voltage. However, the raw signal captured by the piezoelectric sensor is often weak and noisy, mandating signal conditioning before it can be used for practical purposes. We have conducted a pilot study and collected sample BCG signals from 17 subjects mostly students (IRB approval obtained in May 2024). The figure on the right shows our PCSbased BCG acquisition circuit (with a specially-designed wristband built by 3D printing, to house the PCS sensor). It also shows a sample of the collected BCG signals from three subjects. The initial cost breakdown of the BCG acquisition prototype circuit is ultra-low and totals to less than $12 including the Arduino.

3.2 CVD Data Analytics and Machine Learning Innovations 3.2.1 Federated learning for the Prediction of Coronary Heart Diseases (CHD) [3] We propose a federated learning (FL) framework, coined FedCVD, for predicting the coronary heart disease (CHD) due to its wide prevalence among adults. As a first step, we employed simple logistic regression and the more complex support vector machine (SVM). FedCVD leverages the privacy and scalability

94

advantages of federated learning to facilitate collaborative model training, across multiple hospitals, using decentralized patient data. We conducted experiments for predicting the 10-year risk of developing CHD, based on the Framingham Heart Study* dataset (from Kaggle data repository). Data imbalance challenges are addressed through exploring three known techniques, namely Random Over Sampling, Random Under Sampling, and Synthetic Minority Oversampling Technique (SMOTE). For the federated logistic regression with SMOTE, it achieves the best area under the ROC curve (AUC) of 0.704. This comes very close to the best performance of centralized logistic regression, trained on the entire dataset, with an AUC of 0.708, using random under sampling. For federated SVM, an AUC of 0.7340 is achieved using random under sampling. This outperforms all combinations of the studied models and data balancing techniques. These initial findings highlight the merits and great promise of federated learning for privacypreserving CVD prediction in scalable, decentralized settings.

3.2.2 Synthetic ECG Dataset Generation Using Conditional GANs (cGANs) Motivated by the limited access to large and diverse Electrocardiogram (ECG) datasets, which hinders progress in CVD diagnostics using AI, largely because of patient privacy regulations, we are developing a novel framework for synthetic ECG generation, combining FL with conditional Generative Adversarial Networks (cGANs). This approach allows multiple institutions to collaboratively train powerful models without sharing sensitive patient data. Our aim is to create a system capable of simultaneously generating highly realistic, synthetic ECG data reflecting diverse patient populations, improving the accuracy of automated CVD classification and providing clinically relevant, interpretable insights into model predictions. Our team has successfully curated and prepared the necessary heterogeneous ECG datasets, specifically designed to simulate non-independent and identically distributed (non-iid) data conditions, typical of real-world multi-entity settings. We developed initial baseline versions of our core generative AI models in a centralized setting, demonstrating the feasibility of generating synthetic ECG waveforms and performing classification with these architectures. This serves as a benchmark for our federated learning approach currently under development.

3.2.3 BCG-to-ECG Signal Mapping Using Deep Learning [2] High cost, hospital-grade and bulky ECG equipment challenges large-scale CVD diagnosis in developing countries, especially in rural, underserved communities. This thrust presents a novel approach for mapping patient-friendly (i.e. no electrodes) BCG waveforms to the widely adopted ECG, using machine learning.

This is inspired by the inherent correlation between the two waveforms as synchronous responses to the *

https://www.framinghamheartstudy.org/

95

same heart event. More specifically, we propose a novel, hybrid Deep Learning (DL) model, composed of convolutional neural networks (CNNs) and long short-term memory (LSTM), for BCG-to-ECG signal mapping. To train and test the proposed model, a bed-based public dataset is employed. For performance evaluation, we employ the Pearson Correlation Coefficient (PCC) as a widely-accepted metric from statistics for quantitatively assessing signal correlation. Our approach holds a great promise as evidenced by achieving a PCC of 0.97, on the average, when correlating our deep learning-generated ECG waveform to the real ECG captured from the subject (ground truth). BCG-to-ECG signal mapping opens ample room for the integration of ECG’s rich diagnostic knowledge over the years and BCG’s affordable, at-home, continuous monitoring for early diagnosis of CVD, in a low-cost, scalable manner.

3.2.4 ECG Feature Extraction Using Deep Learning – Towards Automated CVD Diagnosis [1] Manual analysis of ECG signals by qualified healthcare providers in rural, underserved areas is generally time-consuming, error-prone, and often inaccessible. This research thrust introduces a novel open-source deep learning model for extracting interpretable morphological and temporal ECG features with high accuracy to aid physicians in diagnosis. Our proposed approach preprocesses raw ECG signals and employs a CNN for feature extraction. Using the PTB-XL+ public dataset, our model exhibits high correlation with the ground truth features, effectively capturing global and lead-specific ECG characteristics. We demonstrate that a single lead (Lead II) performs comparably to all 12 leads, reducing the computational complexity.

Furthermore, a lower sampling frequency (100 Hz) preserves performance, offering an efficient alternative to typical 500 Hz sampling frequency. This ongoing work provides a novel, practical and interpretable ECG feature extraction approach for regions with limited access to qualified healthcare professionals and costly, commercial ECG analysis software. This is work in progress with potential extensions along multiple directions currently being explored.

3.3 Research Dissemination: Public, Online Workshop The project team co-organized a public, online workshop hosted by the AUC: • Title: Vision and Innovations for Digital Health in Africa and beyond.

• Date: Tuesday April 8th, 2025.

• Format: Online, Half-day - 2:00 – 6:15 pm (Cairo Local Time).

96

• Program: Seven talks including three by our Afretec project partners on their research activities and major findings. We also hosted four invited talks by experts in digital health, from Rwanda, Egypt, Italy and the United States. A glimpse of the program is shown to the right.

• Announcement: widely disseminated in Egypt, Africa and beyond.

• Attendance: around 35 participants, mostly from Egypt and Africa.

4. Next Steps • mmWave Radar: our current system can measure cardiac signals at only four points, however, there are more pulse points on the body, e.g., the brachial, femoral and tibial artery. This can help diagnose arterial stiffening along these pathways and provide complete picture of the cardiovascular health. The multi-point pulse transit time is an intermediate step towards extracting coarse multi-point BCG waveform.

• Low-cost PCS-based BCG Acquisition: continue testing and developing the circuit design to filter out noisy oscillations, optimize the circuit design and enhance the quality of the acquired signal. A particular challenge is the tiny wrist pulse site which yields a weak, noisy BCG signal • cGANS: integrate generative models into the FL framework across hospitals. We will conduct comparative studies to evaluate different types of AI architectures, e.g., exploring variations in sequence modeling and generative techniques. The goal is to optimize performance for high-fidelity data generation, high accuracy classification and meaningful interpretability for diagnostic purposes and clinical decisions. We will also integrate and assess differential privacy in the context of our problem.

• BCG-to-ECG signal mapping: develop the proposed BCG-to-ECG signal mapping learning model to address technical challenges, e.g., scale the training process to more and diverse patients’ datasets and other performance metrics, among other practical considerations.

• LLM-based feature extraction: continue developing our novel concept and proposed approach of Large Language Models-based feature extraction for federated learning with heterogeneous clients, i.e. with overlapping, yet, potentially different, sets of attributes collected at different hospitals.

• CVD patients’ rehospitalization prediction: continue working on the important problem of CVD patients’ rehospitalization prediction using FL. The Wits team received an ethics approval in 2024 (Dr. Dineo Mpanya) to utilise a dataset for rehospitalisation of heart failure patients from a hospital in Johannesburg, South Africa. However, the sample size was too small for FL and attempts to augment it with synthetic data had limited success. A related challenge is the limited-quality of public datasets found for training the model. We are in the process of examining different machine learning models on a public dataset as well as exploring other datasets relevant to this problem.

5. References 1. Y. Abuzied, H. Abd-Eltawab, A. Gaber, T. ElBatt, “Towards Automated CVD Diagnosis: Deep Learning-Based ECG Feature Extraction,” accepted to 5th IEEE ICTS4eHealth, Italy, July 2025.

2. H. Abd-Eltawab, A. Gaber, T. ElBatt, “Towards Machine Learning-based Mapping of Low-cost BCG to ECG,” IEEE Healthcom, Nara, Japan, Nov. 2024.

3. A. Gaber, H. Abd-Eltawab, T. ElBatt, “FedCVD: Towards a Scalable, Privacy-Preserving Federated Learning Model for Cardiovascular Diseases Prediction,” 8th ICMLSC, Jan. 2024.(Best Presentation) 4. H. Abd-Eltawab, A. Gaber, Y. Abuzied, T. ElBatt, “Piezoelectric Ceramic Sensor for Wrist-based BCG Acquisition,” AUC CSE Technical Report, April 2025.

97

Wakanda Engineering Smart Solutions for Mother Wakanda AI: AI: Engineering Smart Solutions for Mother Tongues

by Advancing Language Technologies in the African Tongues

by Advancing Language Technologies in theContext African Context

Bhiksha Raj, Ph.D* 1 , Abraham Nyete, Ph.D†2 , and Moise Busogi, Ph.D‡3 1

3

1

Carnegie Mellon University Carnegie Mellon University Africa 2 University of Nairobi

Project Objectives

Despite significant publication efforts by the African Natural Language Processing (NLP) community, many of the published works have seen little to no practical use in real-world applications, especially in resource-constrained environments. Most language translation solutions for the African context are textbased and fail to serve populations where literacy levels are low, languages are primarily spoken rather than written, and technological deployment occurs in highly constrained environments. A substantial portion of the population who speak indigenous languages cannot access or benefit from text-based technologies due to limited literacy skills. Additionally, individuals with visual impairments face immense barriers in accessing written information, this includes the estimated 43 million people who are blind and 295 million with moderate to severe vision impairments. These challenges further increase the digital divide and limit the reach of digital AI solutions for machine translation in Africa.

This project proposes a speech-first framework as a more inclusive and practical alternative. By leveraging speech-to-text, text-to-speech, and speech-to-speech translation systems, we aim to build solutions that align more closely with the communication needs of African communities. The project’s core objective is to develop and deploy a scalable speech translation framework that bridges the gap between cutting-edge research and real-world usability.

To achieve this, we begin with curation of multilingual speech datasets for selected African languages, with an initial focus on Kinyarwanda. Using these datasets, we will fine-tune and adapt stateof-the-art automatic speech recognition/translation and text to speech models, optimizing them for both accuracy and efficiency in resource-constrained settings.

The final deliverable is an end-to-end speech translation system, accompanied by an open-source software toolkit. By focusing on speech applications, this solution will foster digital inclusion of nonliterate users and individuals with visual impairments. Through this initiative, we aim to create impactful and inclusive technologies that move beyond academic benchmarks and deliver real-world value to the African community.

The following objectives define the scope of this project: 1. Identify/Collect and annotate large-scale datasets of spoken (and optionally written) African languages to develop and evaluate machine learning models for speech-related tasks. * Principal Investigator † ‡

Co-Principal Investigator Co-Principal Investigator

1 98

2. Assess the adaptability of current models addressing various aspects involved in seamless communication, including speech-to-speech translation (S2ST), speech-to-text translation (S2TT), textto-speech translation (TTST), text-to-text (T2TT) translation, and automatic speech recognition (ASR), to the selected African languages. This includes Translatotron models and SeamlessM4T among others.

3. Adapt existing speech processing technologies to the linguistic and cultural contexts of African languages, considering dialectal variations, sociolinguistic factors, and linguistic diversity within regions.

4. Define or adapt appropriate evaluation metrics and benchmarks to assess the performance and effectiveness of speech processing systems for African languages, ensuring reliable and consistent evaluation methodologies.

5. Map out key contributors to the speech processing ecosystem for African languages and promote open-source development and collaborative research initiatives to facilitate knowledge sharing, resource sharing, and community-driven innovation in the field of African language processing.

6. Establish standards, guidelines, and best practices for developing, deploying, and evaluating speech and text processing technologies for African languages, fostering interoperability and reproducibility in research and development efforts.

7. Address ethical considerations and societal implications of deploying speech processing technologies for African languages, including issues related to data privacy, cultural sensitivity, and linguistic rights.

8. Ensure that the resulting technologies are actually usable and accessible to diverse user populations, including those with limited literacy or technological literacy.

To ensure the timely realization of the objectives, the project is managed in an agile fashion with several work packages — namely work packages 2, 3, 4 and 5 — involved in both waterfall and iterative processes (see Figure 1 for project PERT chart).

The project begins with WP1 that deals with the identification, collection, and pre-processing of African voice datasets. An important deliverable of this work package is an extensive repository of datasets for speech recognition and translation for African languages, as well as a module for standardizing the datasets.

Model implementation, training, and evaluation are carried out in WP2 and WP3. These work packages constitute a critical stage of the project as they involve most of the design and modeling activities. A significant number of modules, functions, and utilities will be developed in WP2 in order to support training and deployment of the implemented model as part of WP3 and WP4. WP3 also carries out the selection and the monitoring of meaningful performance metrics that will be used as a proxy to success measurement in the project. One significant outcome of Work Package 2 (WP2) is the model software package, consolidating all the modules and individual software components implemented in this project.

In order to assess the usability of the models that will result from WP2 and WP3, a use case demonstration will be carried out in WP4. A more explicit description of those use cases will be provided after completing WP1.

The project management and system engineering activities are carried out in WP5.

2

Recent Progress

Our team has made significant strides in building the foundational elements for the Wakanda AI speech translation framework. This initial phase has focused on critical data sourcing and infrastructure development to support robust model training and evaluation.

2 99

Figure 1: PERT chart showing the dependencies between the technical work packages. WP5 orchestrates all the activities involved in the other work packages.

We have assembled an extensive collection of speech datasets specifically tailored for Automatic Speech Recognition (ASR) and Speech Translation (ST) tasks. These datasets capture the rich linguistic diversity and dialectal variations essential for developing impactful language technologies across Africa.

To establish a feasible scope, we identified 100 priority mother tongue languages spoken throughout the continent. As illustrated in Figure 2, we have successfully gathered speech recognition and translation datasets for 30 languages thus far. We have ensured that the licenses permit our research efforts.

A major technical achievement has been the development of a dataset module that standardizes access to these diverse speech collections. This infrastructure innovation enables researchers and developers to: 1. Access multiple speech datasets through a consistent, well-documented interface; 2. Streamline preprocessing workflows and feature extraction pipelines; 3. Conduct rigorous cross-dataset experiments and benchmarking; and 4. Overcome technical barriers traditionally associated with African language resources.

This unified data access layer represents a critical step toward democratizing African NLP research by significantly reducing the engineering overhead required to work with these languages. By standardizing data access patterns, we have established a foundation that will accelerate model development and enable more efficient fine-tuning of state-of-the-art speech processing systems for diverse African contexts.

The data collection and infrastructure development completed in this phase directly supports our speech-first approach, providing the essential training resources needed for adapting existing speech technologies while ensuring these adaptations remain practical for deployment in resource-constrained environments.

3

Next Targets

Having curated a wide-ranging dataset of speech in several African languages, the next phase of the project would focus on model development, benchmarking, and deployment. Our immediate focus is to begin training models for speech-to-text translation (S2TT) using the curated datasets. These models 3 100

Figure 2: African language coverage map highlighting the countries where we cover at least one mother tongue language spoken. will enable automatic conversion of speech into written text across multiple languages. Alongside model development, we will create custom benchmarks on the curated datasets to systematically evaluate model performance and guide iterative improvements. In parallel, we aim to develop an initial deployment system that allows us to test and demonstrate the usability of these models outside of research settings, making them accessible to users.

Aligned with the broader project objectives, we will also begin expanding beyond speech-to-text by developing models for: • Speech-to-speech translation (S2ST), which allows speech in one language to be directly translated into speech in another.

• Text-to-speech translation (T2ST), enabling written text in one language to be rendered as naturalsounding speech in another.

• Text-to-text translation (T2TT), for cases where written multilingual communication is needed.

These components form the basis of the proposed speech translation framework. Outlined below are the key targets for the next phase of the project: 1. Develop and train speech-to-text translation (S2TT) models using annotated data.

2. Design and implement benchmarks, evaluating model performance and guide iterative improvements.

3. Develop an initial real-world deployed system to demonstrate and test model usability in constrained environments.

4. Initiate development of additional translation components: Speech-to-Speech Translation (S2ST), Text-to-Speech Translation (T2ST), and Text-to-Text Translation (T2TT).

5. Prepare for community testing and feedback to ensure alignment with local user needs and continuous improvement.

4 101

4

Peer-Reviewed Publications Resulting from this Project

No publications to report.

5 102