Publication Archive
Presentations
2025
Grace, Njogu
Empowering Africa’s Health: AI-Driven Solutions for Universal Access Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{GraceNjogu2025,
title = {Empowering Africa’s Health: AI-Driven Solutions for Universal Access},
author = {Njogu Grace},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Njogu_Grace.pdf?generation=1755026546227930&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {In many parts of Africa, access to quality healthcare remains a major challenge due to infrastructural gaps, workforce shortages, and affordability issues. mHealth4everyone is a mobile health initiative that leverages artificial intelligence (AI) to bridge this gap by providing accessible, low-cost, and data-driven healthcare solutions via mobile phones. Our AI-powered platform delivers early diagnostic support, health education, telemedicine, and predictive analytics to underserved communities. This poster presents our methodology, results from pilot implementations, and how our scalable model contributes to achieving Universal Health Coverage (UHC) and SDG 3 in Africa.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
In many parts of Africa, access to quality healthcare remains a major challenge due to infrastructural gaps, workforce shortages, and affordability issues. mHealth4everyone is a mobile health initiative that leverages artificial intelligence (AI) to bridge this gap by providing accessible, low-cost, and data-driven healthcare solutions via mobile phones. Our AI-powered platform delivers early diagnostic support, health education, telemedicine, and predictive analytics to underserved communities. This poster presents our methodology, results from pilot implementations, and how our scalable model contributes to achieving Universal Health Coverage (UHC) and SDG 3 in Africa.
Mutembesa, Daniel
MSense: Dynamic Route Planning and Adaptive Incentive Mechanisms for Mobile Sensors Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{MutembesaDaniel2025,
title = {MSense: Dynamic Route Planning and Adaptive Incentive Mechanisms for Mobile Sensors},
author = {Daniel Mutembesa},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Daniel_Mutembesa.pdf?generation=1755026600833055&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {In dynamic route planning for mobile sensors used in air quality monitoring, incorporating
heterogeneity among sensors is crucial for optimizing coverage, reducing costs, and enhancing data quality.
Heterogeneity considers different sensor capabilities, priorities, roles, and requirements, making the model
more realistic and effective. This paper presents an optimal dynamic route planning model designed for
mobile sensors, incorporating diversity guarantees for data collection and accomodating sensor heterogeneity
under resource-constrained conditions. Our approach introduces heterogeneity by accounting for varying
capabilities, priorities, roles, and requirements of different mobile sensors. Specifically, we model the
sensors’ different speeds, coverage radii, and battery lives, allowing for a comprehensive optimization
of routes based on these heterogeneous characteristics. The algorithm also integrates priority weights to
reflect the differing importance of nodes and areas for each sensor, and includes incentive structures to
encourage efficient coverage. Additionally, we introduce constraints to ensure collocation requirements and
diversity thresholds are met, promoting effective coordination and reducing redundancy among sensors. The
objective function maximizes overall coverage and efficiency while minimizing travel time and overlap.
This dynamic route planning model is crucial for applications where multiple mobile sensors operate under
diverse conditions, providing a robust solution for optimizing sensor deployment and resource utilization.
Simulation experiments on an urban road network graph with two mobile sensors deployed across various
source-destination pairs revealed key differences between three routing models. Model 1, though straight-
forward and efficient, failed to account for sensor path redundancies, resulting in less effective coverage.
Model 2 addressed this by penalizing overlapping routes, improving path diversity at the cost of added
complexity. Model 3, the most complex, considered both path diversity and sensor heterogeneity, offering
the best performance with minimal costs across all scenarios. While Model 1 is suitable for small networks,
Model 2 balances efficiency and complexity for moderate networks, and Model 3 is optimal for large-scale,
critical applications. Future work should focus on incorporating real-time data to enhance adaptability in
dynamic environments.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
In dynamic route planning for mobile sensors used in air quality monitoring, incorporating
heterogeneity among sensors is crucial for optimizing coverage, reducing costs, and enhancing data quality.
Heterogeneity considers different sensor capabilities, priorities, roles, and requirements, making the model
more realistic and effective. This paper presents an optimal dynamic route planning model designed for
mobile sensors, incorporating diversity guarantees for data collection and accomodating sensor heterogeneity
under resource-constrained conditions. Our approach introduces heterogeneity by accounting for varying
capabilities, priorities, roles, and requirements of different mobile sensors. Specifically, we model the
sensors’ different speeds, coverage radii, and battery lives, allowing for a comprehensive optimization
of routes based on these heterogeneous characteristics. The algorithm also integrates priority weights to
reflect the differing importance of nodes and areas for each sensor, and includes incentive structures to
encourage efficient coverage. Additionally, we introduce constraints to ensure collocation requirements and
diversity thresholds are met, promoting effective coordination and reducing redundancy among sensors. The
objective function maximizes overall coverage and efficiency while minimizing travel time and overlap.
This dynamic route planning model is crucial for applications where multiple mobile sensors operate under
diverse conditions, providing a robust solution for optimizing sensor deployment and resource utilization.
Simulation experiments on an urban road network graph with two mobile sensors deployed across various
source-destination pairs revealed key differences between three routing models. Model 1, though straight-
forward and efficient, failed to account for sensor path redundancies, resulting in less effective coverage.
Model 2 addressed this by penalizing overlapping routes, improving path diversity at the cost of added
complexity. Model 3, the most complex, considered both path diversity and sensor heterogeneity, offering
the best performance with minimal costs across all scenarios. While Model 1 is suitable for small networks,
Model 2 balances efficiency and complexity for moderate networks, and Model 3 is optimal for large-scale,
critical applications. Future work should focus on incorporating real-time data to enhance adaptability in
dynamic environments.
heterogeneity among sensors is crucial for optimizing coverage, reducing costs, and enhancing data quality.
Heterogeneity considers different sensor capabilities, priorities, roles, and requirements, making the model
more realistic and effective. This paper presents an optimal dynamic route planning model designed for
mobile sensors, incorporating diversity guarantees for data collection and accomodating sensor heterogeneity
under resource-constrained conditions. Our approach introduces heterogeneity by accounting for varying
capabilities, priorities, roles, and requirements of different mobile sensors. Specifically, we model the
sensors’ different speeds, coverage radii, and battery lives, allowing for a comprehensive optimization
of routes based on these heterogeneous characteristics. The algorithm also integrates priority weights to
reflect the differing importance of nodes and areas for each sensor, and includes incentive structures to
encourage efficient coverage. Additionally, we introduce constraints to ensure collocation requirements and
diversity thresholds are met, promoting effective coordination and reducing redundancy among sensors. The
objective function maximizes overall coverage and efficiency while minimizing travel time and overlap.
This dynamic route planning model is crucial for applications where multiple mobile sensors operate under
diverse conditions, providing a robust solution for optimizing sensor deployment and resource utilization.
Simulation experiments on an urban road network graph with two mobile sensors deployed across various
source-destination pairs revealed key differences between three routing models. Model 1, though straight-
forward and efficient, failed to account for sensor path redundancies, resulting in less effective coverage.
Model 2 addressed this by penalizing overlapping routes, improving path diversity at the cost of added
complexity. Model 3, the most complex, considered both path diversity and sensor heterogeneity, offering
the best performance with minimal costs across all scenarios. While Model 1 is suitable for small networks,
Model 2 balances efficiency and complexity for moderate networks, and Model 3 is optimal for large-scale,
critical applications. Future work should focus on incorporating real-time data to enhance adaptability in
dynamic environments.
serrhini,
Enhancing XSS Detection with LLM-Generated Obfuscation and Graph Neural Networks Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{serrhinimohamed2025b,
title = {Enhancing XSS Detection with LLM-Generated Obfuscation and Graph Neural Networks},
author = {serrhini},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/mohamed_serrhini.pdf?generation=1755026580657340&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Cross-Site Scripting (XSS) remains one of the most persistent and dangerous web vulnerabilities, allowing attackers to inject and execute malicious JavaScript within trusted web pages.
Despite years of research, detection remains challenging due to:
Advanced Obfuscation Techniques: Attackers use encoding, control flow distortion, and variable renaming to hide malicious behavior.
Limitations of Traditional Detection: Signature-based and token-based methods fail to generalize against unseen or obfuscated payloads.
Lack of Diverse Training Data: Models trained on simple or synthetic data often fail in real-world scenarios.
To address these challenges, we propose a robust detection pipeline that combines the code generation capabilities of Large Language Models (LLMs) with the structural learning power of Graph Neural Networks (GNNs).},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Cross-Site Scripting (XSS) remains one of the most persistent and dangerous web vulnerabilities, allowing attackers to inject and execute malicious JavaScript within trusted web pages.
Despite years of research, detection remains challenging due to:
Advanced Obfuscation Techniques: Attackers use encoding, control flow distortion, and variable renaming to hide malicious behavior.
Limitations of Traditional Detection: Signature-based and token-based methods fail to generalize against unseen or obfuscated payloads.
Lack of Diverse Training Data: Models trained on simple or synthetic data often fail in real-world scenarios.
To address these challenges, we propose a robust detection pipeline that combines the code generation capabilities of Large Language Models (LLMs) with the structural learning power of Graph Neural Networks (GNNs).
Despite years of research, detection remains challenging due to:
Advanced Obfuscation Techniques: Attackers use encoding, control flow distortion, and variable renaming to hide malicious behavior.
Limitations of Traditional Detection: Signature-based and token-based methods fail to generalize against unseen or obfuscated payloads.
Lack of Diverse Training Data: Models trained on simple or synthetic data often fail in real-world scenarios.
To address these challenges, we propose a robust detection pipeline that combines the code generation capabilities of Large Language Models (LLMs) with the structural learning power of Graph Neural Networks (GNNs).
Mwaibale, Upendo
Enhancing Detection of Common Bean Diseases Via Fast Gradient Sign Method – Trained Vision Transformers Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{MwaibaleUpendo2025,
title = {Enhancing Detection of Common Bean Diseases Via Fast Gradient Sign Method – Trained Vision Transformers},
author = {Upendo Mwaibale},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Upendo_Mwaibale.pdf?generation=1755026605432672&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Common bean farming is vital to food security in Tanzania, but it is often threatened by diseases such as bean rust and anthracnose. Traditional detection methods are limited in accuracy and speed, especially in rural settings. This study developed a deep learning model enhanced with adversarial training to improve early disease detection under real-world conditions. A farm-collected dataset of 59,072 images from Njombe, Iringa, and Mbeya was expanded to 100,000 through augmentation across four classes; healthy, rust, anthracnose, and other(images not related to common bean leaves). Three models were evaluated; the standard ViT, an adversarially trained CNN, and an adversarially trained ViT, with the latter achieving an accuracy of 99.4%. The model, robust to image noise via FGSM-based training, was deployed in a bilingual mobile app validated by farmers and extension officers. This work presents a practical and scalable tool to support smallholder farmers in managing bean diseases.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Common bean farming is vital to food security in Tanzania, but it is often threatened by diseases such as bean rust and anthracnose. Traditional detection methods are limited in accuracy and speed, especially in rural settings. This study developed a deep learning model enhanced with adversarial training to improve early disease detection under real-world conditions. A farm-collected dataset of 59,072 images from Njombe, Iringa, and Mbeya was expanded to 100,000 through augmentation across four classes; healthy, rust, anthracnose, and other(images not related to common bean leaves). Three models were evaluated; the standard ViT, an adversarially trained CNN, and an adversarially trained ViT, with the latter achieving an accuracy of 99.4%. The model, robust to image noise via FGSM-based training, was deployed in a bilingual mobile app validated by farmers and extension officers. This work presents a practical and scalable tool to support smallholder farmers in managing bean diseases.
Aliyu, Mahi Aminu
Towards Robust Generalization in African AI: Causal Inference and Domain Shift Mitigation Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{AliyuMahiAminu2025,
title = {Towards Robust Generalization in African AI: Causal Inference and Domain Shift Mitigation},
author = {Mahi Aminu Aliyu},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Mahi%20Aminu_Aliyu.pdf?generation=1755026595157904&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {This study explores the effectiveness of causal learning for domain generalization in low-resource African NLP. We introduce Afri-SemEval, a multilingual dataset translated into 17 African languages, and evaluate two causal paradigms: (i) representation learning via the DINER framework and (ii) causal data augmentation using GPT-4o-mini. Experiments on transformer models (XLM-R, Afro-XLMR-Large, Afro-XLMR-Large-76L) show that causal models converge faster and achieve comparable accuracy with fewer training steps. While counterfactual models demonstrate efficiency, their out-of-distribution (OOD) performance is mixed, with notable gains in languages like Yoruba and Igbo but limited generalization in others such as Amharic and Hausa. Results underscore the promise and limitations of causal approaches in low-resource African sentiment classification and highlight the importance of data quality and language-specific alignment.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
This study explores the effectiveness of causal learning for domain generalization in low-resource African NLP. We introduce Afri-SemEval, a multilingual dataset translated into 17 African languages, and evaluate two causal paradigms: (i) representation learning via the DINER framework and (ii) causal data augmentation using GPT-4o-mini. Experiments on transformer models (XLM-R, Afro-XLMR-Large, Afro-XLMR-Large-76L) show that causal models converge faster and achieve comparable accuracy with fewer training steps. While counterfactual models demonstrate efficiency, their out-of-distribution (OOD) performance is mixed, with notable gains in languages like Yoruba and Igbo but limited generalization in others such as Amharic and Hausa. Results underscore the promise and limitations of causal approaches in low-resource African sentiment classification and highlight the importance of data quality and language-specific alignment.
BADRI, Nabil
A Multilingual and Multidialect Deep Learning Approach for Hate and Abusive Speech Classification Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{BADRINabil2025,
title = {A Multilingual and Multidialect Deep Learning Approach for Hate and Abusive Speech Classification},
author = {Nabil BADRI},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Nabil_BADRI.pdf?generation=1755026576229896&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {This poster presents a deep learning approach for the classification of hate and abusive speech in a multilingual and multidialectal context. Our work addresses the challenges posed by linguistic diversity, especially in low-resource languages and dialects, by leveraging transformer-based models and transfer learning techniques. We evaluate our system across multiple datasets containing hate speech in different languages and dialects, demonstrating promising results in cross-lingual generalization and dialect robustness. This research contributes to the development of inclusive, language-aware AI systems capable of supporting safer online communication spaces.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
This poster presents a deep learning approach for the classification of hate and abusive speech in a multilingual and multidialectal context. Our work addresses the challenges posed by linguistic diversity, especially in low-resource languages and dialects, by leveraging transformer-based models and transfer learning techniques. We evaluate our system across multiple datasets containing hate speech in different languages and dialects, demonstrating promising results in cross-lingual generalization and dialect robustness. This research contributes to the development of inclusive, language-aware AI systems capable of supporting safer online communication spaces.
Abdelkader, Mahmoud
Multi-Objective Route Optimization Using Graph Neural Networks Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{AbdelkaderMahmoud2025,
title = {Multi-Objective Route Optimization Using Graph Neural Networks},
author = {Mahmoud Abdelkader},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Mahmoud_Abdelkader.pdf?generation=1755026547629289&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Naira, Abdou Mohamed
DVoice: Open-Source Voice AI for Africa and Beyond Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{NairaAbdouMohamed2025,
title = {DVoice: Open-Source Voice AI for Africa and Beyond},
author = {Abdou Mohamed Naira},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Abdou%20Mohamed_Naira.pdf?generation=1755026553163966&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {DVoice is an open-source, community-driven platform dedicated to addressing the underrepresentation of over 2,000 African languages in the digital world. By collecting, annotating, and sharing voice data, DVoice empowers local communities to contribute to speech technology development while preserving cultural heritage. The platform supports researchers and developers with free datasets, APIs, and models tailored to African linguistic diversity, including speech-to-text (STT), and text-to-speech (TTS) systems. With milestones like deploying 50+ models for 25+ languages by 2030, DVoice aims to foster inclusive innovation, preserve endangered oral traditions, and position Africa as a leader in ethical AI. The initiative seeks partnerships, funding, and community engagement to scale its impact globally.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
DVoice is an open-source, community-driven platform dedicated to addressing the underrepresentation of over 2,000 African languages in the digital world. By collecting, annotating, and sharing voice data, DVoice empowers local communities to contribute to speech technology development while preserving cultural heritage. The platform supports researchers and developers with free datasets, APIs, and models tailored to African linguistic diversity, including speech-to-text (STT), and text-to-speech (TTS) systems. With milestones like deploying 50+ models for 25+ languages by 2030, DVoice aims to foster inclusive innovation, preserve endangered oral traditions, and position Africa as a leader in ethical AI. The initiative seeks partnerships, funding, and community engagement to scale its impact globally.
Awak, Mbuotidem
EfikNLP: Parallel Corpora and Machine Translation System for Digital Inclusion Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{AwakMbuotidem2025,
title = {EfikNLP: Parallel Corpora and Machine Translation System for Digital Inclusion},
author = {Mbuotidem Awak},
url = {https://drive.google.com/file/d/1npGCtIW9dINzvH9vKoPMkUUzWc2zrdgh/view?usp=sharing},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Low-resource languages serve as invaluable repositories of human history, preserving cultural and intellectual diversity. Yet, they remain vastly underrepresented in the development of natural language processing (NLP) tools, particularly in machine translation (MT). Over the past decade, major advancements in MT have significantly improved translation for high-resource languages. However, these advances have not been equitably distributed. Many smaller indigenous languages, such as Efik, remain overlooked in both research and real-world applications, leaving their speakers excluded from the growing benefits of language technologies. This study seeks to change that by developing a comprehensive neural machine translation (NMT) system for the Efik-English language pair. Collaborating with native speakers, we built a culturally grounded parallel corpus of 1,040 sentence pairs. By fine-tuning the M2M100 model, we achieved a BLEU score of 8.437, indicating potential for improvement with more data.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Low-resource languages serve as invaluable repositories of human history, preserving cultural and intellectual diversity. Yet, they remain vastly underrepresented in the development of natural language processing (NLP) tools, particularly in machine translation (MT). Over the past decade, major advancements in MT have significantly improved translation for high-resource languages. However, these advances have not been equitably distributed. Many smaller indigenous languages, such as Efik, remain overlooked in both research and real-world applications, leaving their speakers excluded from the growing benefits of language technologies. This study seeks to change that by developing a comprehensive neural machine translation (NMT) system for the Efik-English language pair. Collaborating with native speakers, we built a culturally grounded parallel corpus of 1,040 sentence pairs. By fine-tuning the M2M100 model, we achieved a BLEU score of 8.437, indicating potential for improvement with more data.
Nakiranda, Proscovia
Detection of Stationary Pollution Sources and Profiling Using Satellite Imagery and Machine Learning Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{NakirandaProscovia2025,
title = {Detection of Stationary Pollution Sources and Profiling Using Satellite Imagery and Machine Learning},
author = {Proscovia Nakiranda},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Proscovia_Nakiranda.pdf?generation=1755026572247295&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Air pollution from stationary sources such as factories significantly impacts cities. Detecting and understanding these sources is challenging due to undocumented, emerging pollution sources and a lack of consolidated data. This research aims to create a comprehensive sources of pollution dataset with supplementary data from satellite data. This dataset is specifically developed to use machine learning algorithms for tasks such as detecting and profiling source point sources in satellite imagery. We trained a U-Net model with an accuracy of 80% on the manually labelled dataset to automatically identify potential stationary sources of pollution. To facilitate integration with other datasets, a post-processing step converts the model's predictions into a geospatial format (GeoJSON) with location information. By combining data on stationary sources of pollution with ground-based monitoring, we enable stakeholders to gain a deeper understanding of a city's air pollution profile. To support this, we have developed a visualization tool that makes the information easily accessible and actionable for decision-makers.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Air pollution from stationary sources such as factories significantly impacts cities. Detecting and understanding these sources is challenging due to undocumented, emerging pollution sources and a lack of consolidated data. This research aims to create a comprehensive sources of pollution dataset with supplementary data from satellite data. This dataset is specifically developed to use machine learning algorithms for tasks such as detecting and profiling source point sources in satellite imagery. We trained a U-Net model with an accuracy of 80% on the manually labelled dataset to automatically identify potential stationary sources of pollution. To facilitate integration with other datasets, a post-processing step converts the model's predictions into a geospatial format (GeoJSON) with location information. By combining data on stationary sources of pollution with ground-based monitoring, we enable stakeholders to gain a deeper understanding of a city's air pollution profile. To support this, we have developed a visualization tool that makes the information easily accessible and actionable for decision-makers.
Mohamed, Benayad
Generative AI for Urban Sustainability: Enhanced Basemaps from High-Resolution Satellite Imagery Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{MohamedBenayad2025,
title = {Generative AI for Urban Sustainability: Enhanced Basemaps from High-Resolution Satellite Imagery},
author = {Benayad Mohamed},
url = {https://docs.google.com/presentation/d/182r9Ch2G9f9dwHIn3nKV8hhkQYKMAy48/edit?usp=drive_link&ouid=112038764170136351774&rtpof=true&sd=true},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {The growing adoption of electric vehicles (EVs) demands the deployment of robust and well-planned charging infrastructure, particularly in emerging regions like Africa, where urban development is rapidly evolving, and spatial data is often underutilized. This study introduces a deep learning-based geospatial analysis framework to support the strategic planning of EV charging stations using high-resolution RGB satellite imagery and land cover/land use (LCLU) data. Our approach utilizes semantic segmentation models to extract detailed urban features from RGB imagery, enabling the identification of suitable charging station sites based on real-world land use characteristics. We constructed a land cover dataset comprising 13 urban classes: roads, buildings, vegetation, trees, bare land, solar PV, sidewalks, tracks, playgrounds, cars, grass, brown soil, and water. Multiple state-of-the-art deep learning models were evaluated, including SegFormer, UNet, PSPNet, and DeepLabV3. Among them, SegFormer achieved the highest performance with 97.3% accuracy, 0.938 F1-score, and 0.898 Intersection over Union (IoU), clearly outperforming the other models. The trained model was deployed on high-resolution satellite imagery of a Moroccan city to generate precise land cover maps. These maps were then analyzed to detect optimal locations for EV charging infrastructure, considering accessibility, available space, and urban activity patterns. The prototype serves as a proof of concept for a scalable, automated planning tool tailored to the African context. By combining remote sensing, land use analysis, and deep learning, this work provides a reproducible and adaptable method for improving EV infrastructure planning across African cities, contributing to smarter, data-driven urban development and more sustainable mobility transitions.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
The growing adoption of electric vehicles (EVs) demands the deployment of robust and well-planned charging infrastructure, particularly in emerging regions like Africa, where urban development is rapidly evolving, and spatial data is often underutilized. This study introduces a deep learning-based geospatial analysis framework to support the strategic planning of EV charging stations using high-resolution RGB satellite imagery and land cover/land use (LCLU) data. Our approach utilizes semantic segmentation models to extract detailed urban features from RGB imagery, enabling the identification of suitable charging station sites based on real-world land use characteristics. We constructed a land cover dataset comprising 13 urban classes: roads, buildings, vegetation, trees, bare land, solar PV, sidewalks, tracks, playgrounds, cars, grass, brown soil, and water. Multiple state-of-the-art deep learning models were evaluated, including SegFormer, UNet, PSPNet, and DeepLabV3. Among them, SegFormer achieved the highest performance with 97.3% accuracy, 0.938 F1-score, and 0.898 Intersection over Union (IoU), clearly outperforming the other models. The trained model was deployed on high-resolution satellite imagery of a Moroccan city to generate precise land cover maps. These maps were then analyzed to detect optimal locations for EV charging infrastructure, considering accessibility, available space, and urban activity patterns. The prototype serves as a proof of concept for a scalable, automated planning tool tailored to the African context. By combining remote sensing, land use analysis, and deep learning, this work provides a reproducible and adaptable method for improving EV infrastructure planning across African cities, contributing to smarter, data-driven urban development and more sustainable mobility transitions.
Mwaba, Natasha
Empowering communities in Zambia through innovation Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{MwabaNatasha2025,
title = {Empowering communities in Zambia through innovation},
author = {Natasha Mwaba},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Natasha_Mwaba.pdf?generation=1755026616510062&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Abstract:
Artificial Intelligence (AI) is emerging as a powerful tool for social good in Zambia, offering innovative solutions to some of the country’s pressing development challenges. In healthcare, AI-driven mobile applications like image-based diagnostic tools are helping rural clinics detect diseases such as malaria and tuberculosis more accurately and efficiently. In agriculture, AI-powered platforms provide small-scale farmers with real-time weather forecasts, crop disease alerts, and personalized farming advice, improving food security and productivity. In education, AI chatbots and language processing tools are being used to deliver interactive learning content in local languages, bridging the digital divide for students in remote areas. Additionally, AI is supporting data analysis for disaster response and public policy planning, enabling quicker, evidence-based decisions. These applications highlight how AI, when guided by ethical standards and inclusive policies, can accelerate Zambia’s progress toward sustainable development and improved quality of life for all.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Abstract:
Artificial Intelligence (AI) is emerging as a powerful tool for social good in Zambia, offering innovative solutions to some of the country’s pressing development challenges. In healthcare, AI-driven mobile applications like image-based diagnostic tools are helping rural clinics detect diseases such as malaria and tuberculosis more accurately and efficiently. In agriculture, AI-powered platforms provide small-scale farmers with real-time weather forecasts, crop disease alerts, and personalized farming advice, improving food security and productivity. In education, AI chatbots and language processing tools are being used to deliver interactive learning content in local languages, bridging the digital divide for students in remote areas. Additionally, AI is supporting data analysis for disaster response and public policy planning, enabling quicker, evidence-based decisions. These applications highlight how AI, when guided by ethical standards and inclusive policies, can accelerate Zambia’s progress toward sustainable development and improved quality of life for all.
Artificial Intelligence (AI) is emerging as a powerful tool for social good in Zambia, offering innovative solutions to some of the country’s pressing development challenges. In healthcare, AI-driven mobile applications like image-based diagnostic tools are helping rural clinics detect diseases such as malaria and tuberculosis more accurately and efficiently. In agriculture, AI-powered platforms provide small-scale farmers with real-time weather forecasts, crop disease alerts, and personalized farming advice, improving food security and productivity. In education, AI chatbots and language processing tools are being used to deliver interactive learning content in local languages, bridging the digital divide for students in remote areas. Additionally, AI is supporting data analysis for disaster response and public policy planning, enabling quicker, evidence-based decisions. These applications highlight how AI, when guided by ethical standards and inclusive policies, can accelerate Zambia’s progress toward sustainable development and improved quality of life for all.
Weya, Melissah
Building African Stereotypes Datasets For Responsible AI Evaluation Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{WeyaMelissah2025,
title = {Building African Stereotypes Datasets For Responsible AI Evaluation},
author = {Melissah Weya},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Melissah_Weya.pdf?generation=1755026589852217&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {AI risk assessments often overlook local socio-cultural perspectives, especially in underrepresented African regions (1–2% of NLP data). This results in biased AI outputs, reinforcing harmful stereotypes with real-world consequences in health, finance, and education, like misdiagnoses or loan denials. To bridge this critical gap, we are introducing an open-source, socio-culturally grounded extension of existing stereotype evaluation resources. Building on prior work (Dev et al., 2023; Davani et al., 2025), we surveyed participants in Senegal, Kenya, and Nigeria to capture top-of-mind societal associations and cultural stereotypes. While this method enabled authentic responses, it occasionally yielded superficial or overtly biased data, revealing both the richness and challenges of the format. For this pilot, we collected 1164 stereotypes from 107 respondents across the three countries, classifying responses by gender, religion, and ethnicity. We will continuously refine and share this dataset, incorporating local languages and voice-based responses for more diverse and culturally relevant data. We'll leverage on-the-ground surveyors and community collaborations for scalable data collection, allowing us to rapidly respond to evolving biases and expand to new countries. Our initial evaluation assesses language models' tendency to reflect these societal biases using Stereo Anti-Stereo (S-AS) pairs (Nangia et al., 2020). We will also explore complementary methods like the NLI-based framework (Dev et al., 2019) for stereotypical inferences. Beyond identification, future efforts will classify the types and degrees of harm these biases inflict across various sectors, aiming for a more nuanced and inclusive understanding of cultural stereotypes in Africa.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
AI risk assessments often overlook local socio-cultural perspectives, especially in underrepresented African regions (1–2% of NLP data). This results in biased AI outputs, reinforcing harmful stereotypes with real-world consequences in health, finance, and education, like misdiagnoses or loan denials. To bridge this critical gap, we are introducing an open-source, socio-culturally grounded extension of existing stereotype evaluation resources. Building on prior work (Dev et al., 2023; Davani et al., 2025), we surveyed participants in Senegal, Kenya, and Nigeria to capture top-of-mind societal associations and cultural stereotypes. While this method enabled authentic responses, it occasionally yielded superficial or overtly biased data, revealing both the richness and challenges of the format. For this pilot, we collected 1164 stereotypes from 107 respondents across the three countries, classifying responses by gender, religion, and ethnicity. We will continuously refine and share this dataset, incorporating local languages and voice-based responses for more diverse and culturally relevant data. We'll leverage on-the-ground surveyors and community collaborations for scalable data collection, allowing us to rapidly respond to evolving biases and expand to new countries. Our initial evaluation assesses language models' tendency to reflect these societal biases using Stereo Anti-Stereo (S-AS) pairs (Nangia et al., 2020). We will also explore complementary methods like the NLI-based framework (Dev et al., 2019) for stereotypical inferences. Beyond identification, future efforts will classify the types and degrees of harm these biases inflict across various sectors, aiming for a more nuanced and inclusive understanding of cultural stereotypes in Africa.
Simon, Nebiyu
Large Vocabulary Read-Mode Speech Corpora for Low-Resourced Ometo Languages: Gamo, Gofa, Dawuro and Wolaita Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{SimonNebiyu2025,
title = {Large Vocabulary Read-Mode Speech Corpora for Low-Resourced Ometo Languages: Gamo, Gofa, Dawuro and Wolaita},
author = {Nebiyu Simon},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Nebiyu_Simon.pdf?generation=1755026557089972&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Speech is a fundamental mode of human communication and a key interface for interacting with ma‐
chines. Automatic Speech Recognition (ASR) enables the conversion of spoken language into text, but
developing ASR systems requires large, high‐quality speech corpora.
For low‐resource languages like the Ometo languages (Gamo, Gofa, Dawuro, and Wolaita), such
datasets are scarce due to the high cost of data collection and limited technological resources.
In this study, we developed a 24.35‐hour multilingual speech corpus with corresponding transcriptions
for four Ometo languages. Using deep learning techniques, we built baseline ASR systems for each
language, achieving word error rates (WER) of: Gamo: 72.00%, Gofa: 57.94%, Dawuro: 62.22%,
Wolaita: 64.71%
The results confirm the usability of the corpus and its potential for further research and development
of ASR systems for underrepresented languages.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Speech is a fundamental mode of human communication and a key interface for interacting with ma‐
chines. Automatic Speech Recognition (ASR) enables the conversion of spoken language into text, but
developing ASR systems requires large, high‐quality speech corpora.
For low‐resource languages like the Ometo languages (Gamo, Gofa, Dawuro, and Wolaita), such
datasets are scarce due to the high cost of data collection and limited technological resources.
In this study, we developed a 24.35‐hour multilingual speech corpus with corresponding transcriptions
for four Ometo languages. Using deep learning techniques, we built baseline ASR systems for each
language, achieving word error rates (WER) of: Gamo: 72.00%, Gofa: 57.94%, Dawuro: 62.22%,
Wolaita: 64.71%
The results confirm the usability of the corpus and its potential for further research and development
of ASR systems for underrepresented languages.
chines. Automatic Speech Recognition (ASR) enables the conversion of spoken language into text, but
developing ASR systems requires large, high‐quality speech corpora.
For low‐resource languages like the Ometo languages (Gamo, Gofa, Dawuro, and Wolaita), such
datasets are scarce due to the high cost of data collection and limited technological resources.
In this study, we developed a 24.35‐hour multilingual speech corpus with corresponding transcriptions
for four Ometo languages. Using deep learning techniques, we built baseline ASR systems for each
language, achieving word error rates (WER) of: Gamo: 72.00%, Gofa: 57.94%, Dawuro: 62.22%,
Wolaita: 64.71%
The results confirm the usability of the corpus and its potential for further research and development
of ASR systems for underrepresented languages.
NINYIM, Astride Melvin FOKAM
Federated Learning for Respiratory Disease Forecasting in Africa Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{FOKAMNINYIMAstrideMelvin2025,
title = {Federated Learning for Respiratory Disease Forecasting in Africa},
author = {Astride Melvin FOKAM NINYIM},
url = {https://drive.google.com/drive/folders/1udB7WMF5Te6Rn36vaX3Sj3fjjwVTvqnY?usp=drive_link},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Respiratory diseases, including COPD, asthma, and pneumonia, claim over 6 million lives annually in Africa, representing 80% of the continent's infectious disease burden and causing $450 billion in losses from premature deaths plus $800 billion in productivity losses yearly. Driven by climate change and air pollution, the increasing frequency of extreme weather events underscores the urgent need for accurate outbreak prediction systems.
This study proposes a federated learning framework combined with deep learning to forecast climate-driven respiratory disease outbreaks across Africa. The approach addresses key challenges including data availability and quality issues, spatial and temporal variability in disease patterns, real-time prediction requirements, and forecasting uncertainty. The framework enables privacy-preserving model training on decentralized health and environmental data, addressing data sensitivity and infrastructure constraints.
The methodology integrates environmental data from ERA5, NASA POWER, and OpenAQ with health data from Synthea, OpenMRS, and WHO sources. A predictive LSTM model incorporates environmental variables (air quality, pollution) and health data (asthma, COPD diagnoses) using federated learning. Results demonstrate strong correlations between environmental factors and health outcomes, with PM2.5 showing 0.87 correlation and temperature 0.74 correlation with respiratory cases. Model optimization revealed that a 20-day input window maximizes early warning performance, achieving consistent predictive accuracy across both urban and rural settings with 1-week lead time.
This scalable, privacy-preserving solution supports multiple UN Sustainable Development Goals (SDGs 3, 9, 11, 13, and 17) and provides a foundation for timely interventions to reduce respiratory disease burden across Africa.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Respiratory diseases, including COPD, asthma, and pneumonia, claim over 6 million lives annually in Africa, representing 80% of the continent's infectious disease burden and causing $450 billion in losses from premature deaths plus $800 billion in productivity losses yearly. Driven by climate change and air pollution, the increasing frequency of extreme weather events underscores the urgent need for accurate outbreak prediction systems.
This study proposes a federated learning framework combined with deep learning to forecast climate-driven respiratory disease outbreaks across Africa. The approach addresses key challenges including data availability and quality issues, spatial and temporal variability in disease patterns, real-time prediction requirements, and forecasting uncertainty. The framework enables privacy-preserving model training on decentralized health and environmental data, addressing data sensitivity and infrastructure constraints.
The methodology integrates environmental data from ERA5, NASA POWER, and OpenAQ with health data from Synthea, OpenMRS, and WHO sources. A predictive LSTM model incorporates environmental variables (air quality, pollution) and health data (asthma, COPD diagnoses) using federated learning. Results demonstrate strong correlations between environmental factors and health outcomes, with PM2.5 showing 0.87 correlation and temperature 0.74 correlation with respiratory cases. Model optimization revealed that a 20-day input window maximizes early warning performance, achieving consistent predictive accuracy across both urban and rural settings with 1-week lead time.
This scalable, privacy-preserving solution supports multiple UN Sustainable Development Goals (SDGs 3, 9, 11, 13, and 17) and provides a foundation for timely interventions to reduce respiratory disease burden across Africa.
This study proposes a federated learning framework combined with deep learning to forecast climate-driven respiratory disease outbreaks across Africa. The approach addresses key challenges including data availability and quality issues, spatial and temporal variability in disease patterns, real-time prediction requirements, and forecasting uncertainty. The framework enables privacy-preserving model training on decentralized health and environmental data, addressing data sensitivity and infrastructure constraints.
The methodology integrates environmental data from ERA5, NASA POWER, and OpenAQ with health data from Synthea, OpenMRS, and WHO sources. A predictive LSTM model incorporates environmental variables (air quality, pollution) and health data (asthma, COPD diagnoses) using federated learning. Results demonstrate strong correlations between environmental factors and health outcomes, with PM2.5 showing 0.87 correlation and temperature 0.74 correlation with respiratory cases. Model optimization revealed that a 20-day input window maximizes early warning performance, achieving consistent predictive accuracy across both urban and rural settings with 1-week lead time.
This scalable, privacy-preserving solution supports multiple UN Sustainable Development Goals (SDGs 3, 9, 11, 13, and 17) and provides a foundation for timely interventions to reduce respiratory disease burden across Africa.
KABORE, Nematou
Virtual reality and anatomy learning with voice recognition Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{KABORENematou2025,
title = {Virtual reality and anatomy learning with voice recognition},
author = {Nematou KABORE},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Nematou_KABORE.pdf?generation=1755026576451325&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {This project aims to revolutionize the teaching of anatomy by developing a virtual reality (VR) educational application with voice recognition. It is designed for use with VR headsets such as the Oculus Quest 2 and the HP Reverb G2. The application incorporates interactive 3D anatomical models. A key dimension is the integration of AI voice recognition to navigate and interact with the anatomical models, allowing students to formulate voice commands to examine different body structures. The aim is to assess the impact of this VR technology on anatomy learning in terms of engagement and understanding. To do this, we used Unity, a powerful and flexible game engine, to create immersive 3D environments and intuitive interactions. Voice recognition is integrated using Wit.AI, allowing users to make voice requests to view anatomical models. An interactive avatar, developed with ConvAI, is also included to provide detailed explanations of different body parts, making learning more engaging and interactive.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
This project aims to revolutionize the teaching of anatomy by developing a virtual reality (VR) educational application with voice recognition. It is designed for use with VR headsets such as the Oculus Quest 2 and the HP Reverb G2. The application incorporates interactive 3D anatomical models. A key dimension is the integration of AI voice recognition to navigate and interact with the anatomical models, allowing students to formulate voice commands to examine different body structures. The aim is to assess the impact of this VR technology on anatomy learning in terms of engagement and understanding. To do this, we used Unity, a powerful and flexible game engine, to create immersive 3D environments and intuitive interactions. Voice recognition is integrated using Wit.AI, allowing users to make voice requests to view anatomical models. An interactive avatar, developed with ConvAI, is also included to provide detailed explanations of different body parts, making learning more engaging and interactive.
Lucas, Mgasa
Video Understanding Using LLMs for Intelligent CCTV Surveillance Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{LucasMgasa2025,
title = {Video Understanding Using LLMs for Intelligent CCTV Surveillance},
author = {Mgasa Lucas},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Mgasa_Lucas.pdf?generation=1755026609502337&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {The rapid expansion of surveillance systems has led to an increased demand
for intelligent video analysis. This research focuses on developing an AI-driven video
understanding system for CCTV cameras, leveraging machine learning (ML) and large
language models (LLMs) to enhance real-time security, anomaly detection, and
behavioral analysis. Our contribution involves creating an end-to-end deep learning
pipeline that integrates object detection, action recognition, and event summarization to
improve security monitoring efficiency. We employ state-of-the-art vision transformers
and open-source large language models to extract meaningful insights from live CCTV
footage. This research aims to refine our models further and explore their deployment in
smart city infrastructure, enterprise security, and home and public safety applications.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
The rapid expansion of surveillance systems has led to an increased demand
for intelligent video analysis. This research focuses on developing an AI-driven video
understanding system for CCTV cameras, leveraging machine learning (ML) and large
language models (LLMs) to enhance real-time security, anomaly detection, and
behavioral analysis. Our contribution involves creating an end-to-end deep learning
pipeline that integrates object detection, action recognition, and event summarization to
improve security monitoring efficiency. We employ state-of-the-art vision transformers
and open-source large language models to extract meaningful insights from live CCTV
footage. This research aims to refine our models further and explore their deployment in
smart city infrastructure, enterprise security, and home and public safety applications.
for intelligent video analysis. This research focuses on developing an AI-driven video
understanding system for CCTV cameras, leveraging machine learning (ML) and large
language models (LLMs) to enhance real-time security, anomaly detection, and
behavioral analysis. Our contribution involves creating an end-to-end deep learning
pipeline that integrates object detection, action recognition, and event summarization to
improve security monitoring efficiency. We employ state-of-the-art vision transformers
and open-source large language models to extract meaningful insights from live CCTV
footage. This research aims to refine our models further and explore their deployment in
smart city infrastructure, enterprise security, and home and public safety applications.
NIBIGIRA, Nadine
Real time cardiac monitoring system based on Artificial Intelligence Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{NIBIGIRANadine2025,
title = {Real time cardiac monitoring system based on Artificial Intelligence},
author = {Nadine NIBIGIRA},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Nadine_NIBIGIRA.pdf?generation=1755026568915876&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Cardiovascular diseases (CVDs) remain the leading cause of mortality globally, disproportionately affecting populations in low- and middle-income countries due to limited access to timely diagnosis and care.
This study presents UMUTIMA, a real-time cardiac monitoring system that leverages the Internet of Things (IoT) and Artificial Intelligence (AI) to provide an end-to-end framework for remote cardiovascular healthcare. The system architecture combines wearable IoT sensors to continuously collect vital signs such as heart rate, blood pressure, and cardiac rhythm and transmits the data securely via MQTT with TLS encryption to a cloud-based server. In the cloud, lightweight AI models perform real-time analysis to detect anomalies and trigger alerts.
A hybrid AI model combining Random Forest and CNN+LSTM achieved a training accuracy of 98.6% and test accuracy of 93.8%, demonstrating strong predictive performance. Another model, the Bagging Classifier, reached 89.7% on the training set but only 78.5% on the test set, indicating overfitting. The system integrates explainable AI techniques to identify and visualize the features contributing to each alert, thereby enhancing clinical transparency and decision-making.
UMUTIMA supports early detection of cardiac anomalies, enables continuous out-of-hospital monitoring, and reduces healthcare costs. It features both mobile and web interfaces to deliver notifications via SMS, email, and push messages, and is designed to interoperate with existing telemedicine platforms and Electronic Health Records (EHRs). Despite its promise, challenges remain, including ensuring data privacy, regulatory compliance, stable wireless connectivity, and energy-efficient operation.
Overall, UMUTIMA illustrates the transformative potential of integrating AI and IoT in cardiovascular care. With continued development and clinical validation, it could significantly improve cardiac health outcomes, particularly in resource-limited settings.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Cardiovascular diseases (CVDs) remain the leading cause of mortality globally, disproportionately affecting populations in low- and middle-income countries due to limited access to timely diagnosis and care.
This study presents UMUTIMA, a real-time cardiac monitoring system that leverages the Internet of Things (IoT) and Artificial Intelligence (AI) to provide an end-to-end framework for remote cardiovascular healthcare. The system architecture combines wearable IoT sensors to continuously collect vital signs such as heart rate, blood pressure, and cardiac rhythm and transmits the data securely via MQTT with TLS encryption to a cloud-based server. In the cloud, lightweight AI models perform real-time analysis to detect anomalies and trigger alerts.
A hybrid AI model combining Random Forest and CNN+LSTM achieved a training accuracy of 98.6% and test accuracy of 93.8%, demonstrating strong predictive performance. Another model, the Bagging Classifier, reached 89.7% on the training set but only 78.5% on the test set, indicating overfitting. The system integrates explainable AI techniques to identify and visualize the features contributing to each alert, thereby enhancing clinical transparency and decision-making.
UMUTIMA supports early detection of cardiac anomalies, enables continuous out-of-hospital monitoring, and reduces healthcare costs. It features both mobile and web interfaces to deliver notifications via SMS, email, and push messages, and is designed to interoperate with existing telemedicine platforms and Electronic Health Records (EHRs). Despite its promise, challenges remain, including ensuring data privacy, regulatory compliance, stable wireless connectivity, and energy-efficient operation.
Overall, UMUTIMA illustrates the transformative potential of integrating AI and IoT in cardiovascular care. With continued development and clinical validation, it could significantly improve cardiac health outcomes, particularly in resource-limited settings.
This study presents UMUTIMA, a real-time cardiac monitoring system that leverages the Internet of Things (IoT) and Artificial Intelligence (AI) to provide an end-to-end framework for remote cardiovascular healthcare. The system architecture combines wearable IoT sensors to continuously collect vital signs such as heart rate, blood pressure, and cardiac rhythm and transmits the data securely via MQTT with TLS encryption to a cloud-based server. In the cloud, lightweight AI models perform real-time analysis to detect anomalies and trigger alerts.
A hybrid AI model combining Random Forest and CNN+LSTM achieved a training accuracy of 98.6% and test accuracy of 93.8%, demonstrating strong predictive performance. Another model, the Bagging Classifier, reached 89.7% on the training set but only 78.5% on the test set, indicating overfitting. The system integrates explainable AI techniques to identify and visualize the features contributing to each alert, thereby enhancing clinical transparency and decision-making.
UMUTIMA supports early detection of cardiac anomalies, enables continuous out-of-hospital monitoring, and reduces healthcare costs. It features both mobile and web interfaces to deliver notifications via SMS, email, and push messages, and is designed to interoperate with existing telemedicine platforms and Electronic Health Records (EHRs). Despite its promise, challenges remain, including ensuring data privacy, regulatory compliance, stable wireless connectivity, and energy-efficient operation.
Overall, UMUTIMA illustrates the transformative potential of integrating AI and IoT in cardiovascular care. With continued development and clinical validation, it could significantly improve cardiac health outcomes, particularly in resource-limited settings.
Omer, Muhammad Abdulghaffar Muhammad
Rule-Based Reward Modeling for Large Reasoning Models Post-Training Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{MuhammadOmerMuhammadAbdulghaffar2025,
title = {Rule-Based Reward Modeling for Large Reasoning Models Post-Training},
author = {Muhammad Abdulghaffar Muhammad Omer},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Muhammad%20Abdulghaffar%20_Muhammad%20Omer%20.pdf?generation=1755026567278976&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Recent developments in the field of LLMs and the emergence of reasoning capabilities from foundational
models has sparked a wave of specialized models for reasoning and math tasks employing novel and
sophisticated prompting, training and finetuning techniques, these models are often referred to as Large
Reasoning Models LRMs. Some of the most prominent models tailored for such tasks are Open AI’s
o-series models, Llama Nemotron, Deepseek R1, and Qwen-Math model series. At the center of these
models’ development are ideas such as Chain of Thought (CoT) and Tree of Thought (ToT) prompting as
well as RLHF post-training. This work investigates the potential of simple reward modeling using rule
based techniques to enhance finetuning without using labeled examples. Our work builds on a recent
method called ”Test Time Reinforcement Learning (TTRL)”, that performs finetuning on LLMs using RL
with unlabeled data. TTRL takes a finite set of samples from the model during inference and uses majority
voting to construct binary rewards that are fed to the RL pipeline for finetuning. Our work develops an
additional intermediate step that adds or subtracts additional reward signals based on a set of rules such
as prompt-to-response length ratio, compression ratio, and the presence of code and other patterns in
the response. Two of our methods show significant improvement in reward and reward accuracy over
two math tasks, AIME 2024 and AMC.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Recent developments in the field of LLMs and the emergence of reasoning capabilities from foundational
models has sparked a wave of specialized models for reasoning and math tasks employing novel and
sophisticated prompting, training and finetuning techniques, these models are often referred to as Large
Reasoning Models LRMs. Some of the most prominent models tailored for such tasks are Open AI’s
o-series models, Llama Nemotron, Deepseek R1, and Qwen-Math model series. At the center of these
models’ development are ideas such as Chain of Thought (CoT) and Tree of Thought (ToT) prompting as
well as RLHF post-training. This work investigates the potential of simple reward modeling using rule
based techniques to enhance finetuning without using labeled examples. Our work builds on a recent
method called ”Test Time Reinforcement Learning (TTRL)”, that performs finetuning on LLMs using RL
with unlabeled data. TTRL takes a finite set of samples from the model during inference and uses majority
voting to construct binary rewards that are fed to the RL pipeline for finetuning. Our work develops an
additional intermediate step that adds or subtracts additional reward signals based on a set of rules such
as prompt-to-response length ratio, compression ratio, and the presence of code and other patterns in
the response. Two of our methods show significant improvement in reward and reward accuracy over
two math tasks, AIME 2024 and AMC.
models has sparked a wave of specialized models for reasoning and math tasks employing novel and
sophisticated prompting, training and finetuning techniques, these models are often referred to as Large
Reasoning Models LRMs. Some of the most prominent models tailored for such tasks are Open AI’s
o-series models, Llama Nemotron, Deepseek R1, and Qwen-Math model series. At the center of these
models’ development are ideas such as Chain of Thought (CoT) and Tree of Thought (ToT) prompting as
well as RLHF post-training. This work investigates the potential of simple reward modeling using rule
based techniques to enhance finetuning without using labeled examples. Our work builds on a recent
method called ”Test Time Reinforcement Learning (TTRL)”, that performs finetuning on LLMs using RL
with unlabeled data. TTRL takes a finite set of samples from the model during inference and uses majority
voting to construct binary rewards that are fed to the RL pipeline for finetuning. Our work develops an
additional intermediate step that adds or subtracts additional reward signals based on a set of rules such
as prompt-to-response length ratio, compression ratio, and the presence of code and other patterns in
the response. Two of our methods show significant improvement in reward and reward accuracy over
two math tasks, AIME 2024 and AMC.
Iliya, Nasiru
Targeting the Right Farmer: Predictive Analytics for Agricultural Input Subsidy Allocation in Sub-Saharan Africa Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{IliyaNasiru2025,
title = {Targeting the Right Farmer: Predictive Analytics for Agricultural Input Subsidy Allocation in Sub-Saharan Africa},
author = {Nasiru Iliya},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Nasiru_Iliya.pdf?generation=1755026565083581&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Policymakers in Sub-Saharan Africa (SSA) are renewing interest in agricultural input subsidy programs (ISPs) for their potential to reduce poverty, enhance food security, and promote gender empowerment. However, existing ISPs often use blanket distribution methods lacking targeting precision and graduation mechanisms, leading to inefficiencies and unsustainable resource use. This study proposes a machine learning–driven model to optimize subsidy allocation by identifying eligible beneficiaries based on features such as prior input use, demographics, and household vulnerability. Using farm-level data from 422 maize farmers in Taita Taveta, Kenya, we develop a smallholder index train four machine learning models. This study highlights key predictors - such as reliance on rain-fed farming, landholdings under one acre, and large household size - using SHAP value analysis. This data-driven approach supports more equitable, efficient, and scalable ISP targeting in resource-constrained settings.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Policymakers in Sub-Saharan Africa (SSA) are renewing interest in agricultural input subsidy programs (ISPs) for their potential to reduce poverty, enhance food security, and promote gender empowerment. However, existing ISPs often use blanket distribution methods lacking targeting precision and graduation mechanisms, leading to inefficiencies and unsustainable resource use. This study proposes a machine learning–driven model to optimize subsidy allocation by identifying eligible beneficiaries based on features such as prior input use, demographics, and household vulnerability. Using farm-level data from 422 maize farmers in Taita Taveta, Kenya, we develop a smallholder index train four machine learning models. This study highlights key predictors - such as reliance on rain-fed farming, landholdings under one acre, and large household size - using SHAP value analysis. This data-driven approach supports more equitable, efficient, and scalable ISP targeting in resource-constrained settings.
MKUMBO, HAPPINESS
Multilingual Natural Language Processing Conversational Platform for Promoting Blood Donation in Tanzania Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{MKUMBOHAPPINESS2025,
title = {Multilingual Natural Language Processing Conversational Platform for Promoting Blood Donation in Tanzania},
author = {HAPPINESS MKUMBO},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/HAPPINESS_MKUMBO.pdf?generation=1755026597782770&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Blood donations continue to become a public health challenge worldwide, and in many countries, including Tanzania, due to low donor retention and limited public awareness. To address this, this study presents a multilingual conversational AI platform that encourages and promotes voluntary blood donation, developed with a hybrid architecture that integrates Retrieval Augmented Generation (RAG) LaBSE embeddings, Normalized Compression Distance (NCD), and Llama-based response generation. The chatbot provided accurate, bilingual responses (in English and Kiswahili) tailored for Tanzanian users. The system achieves 91% accuracy and a cosine similarity score of 0.93 on standard questions, representing high performance in both information retrieval and natural language generation. This work contributes to digital public health by presenting a scalable, low-cost, and culturally adapted solution for improving donor engagement},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Blood donations continue to become a public health challenge worldwide, and in many countries, including Tanzania, due to low donor retention and limited public awareness. To address this, this study presents a multilingual conversational AI platform that encourages and promotes voluntary blood donation, developed with a hybrid architecture that integrates Retrieval Augmented Generation (RAG) LaBSE embeddings, Normalized Compression Distance (NCD), and Llama-based response generation. The chatbot provided accurate, bilingual responses (in English and Kiswahili) tailored for Tanzanian users. The system achieves 91% accuracy and a cosine similarity score of 0.93 on standard questions, representing high performance in both information retrieval and natural language generation. This work contributes to digital public health by presenting a scalable, low-cost, and culturally adapted solution for improving donor engagement
Nimo, Charles
Africa Health Check: Probing Cultural Bias in Medical LLMs Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{NimoCharles2025,
title = {Africa Health Check: Probing Cultural Bias in Medical LLMs},
author = {Charles Nimo},
url = {https://drive.google.com/drive/folders/12qOi6ZJI8ePvTMs9D5zYrB3wDbt245Wr?usp=sharing},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Large language models (LLMs) are increasingly deployed in global healthcare, yet their outputs often reflect Western-centric training data and omit indigenous medical systems and region-specific treatments. This study investigates cultural bias in instruction-tuned medical LLMs using a curated dataset of African Traditional Herbal Medicine. We evaluate model behavior across two complementary tasks, multiple-choice questions and fill-in-the-blank completions, designed to capture both treatment preferences and responsiveness to cultural context. To quantify outcome preferences and prompt influences, we apply two complementary metrics: Cultural Bias Score (CBS) and Cultural Bias Attribution (CBA). Our results show that while prompt adaptation can reduce inherent bias and enhance cultural alignment, models vary in how responsive they are to contextual guidance. Persistent default to allopathic (Western) treatments in zero-shot scenarios suggests that many biases remain embedded in model training. These findings underscore the need for culturally informed evaluation strategies to guide the development of AI systems that equitably serve diverse global health contexts. By releasing our dataset and providing a dual-metric evaluation approach, we offer practical tools for developing more culturally aware and clinically grounded AI systems for healthcare settings in the Global South.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Large language models (LLMs) are increasingly deployed in global healthcare, yet their outputs often reflect Western-centric training data and omit indigenous medical systems and region-specific treatments. This study investigates cultural bias in instruction-tuned medical LLMs using a curated dataset of African Traditional Herbal Medicine. We evaluate model behavior across two complementary tasks, multiple-choice questions and fill-in-the-blank completions, designed to capture both treatment preferences and responsiveness to cultural context. To quantify outcome preferences and prompt influences, we apply two complementary metrics: Cultural Bias Score (CBS) and Cultural Bias Attribution (CBA). Our results show that while prompt adaptation can reduce inherent bias and enhance cultural alignment, models vary in how responsive they are to contextual guidance. Persistent default to allopathic (Western) treatments in zero-shot scenarios suggests that many biases remain embedded in model training. These findings underscore the need for culturally informed evaluation strategies to guide the development of AI systems that equitably serve diverse global health contexts. By releasing our dataset and providing a dual-metric evaluation approach, we offer practical tools for developing more culturally aware and clinically grounded AI systems for healthcare settings in the Global South.
Leventhal, Michael
AI for Universal Literacy in the Languages People Speak Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{LeventhalMichael2025,
title = {AI for Universal Literacy in the Languages People Speak},
author = {Michael Leventhal},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Michael_Leventhal.pdf?generation=1755026549545263&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Overview of Malian language NLP projects at RobotsMali including creative and educational content using generative AI, a reading tutor App using a small ASR models, an LLM-based writing assistant, use of generative AI to create multi-modal digital skills, data science, and AI curriculum emphasizing orality, and related development projects including a national reading competition and development of AI curriculum for Malian universities.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Overview of Malian language NLP projects at RobotsMali including creative and educational content using generative AI, a reading tutor App using a small ASR models, an LLM-based writing assistant, use of generative AI to create multi-modal digital skills, data science, and AI curriculum emphasizing orality, and related development projects including a national reading competition and development of AI curriculum for Malian universities.
Onyango, Nelson
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{OnyangoNelson2025,
title = {A BI-DIRECTIONAL LONG SHORT-TERM MEMORY BASED DEEP LEARNING MODEL FOR POLITICAL HATE SPEECH DETECTION IN SWAHILI AND CODE-SWITCHED ENGLISH-SWAHILI TEXTUAL DATA},
author = {Nelson Onyango},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Nelson%20_Onyango.pdf?generation=1755026544679430&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {The study successfully curated a novel hate speech dataset composed of 19,369 tweets from three key datasets such as Politikweli, Hate_Speech_Kenya, and AfriSenti. These were cleaned, de-duplicated, language-labeled, and categorized into three primary classes: hate, offensive, and neither. Hate texts were further annotated across eight target categories including politics, ethnicity, gender, religion, social status, nationality, disability, and other. The dataset achieved a Randolph’s kappa of 0.566, indicating moderate agreement, and was validated through majority-voting and deep learning models such as Bi-LSTM together with SwahBERT and non-linear SVM models. Using GloVe, FastText, and TF-IDF embeddings, Bi-LSTM model achieved 92.5% accuracy and 90.1% F1-score, outperforming SwahBERT and SVM baselines. The research highlights the model’s effectiveness in handling code-switched language and contributes significantly to low-resource NLP.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
The study successfully curated a novel hate speech dataset composed of 19,369 tweets from three key datasets such as Politikweli, Hate_Speech_Kenya, and AfriSenti. These were cleaned, de-duplicated, language-labeled, and categorized into three primary classes: hate, offensive, and neither. Hate texts were further annotated across eight target categories including politics, ethnicity, gender, religion, social status, nationality, disability, and other. The dataset achieved a Randolph’s kappa of 0.566, indicating moderate agreement, and was validated through majority-voting and deep learning models such as Bi-LSTM together with SwahBERT and non-linear SVM models. Using GloVe, FastText, and TF-IDF embeddings, Bi-LSTM model achieved 92.5% accuracy and 90.1% F1-score, outperforming SwahBERT and SVM baselines. The research highlights the model’s effectiveness in handling code-switched language and contributes significantly to low-resource NLP.
Mayienga, Marlyn
Predicting Electoral Violence Using Integrated Conflict Data Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{MayiengaMarlyn2025,
title = {Predicting Electoral Violence Using Integrated Conflict Data},
author = {Marlyn Mayienga},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Marlyn_Mayienga.pdf?generation=1755026564366368&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Electoral violence severely undermines democratic stability, with approximately 60% of elections in fragile states experiencing conflict, from voter intimidation to lethal clashes
that claim thousands of lives. The 2007 Kenyan election, for instance, resulted in over 1,000 deaths and displaced 300,000 people, highlighting the urgent need for predictive
tools.[1].With over 50 countries facing elections in 2025, many in high-risk regions, this issue demands immediate attention. Current models often fail to capture electoral violence’s distinct temporal and spatial dynamics, such as clustering around election periods, limiting early warning capabilities. This study integrates institutional, economic, and event-based data from the Electoral Contention and Violence (ECAV) dataset, Uppsala Deadly Electoral Conflict Dataset (DECO), Varieties of Democracy (V-Dem), and World Bank indicators to forecast violence at the election-year-country level.
Using machine learning models such as Logistic Regression, K-Nearest Neighbors, XGBoost, Long Short-Term Memory (LSTM), and a Bayesian Ensemble, we achieve an Area
Under the Curve (AUC) of 0.8124 with the ensemble, outperforming baselines like Logistic Regression (AUC 0.61). Analysis reveals 67% of incidents occur within 30 days of
elections, with institutional factors like electoral intimidation and social group exclusion
as key predictors. These findings enable targeted interventions, such as enhanced election monitoring by electoral bodies and international organizations, enhancing violence prevention in high-risk 2025 elections.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Electoral violence severely undermines democratic stability, with approximately 60% of elections in fragile states experiencing conflict, from voter intimidation to lethal clashes
that claim thousands of lives. The 2007 Kenyan election, for instance, resulted in over 1,000 deaths and displaced 300,000 people, highlighting the urgent need for predictive
tools.[1].With over 50 countries facing elections in 2025, many in high-risk regions, this issue demands immediate attention. Current models often fail to capture electoral violence’s distinct temporal and spatial dynamics, such as clustering around election periods, limiting early warning capabilities. This study integrates institutional, economic, and event-based data from the Electoral Contention and Violence (ECAV) dataset, Uppsala Deadly Electoral Conflict Dataset (DECO), Varieties of Democracy (V-Dem), and World Bank indicators to forecast violence at the election-year-country level.
Using machine learning models such as Logistic Regression, K-Nearest Neighbors, XGBoost, Long Short-Term Memory (LSTM), and a Bayesian Ensemble, we achieve an Area
Under the Curve (AUC) of 0.8124 with the ensemble, outperforming baselines like Logistic Regression (AUC 0.61). Analysis reveals 67% of incidents occur within 30 days of
elections, with institutional factors like electoral intimidation and social group exclusion
as key predictors. These findings enable targeted interventions, such as enhanced election monitoring by electoral bodies and international organizations, enhancing violence prevention in high-risk 2025 elections.
that claim thousands of lives. The 2007 Kenyan election, for instance, resulted in over 1,000 deaths and displaced 300,000 people, highlighting the urgent need for predictive
tools.[1].With over 50 countries facing elections in 2025, many in high-risk regions, this issue demands immediate attention. Current models often fail to capture electoral violence’s distinct temporal and spatial dynamics, such as clustering around election periods, limiting early warning capabilities. This study integrates institutional, economic, and event-based data from the Electoral Contention and Violence (ECAV) dataset, Uppsala Deadly Electoral Conflict Dataset (DECO), Varieties of Democracy (V-Dem), and World Bank indicators to forecast violence at the election-year-country level.
Using machine learning models such as Logistic Regression, K-Nearest Neighbors, XGBoost, Long Short-Term Memory (LSTM), and a Bayesian Ensemble, we achieve an Area
Under the Curve (AUC) of 0.8124 with the ensemble, outperforming baselines like Logistic Regression (AUC 0.61). Analysis reveals 67% of incidents occur within 30 days of
elections, with institutional factors like electoral intimidation and social group exclusion
as key predictors. These findings enable targeted interventions, such as enhanced election monitoring by electoral bodies and international organizations, enhancing violence prevention in high-risk 2025 elections.
Ochieng, Ronnie
INTELLIGENT DIGITAL STETHOSCOPE FOR AUTOMATED LUNG SOUND ANALYSIS AND RESPIRATORY DISEASE CLASSIFICATION Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{OchiengRonnie2025,
title = {INTELLIGENT DIGITAL STETHOSCOPE FOR AUTOMATED LUNG SOUND ANALYSIS AND RESPIRATORY DISEASE CLASSIFICATION},
author = {Ronnie Ochieng},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Ronnie_Ochieng.pdf?generation=1755026603209913&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Respiratory disorders significantly impact global health, accounting for approximately 6.97% of all deaths worldwide. Chronic obstructive pulmonary disease (COPD) alone caused 3.5 million deaths in 2021, representing about 5% of all global deaths. Notably, nearly 90% of COPD deaths in individuals under 70 occur in low-and middle-income countries (LMICs), where limited access to healthcare and diagnostic tools, such as chest X-rays, often necessitates reliance on clinical expertise and auscultation for diagnosis. However, accurate interpretation of lung sounds requires specialized training and traditional stethoscopes usually present challenges due to low signal levels and interference from bodily noises. These factors contribute to potential misdiagnosis and underdiagnosis, exacerbating the burden of respiratory diseases in LMICs.
This project presents an intelligent digital stethoscope integrated with advanced machine-learning algorithms to analyze lung sounds and classify respiratory diseases. Using a Boya M1 microphone, the device captures the patient's respiratory sounds and leverages machine-learning algorithms to classify lung sounds such as crackles or wheezing and identify patterns indicative of common lung diseases like chronic obstructive pulmonary disease (COPD) and pneumonia. The classification results and recorded sounds are displayed in real-time on the device’s screen, aiding healthcare providers, especially in resource-limited settings, to make informed diagnoses, facilitating early diagnosis and management of respiratory conditions.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Respiratory disorders significantly impact global health, accounting for approximately 6.97% of all deaths worldwide. Chronic obstructive pulmonary disease (COPD) alone caused 3.5 million deaths in 2021, representing about 5% of all global deaths. Notably, nearly 90% of COPD deaths in individuals under 70 occur in low-and middle-income countries (LMICs), where limited access to healthcare and diagnostic tools, such as chest X-rays, often necessitates reliance on clinical expertise and auscultation for diagnosis. However, accurate interpretation of lung sounds requires specialized training and traditional stethoscopes usually present challenges due to low signal levels and interference from bodily noises. These factors contribute to potential misdiagnosis and underdiagnosis, exacerbating the burden of respiratory diseases in LMICs.
This project presents an intelligent digital stethoscope integrated with advanced machine-learning algorithms to analyze lung sounds and classify respiratory diseases. Using a Boya M1 microphone, the device captures the patient's respiratory sounds and leverages machine-learning algorithms to classify lung sounds such as crackles or wheezing and identify patterns indicative of common lung diseases like chronic obstructive pulmonary disease (COPD) and pneumonia. The classification results and recorded sounds are displayed in real-time on the device’s screen, aiding healthcare providers, especially in resource-limited settings, to make informed diagnoses, facilitating early diagnosis and management of respiratory conditions.
This project presents an intelligent digital stethoscope integrated with advanced machine-learning algorithms to analyze lung sounds and classify respiratory diseases. Using a Boya M1 microphone, the device captures the patient's respiratory sounds and leverages machine-learning algorithms to classify lung sounds such as crackles or wheezing and identify patterns indicative of common lung diseases like chronic obstructive pulmonary disease (COPD) and pneumonia. The classification results and recorded sounds are displayed in real-time on the device’s screen, aiding healthcare providers, especially in resource-limited settings, to make informed diagnoses, facilitating early diagnosis and management of respiratory conditions.
Mnyawami, Yuda
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{MnyawamiYuda2025,
title = {Predictive System for Characterizing Student Dropout using K-Nearest Oracle with Automated Machine Learning (KNORA-AutoML) Model},
author = {Yuda Mnyawami},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Yuda_Mnyawami.pdf?generation=1755026602458589&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Students continue to drop out of basic education in developing countries because student dropout features are dynamic. It is challenging to reduce student dropout in developing countries because factors keep changing periodically. Conventional machine learning models have been used to address the persisting problem, but still, student dropout in developing countries, particularly Tanzania is taking place. Conventional machine learning models can not accurately determine features leading to student dropouts. This study used the Twaweza information repository to establish a suitable dataset using the KNORA-AutoML prediction model. The KNORA-AutoML model demonstrated 97% accuracy and 87% AUC, outperforming previous studies. The suggested model was used to develop the predictive system. Results reveal that students who drop out due to the number of household children walk a long distance to school in rural areas. In urban areas, students travel more than 11 kilometers to school, which makes them fail to do their homework accurately. Parents’ occupations, particularly housewives with more than five children, have a likelihood of not supporting their children. Features such as distance, household size, household children, modes of transport, and parents' occupations highly contribute to student dropouts. The proposed predictive system reveals its effectiveness in identifying students at risk of dropping out and proposes early interventions.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Students continue to drop out of basic education in developing countries because student dropout features are dynamic. It is challenging to reduce student dropout in developing countries because factors keep changing periodically. Conventional machine learning models have been used to address the persisting problem, but still, student dropout in developing countries, particularly Tanzania is taking place. Conventional machine learning models can not accurately determine features leading to student dropouts. This study used the Twaweza information repository to establish a suitable dataset using the KNORA-AutoML prediction model. The KNORA-AutoML model demonstrated 97% accuracy and 87% AUC, outperforming previous studies. The suggested model was used to develop the predictive system. Results reveal that students who drop out due to the number of household children walk a long distance to school in rural areas. In urban areas, students travel more than 11 kilometers to school, which makes them fail to do their homework accurately. Parents’ occupations, particularly housewives with more than five children, have a likelihood of not supporting their children. Features such as distance, household size, household children, modes of transport, and parents' occupations highly contribute to student dropouts. The proposed predictive system reveals its effectiveness in identifying students at risk of dropping out and proposes early interventions.
Okewunmi, Paul
Evaluating Robustness of LLMs to Typographical Noise in Yorùbá QA Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{OkewunmiPaul2025,
title = {Evaluating Robustness of LLMs to Typographical Noise in Yorùbá QA},
author = {Paul Okewunmi},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Paul_Okewunmi%20.pdf?generation=1755026564101424&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Generative AI models are primarily accessed through chat interfaces, where user queries often contain typographical errors. While these models perform well in English, their robustness to noisy inputs in low-resource languages like Yorùbá remains underexplored. This work investigates a Yorùbá question-answering (QA) task by introducing synthetic typographical noise into clean inputs. We design a probabilistic noise injection strategy that simulates realistic human typos. In our experiments, each character in a clean sentence is independently altered, with noise levels ranging from 10% to 40%. We evaluate performance across three strong multilingual models using two complementary metrics: (1) a multilingual BERTScore to assess semantic similarity between outputs on clean and noisy inputs, and (2) an LLM-as-judge approach, where the best Yorùbá-capable model rates fluency, comprehension, and accuracy on a 1–5 scale. Results show that while English QA performance degrades gradually, Yorùbá QA suffers a sharper decline. At 40% noise, GPT-4o experiences over a 50% drop in comprehension ability, with similar declines for Gemini 2.0 Flash and Claude 3.7 Sonnet. We conclude with recommendations for noise-aware training and dedicated noisy Yorùbá benchmarks to enhance LLM robustness in low-resource settings.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Generative AI models are primarily accessed through chat interfaces, where user queries often contain typographical errors. While these models perform well in English, their robustness to noisy inputs in low-resource languages like Yorùbá remains underexplored. This work investigates a Yorùbá question-answering (QA) task by introducing synthetic typographical noise into clean inputs. We design a probabilistic noise injection strategy that simulates realistic human typos. In our experiments, each character in a clean sentence is independently altered, with noise levels ranging from 10% to 40%. We evaluate performance across three strong multilingual models using two complementary metrics: (1) a multilingual BERTScore to assess semantic similarity between outputs on clean and noisy inputs, and (2) an LLM-as-judge approach, where the best Yorùbá-capable model rates fluency, comprehension, and accuracy on a 1–5 scale. Results show that while English QA performance degrades gradually, Yorùbá QA suffers a sharper decline. At 40% noise, GPT-4o experiences over a 50% drop in comprehension ability, with similar declines for Gemini 2.0 Flash and Claude 3.7 Sonnet. We conclude with recommendations for noise-aware training and dedicated noisy Yorùbá benchmarks to enhance LLM robustness in low-resource settings.
Ochieng, Millicent
Benchmarking LLMs: From Standard NLP Tasks to Real-World Multilingual Challenges Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{OchiengMillicent2025,
title = {Benchmarking LLMs: From Standard NLP Tasks to Real-World Multilingual Challenges},
author = {Millicent Ochieng},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Millicent_Ochieng.pdf?generation=1755026562060332&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {The effectiveness of large language models (LLMs) in multicultural, multilingual, and low-resource real-world settings remains largely underexplored. While benchmark metrics like F1 scores provide a superficial measure of performance, they fail to capture critical cultural and contextual understanding. This study evaluates seven leading LLMs, including GPT-4, Mistral-7b, and Llama-2-70b, on a dataset comprising real-world WhatsApp conversations featuring English, Swahili, and Sheng. Through both quantitative (F1 scores) and qualitative (explanation quality) analyses, we reveal that despite strong metric performance, models often misunderstand cultural nuances, code-mixing, and contextual sentiment. Our results suggest that human-centered evaluation is crucial for responsible and effective deployment of LLMs in African and other low-resource settings.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
The effectiveness of large language models (LLMs) in multicultural, multilingual, and low-resource real-world settings remains largely underexplored. While benchmark metrics like F1 scores provide a superficial measure of performance, they fail to capture critical cultural and contextual understanding. This study evaluates seven leading LLMs, including GPT-4, Mistral-7b, and Llama-2-70b, on a dataset comprising real-world WhatsApp conversations featuring English, Swahili, and Sheng. Through both quantitative (F1 scores) and qualitative (explanation quality) analyses, we reveal that despite strong metric performance, models often misunderstand cultural nuances, code-mixing, and contextual sentiment. Our results suggest that human-centered evaluation is crucial for responsible and effective deployment of LLMs in African and other low-resource settings.
Sowole, Oladimeji Samuel Sowole
Integrating Network Curvature into Epidemic Dynamics: A Curvature-Aware SIR Model Framework Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{SowoleOladimejiSamuelSowole2025,
title = {Integrating Network Curvature into Epidemic Dynamics: A Curvature-Aware SIR Model Framework},
author = {Oladimeji Samuel Sowole Sowole},
url = {https://drive.google.com/file/d/1fA2dCCbbdgCNUjVSSphTlCH5d3_7ag3d/view?usp=sharing},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Rabothata, Moyahabo
Supervised Machine Learning and Deep Learning Techchniques for Legal Text Classification Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{RabothataMoyahabo2025,
title = {Supervised Machine Learning and Deep Learning Techchniques for Legal Text Classification},
author = {Moyahabo Rabothata},
url = {https://drive.google.com/file/d/10jWrjh9iuF_5R0zznwy6gPMGFsHLRhdm/view?usp=drivesdk},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Rakotondranisa, Onintsoa Anjara
Constraining the Hubble Tension with Fast Radio Bursts using Machine Learning Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{AnjaraRakotondranisaOnintsoa2025,
title = {Constraining the Hubble Tension with Fast Radio Bursts using Machine Learning},
author = {Onintsoa Anjara Rakotondranisa},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Onintsoa_Anjara%20Rakotondranisa.pdf?generation=1755026592117036&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Fast Radio Bursts (FRBs) are transient events used as valuable tools to probe the cosmic expansion of the Universe. The main property containing the cosmological information is in the dispersion measure from the intergalactic medium DMIGM. In this work, we aim to constrain the Hubble tension by estimating the Hubble constant H0, using FRB data, without assuming any specific cosmological model. As a baseline, we used three machine learning models to predict〈DMIGM〉. Then, the Hubble parameter H(z) was subsequently reconstructed. We found that the ANN gives the most reliable H0 estimation, which is consistent with late-time measurements. We also found a non-physical offset of 100 pc cm−3 at z = 0. Additional experiments were conducted to refine our results. Our findings suggest that Physics-Informed Neural Networks could potentially provide more reliable estimations.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Fast Radio Bursts (FRBs) are transient events used as valuable tools to probe the cosmic expansion of the Universe. The main property containing the cosmological information is in the dispersion measure from the intergalactic medium DMIGM. In this work, we aim to constrain the Hubble tension by estimating the Hubble constant H0, using FRB data, without assuming any specific cosmological model. As a baseline, we used three machine learning models to predict〈DMIGM〉. Then, the Hubble parameter H(z) was subsequently reconstructed. We found that the ANN gives the most reliable H0 estimation, which is consistent with late-time measurements. We also found a non-physical offset of 100 pc cm−3 at z = 0. Additional experiments were conducted to refine our results. Our findings suggest that Physics-Informed Neural Networks could potentially provide more reliable estimations.
Rabothata, Moyahabo Muriel
Network Analysis and Topic Modelling to Identify Influential Authors Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{RabothataMoyahaboMuriel2025,
title = {Network Analysis and Topic Modelling to Identify Influential Authors},
author = {Moyahabo Muriel Rabothata},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Moyahabo%20Muriel_Rabothata.pdf?generation=1755026588075740&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {In assisting the University of Pretoria faculty of private law to understand the underlying insight of the Fontes Juris data consisting of law sources noted in South African court cases from 1825 to 2015. The project aimed at using machine learning (ML) techniques to assess the impact of Law Research in South African (SA) courts. The project employed Topic modelling to identify topics in the corpus and used Network Analysis to model author and judgment pair and author and topic pair using various centrality measures. Degree of centrality measure has shown interesting results emerging from the data indicating most influential authors. Topic modelling has revealed the decrease in publications cited in court although they are not topic specific. Topic modelling also shows large influence of Old Laws, Roman-Dutch and English Laws on their mission, the project aims at using machine learning (ML) techniques to assess the impact of Law Research in South African (SA) courts. The project employed Topic modelling to identify topics in the corpus and used Network Analysis to model author and judgment pair and author and topic pair using various centrality measures. Degree of centrality measure has shown interesting results emerging from the data indicating most influential authors. Topic modelling has revealed the decrease in publications cited in court although they are not topic specific. Topic modelling also shows large influence of Old Laws, Roman-Dutch and English Laws.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
In assisting the University of Pretoria faculty of private law to understand the underlying insight of the Fontes Juris data consisting of law sources noted in South African court cases from 1825 to 2015. The project aimed at using machine learning (ML) techniques to assess the impact of Law Research in South African (SA) courts. The project employed Topic modelling to identify topics in the corpus and used Network Analysis to model author and judgment pair and author and topic pair using various centrality measures. Degree of centrality measure has shown interesting results emerging from the data indicating most influential authors. Topic modelling has revealed the decrease in publications cited in court although they are not topic specific. Topic modelling also shows large influence of Old Laws, Roman-Dutch and English Laws on their mission, the project aims at using machine learning (ML) techniques to assess the impact of Law Research in South African (SA) courts. The project employed Topic modelling to identify topics in the corpus and used Network Analysis to model author and judgment pair and author and topic pair using various centrality measures. Degree of centrality measure has shown interesting results emerging from the data indicating most influential authors. Topic modelling has revealed the decrease in publications cited in court although they are not topic specific. Topic modelling also shows large influence of Old Laws, Roman-Dutch and English Laws.
Ooko, Samson
TinyML-Based Acoustic Bird Detection for Crop Protection in African Farms Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{OokoSamson2025,
title = {TinyML-Based Acoustic Bird Detection for Crop Protection in African Farms},
author = {Samson Ooko},
url = {https://drive.google.com/file/d/1dPdlA3xJ0PswwMp9_2DhygJgZ1-xiiYS/view?usp=sharing},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Musinguzi, Denis
PaliGemma-CXR: Multitask Multimodal for Chest X-ray Interpretation Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{MusinguziDenis2025,
title = {PaliGemma-CXR: Multitask Multimodal for Chest X-ray Interpretation},
author = {Denis Musinguzi},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Denis_Musinguzi.pdf?generation=1755026591481830&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Tuberculosis is a significant global health challenge. Chest Xrays play a critical role in TB screening, yet many countries face shortage of radiologists capable of interpreting these images. In addition, the interpretations are prone to variability. Machine learning has shown promise as an alternative. However, traditional machine learning approaches rely on task-specific models, which fail to utilize the interdependence between medical image interpretation tasks. Moreover, learning multiple tasks within the same model improves generalization.We propose PaliGemma- CXR, a multi-task multimodal model that jointly learns TB diagnosis, object detection, segmentation, report generation, and VQA in a supervised manner. Starting with a dataset of chest X-ray images annotated with diagnosis labels and segmentation masks, we curated multimodal datasets for detection, report generation and VQA. Our experiments with the dataset confirm the effectiveness of PaliGemma-CXR in performing all the 5 tasks over training individual models for each task.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Tuberculosis is a significant global health challenge. Chest Xrays play a critical role in TB screening, yet many countries face shortage of radiologists capable of interpreting these images. In addition, the interpretations are prone to variability. Machine learning has shown promise as an alternative. However, traditional machine learning approaches rely on task-specific models, which fail to utilize the interdependence between medical image interpretation tasks. Moreover, learning multiple tasks within the same model improves generalization.We propose PaliGemma- CXR, a multi-task multimodal model that jointly learns TB diagnosis, object detection, segmentation, report generation, and VQA in a supervised manner. Starting with a dataset of chest X-ray images annotated with diagnosis labels and segmentation masks, we curated multimodal datasets for detection, report generation and VQA. Our experiments with the dataset confirm the effectiveness of PaliGemma-CXR in performing all the 5 tasks over training individual models for each task.
Oyelaja, Iremide
SECURE AND SCALABLE HORIZONTAL FEDERATED LEARNING FOR BANK FRAUD DETECTION Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{OyelajaIremide2025,
title = {SECURE AND SCALABLE HORIZONTAL FEDERATED LEARNING FOR BANK FRAUD DETECTION},
author = {Iremide Oyelaja},
url = {https://drive.google.com/file/d/1AgEXywzgjijyIwFTglqqxiDeaFyepTBb/view?usp=sharing},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Akanni, Comfort
AI-BASED EARLY DETECTION OF DEVELOPMENTAL DELAYS IN CHILDREN WITH SICKLE CELL DISEASE (AGES 0–5) Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{AkanniComfort2025,
title = {AI-BASED EARLY DETECTION OF DEVELOPMENTAL DELAYS IN CHILDREN WITH SICKLE CELL DISEASE (AGES 0–5)},
author = {Comfort Akanni},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Comfort_Akanni.pdf?generation=1755026596315997&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Children under the age of 5 years with Sickle Cell Disease (SCD) are at high risk of having developmental delay. This is of a major concern because the delay is due to the frequent episodes of chronic anemia, crises and cerebrovascular complications on cognitive and motor development. Early detection is essential for timely intervention but the current screening methods lack predictive accuracy and accessibility.
This study aims to develop and evaluate DevSickleNet, a multimodal deep learning model designed to predict developmental delays in children with Sickle Cell Disease (SCD) using clinical, caregiver, and developmental milestone data.
This study proposes DevSickleNet, a multimodal deep learning model that integrates clinical, caregiver, and milestone time-series data to predict developmental delays in children with SCD. A synthetic dataset of 300 samples was generated. This incorporates clinical features (hemoglobin levels, pain crises history,stroke history), caregivers factors (educational factors, socioeconomic status), and developmental milestones progression over a span of 12 months. These features were encoded and normalized using standard preprocessing pipelines. The multimodal data was input into DevSickleNet, which combines an LSTM network for milestone time-series, fully connected layers for static inputs, and an attention-based fusion layer. The model was trained using cross-entropy loss and evaluated using accuracy, precision, recall, F1-score, and ROC-AUC.
The DevSickleNet achieved an accuracy of 85.4, a precision of 82.9, a recall of 79.5, a F1-score of 81.1 and a ROC-AUC of 0.87, outperforming traditional models like Logistic Regression, Random Forest and XGBoost. The feature importance analysis identified pain crises frequency, levels of hemoglobin and the educational status of the caregivers as the key predictors of developmental delays. The model’s performance is attributed to its ability to process sequential milestone data through LSTM and combine it with static clinical and caregiver features using attention-based fusion. This architecture allowed DevSickleNet to capture both temporal progression and static risk factors effectively.
These results highlight the potential of AI-driven multimodal learning for the early developmental delay screening in SCD patients. Even though the findings are promising, there is still a need to validate the model using a clinical dataset gotten from tertiary healthcare facilities in order to confirm its clinical applicability. DevSickleNet therefore provides for AI-powered early detection of developmental delay in SCD children and intervention strategies aiming to improve their developmental outcome.
Keywords: Sickle Cell Disease, Developmental delays, Deep learning, DevSickleNet, Pediatric AI, Early detection, Multimodal AI.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Children under the age of 5 years with Sickle Cell Disease (SCD) are at high risk of having developmental delay. This is of a major concern because the delay is due to the frequent episodes of chronic anemia, crises and cerebrovascular complications on cognitive and motor development. Early detection is essential for timely intervention but the current screening methods lack predictive accuracy and accessibility.
This study aims to develop and evaluate DevSickleNet, a multimodal deep learning model designed to predict developmental delays in children with Sickle Cell Disease (SCD) using clinical, caregiver, and developmental milestone data.
This study proposes DevSickleNet, a multimodal deep learning model that integrates clinical, caregiver, and milestone time-series data to predict developmental delays in children with SCD. A synthetic dataset of 300 samples was generated. This incorporates clinical features (hemoglobin levels, pain crises history,stroke history), caregivers factors (educational factors, socioeconomic status), and developmental milestones progression over a span of 12 months. These features were encoded and normalized using standard preprocessing pipelines. The multimodal data was input into DevSickleNet, which combines an LSTM network for milestone time-series, fully connected layers for static inputs, and an attention-based fusion layer. The model was trained using cross-entropy loss and evaluated using accuracy, precision, recall, F1-score, and ROC-AUC.
The DevSickleNet achieved an accuracy of 85.4, a precision of 82.9, a recall of 79.5, a F1-score of 81.1 and a ROC-AUC of 0.87, outperforming traditional models like Logistic Regression, Random Forest and XGBoost. The feature importance analysis identified pain crises frequency, levels of hemoglobin and the educational status of the caregivers as the key predictors of developmental delays. The model’s performance is attributed to its ability to process sequential milestone data through LSTM and combine it with static clinical and caregiver features using attention-based fusion. This architecture allowed DevSickleNet to capture both temporal progression and static risk factors effectively.
These results highlight the potential of AI-driven multimodal learning for the early developmental delay screening in SCD patients. Even though the findings are promising, there is still a need to validate the model using a clinical dataset gotten from tertiary healthcare facilities in order to confirm its clinical applicability. DevSickleNet therefore provides for AI-powered early detection of developmental delay in SCD children and intervention strategies aiming to improve their developmental outcome.
Keywords: Sickle Cell Disease, Developmental delays, Deep learning, DevSickleNet, Pediatric AI, Early detection, Multimodal AI.
This study aims to develop and evaluate DevSickleNet, a multimodal deep learning model designed to predict developmental delays in children with Sickle Cell Disease (SCD) using clinical, caregiver, and developmental milestone data.
This study proposes DevSickleNet, a multimodal deep learning model that integrates clinical, caregiver, and milestone time-series data to predict developmental delays in children with SCD. A synthetic dataset of 300 samples was generated. This incorporates clinical features (hemoglobin levels, pain crises history,stroke history), caregivers factors (educational factors, socioeconomic status), and developmental milestones progression over a span of 12 months. These features were encoded and normalized using standard preprocessing pipelines. The multimodal data was input into DevSickleNet, which combines an LSTM network for milestone time-series, fully connected layers for static inputs, and an attention-based fusion layer. The model was trained using cross-entropy loss and evaluated using accuracy, precision, recall, F1-score, and ROC-AUC.
The DevSickleNet achieved an accuracy of 85.4, a precision of 82.9, a recall of 79.5, a F1-score of 81.1 and a ROC-AUC of 0.87, outperforming traditional models like Logistic Regression, Random Forest and XGBoost. The feature importance analysis identified pain crises frequency, levels of hemoglobin and the educational status of the caregivers as the key predictors of developmental delays. The model’s performance is attributed to its ability to process sequential milestone data through LSTM and combine it with static clinical and caregiver features using attention-based fusion. This architecture allowed DevSickleNet to capture both temporal progression and static risk factors effectively.
These results highlight the potential of AI-driven multimodal learning for the early developmental delay screening in SCD patients. Even though the findings are promising, there is still a need to validate the model using a clinical dataset gotten from tertiary healthcare facilities in order to confirm its clinical applicability. DevSickleNet therefore provides for AI-powered early detection of developmental delay in SCD children and intervention strategies aiming to improve their developmental outcome.
Keywords: Sickle Cell Disease, Developmental delays, Deep learning, DevSickleNet, Pediatric AI, Early detection, Multimodal AI.
Mba, Patience
Estimation of physico-chemical properties of soil using machine learning Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{MbaPatience2025,
title = {Estimation of physico-chemical properties of soil using machine learning},
author = {Patience Mba},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Patience_Mba.pdf?generation=1755026584622161&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Soil quality assessments are essential for agricultural productivity, yet traditional methods are often slow and expensive. This study presents a machine learning-based approach for predicting the physical and chemical properties of soil using RGB images. A dataset of 1388 soil samples collected from Benue State, Nigeria, was analysed using texture features extracted via GLCM and Gabor filters. Multiple models, including SVR, CNN, an optimized CNN, and a hybrid ML-CNN stack, were evaluated. The ML-CNN stack showed superior performance, achieving high predictive accuracy with reduced error rates. These findings demonstrate the feasibility of using ML techniques for efficient, low-cost soil analysis.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Soil quality assessments are essential for agricultural productivity, yet traditional methods are often slow and expensive. This study presents a machine learning-based approach for predicting the physical and chemical properties of soil using RGB images. A dataset of 1388 soil samples collected from Benue State, Nigeria, was analysed using texture features extracted via GLCM and Gabor filters. Multiple models, including SVR, CNN, an optimized CNN, and a hybrid ML-CNN stack, were evaluated. The ML-CNN stack showed superior performance, achieving high predictive accuracy with reduced error rates. These findings demonstrate the feasibility of using ML techniques for efficient, low-cost soil analysis.
Talotsing, Gaelle Patricia
A stacking Ensemble Machine Learning Model for Emergency Call Forecasting Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{TalotsingGaellePatricia2025,
title = {A stacking Ensemble Machine Learning Model for Emergency Call Forecasting},
author = {Gaelle Patricia Talotsing},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Gaelle%20Patricia_Talotsing.pdf?generation=1755026579961972&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {One of the greatest challenges of Emergency medical services providers is to handle the large number of Emergency Medical Service (EMS) calls coming from the population. An accurate forecast of EMS calls is involved in ambulance fleet dispatching and routing to minimize response times to emergency calls and enhance the efficacy of assistance. Yet, the demand for emergency services exhibits significant variability, posing a challenge in accurately predicting the future occurrence of emergency calls and their spatial-temporal distribution. Here, we propose a stacking ensemble machine learning model to forecast EMS calls, combining different base learners to enhance the overall performance of generalization. Additionally, we conducted experiments using Boruta, Lasso, RFFI and SHAP feature selection methods to identify the most informative attributes from the EMS dataset. The proposed ensemble model integrates a base layer and a meta layer. In the base layer, we applied four base learners: Decision Tree, Gradient Boosting Regression Tree, Light Gradient Boosting Machine and Random Forest. In the meta layer, we used an optimized Random Forest model to integrate the outputs of base learners. We evaluate the performance of our proposed model using the R2 -score and four different error metrics. Based on a real data set including spatial, temporal and weather features, the findings of this study demonstrated that the proposed stacking-based ensemble model showed a better score and the minimum errors compared to the traditional single algorithms, online machine learning methods and voting ensemble methods. We achieved a higher score of 0.9954, mse of 0.8938, rmse of 0.9454, mae of 0.2923 and mape of 0.0724 compared to state-of-the-art models. This work is an aid for emergency managers in making well-informed decisions, improving outcomes for ambulance dispatch and routing, and enhancing ambulance response time.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
One of the greatest challenges of Emergency medical services providers is to handle the large number of Emergency Medical Service (EMS) calls coming from the population. An accurate forecast of EMS calls is involved in ambulance fleet dispatching and routing to minimize response times to emergency calls and enhance the efficacy of assistance. Yet, the demand for emergency services exhibits significant variability, posing a challenge in accurately predicting the future occurrence of emergency calls and their spatial-temporal distribution. Here, we propose a stacking ensemble machine learning model to forecast EMS calls, combining different base learners to enhance the overall performance of generalization. Additionally, we conducted experiments using Boruta, Lasso, RFFI and SHAP feature selection methods to identify the most informative attributes from the EMS dataset. The proposed ensemble model integrates a base layer and a meta layer. In the base layer, we applied four base learners: Decision Tree, Gradient Boosting Regression Tree, Light Gradient Boosting Machine and Random Forest. In the meta layer, we used an optimized Random Forest model to integrate the outputs of base learners. We evaluate the performance of our proposed model using the R2 -score and four different error metrics. Based on a real data set including spatial, temporal and weather features, the findings of this study demonstrated that the proposed stacking-based ensemble model showed a better score and the minimum errors compared to the traditional single algorithms, online machine learning methods and voting ensemble methods. We achieved a higher score of 0.9954, mse of 0.8938, rmse of 0.9454, mae of 0.2923 and mape of 0.0724 compared to state-of-the-art models. This work is an aid for emergency managers in making well-informed decisions, improving outcomes for ambulance dispatch and routing, and enhancing ambulance response time.
Falola, Peace
From Play to Preservation: KọÈdè– Where Tech Meets Language Learning Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{FalolaPeace2025,
title = {From Play to Preservation: KọÈdè– Where Tech Meets Language Learning},
author = {Peace Falola},
url = {https://drive.google.com/file/d/1JQQtxSgl_lUyJrHdHD1XzusaGDWYpIje/view?usp=drive_link},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {This research introduces a high-quality, multi-domain Named Entity Recognition (NER) dataset for the Yoruba language, aimed at expanding NLP resources for low-resource African languages. The dataset consists of 4,767 sentences and over 100,000 tokens, annotated across five domains: blogs, Bible, movies, radio broadcasts, and Wikipedia. Each sentence was labeled by three native Yoruba speakers following consistent guidelines. The dataset supports three entity types: Person, Location, and Organization. Inter-annotator agreement scores (Fleiss’ Kappa) were consistently high across all domains, with no disagreements from entity type mismatch. This dataset fills a critical gap in Yoruba NLP and enables research in cross-domain evaluation, domain-adaptive modeling, and the development of baseline and unified NER systems.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
This research introduces a high-quality, multi-domain Named Entity Recognition (NER) dataset for the Yoruba language, aimed at expanding NLP resources for low-resource African languages. The dataset consists of 4,767 sentences and over 100,000 tokens, annotated across five domains: blogs, Bible, movies, radio broadcasts, and Wikipedia. Each sentence was labeled by three native Yoruba speakers following consistent guidelines. The dataset supports three entity types: Person, Location, and Organization. Inter-annotator agreement scores (Fleiss’ Kappa) were consistently high across all domains, with no disagreements from entity type mismatch. This dataset fills a critical gap in Yoruba NLP and enables research in cross-domain evaluation, domain-adaptive modeling, and the development of baseline and unified NER systems.
Falola, Peace
A Multi-Domain Annotated Yoruba NER Dataset: Expanding NLP Resources for Low-Resource African Languages Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{FalolaPeace2025b,
title = {A Multi-Domain Annotated Yoruba NER Dataset: Expanding NLP Resources for Low-Resource African Languages},
author = {Peace Falola},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Peace%20_Falola.pdf?generation=1755026581717566&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {This research introduces a high-quality, multi-domain Named Entity Recognition (NER) dataset for the Yoruba language, aimed at expanding NLP resources for low-resource African languages. The dataset consists of 4,767 sentences and over 100,000 tokens, annotated across five domains: blogs, Bible, movies, radio broadcasts, and Wikipedia. Each sentence was labeled by three native Yoruba speakers following consistent guidelines. The dataset supports three entity types: Person, Location, and Organization. Inter-annotator agreement scores (Fleiss’ Kappa) were consistently high across all domains, with no disagreements from entity type mismatch. This dataset fills a critical gap in Yoruba NLP and enables research in cross-domain evaluation, domain-adaptive modeling, and the development of baseline and unified NER systems.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
This research introduces a high-quality, multi-domain Named Entity Recognition (NER) dataset for the Yoruba language, aimed at expanding NLP resources for low-resource African languages. The dataset consists of 4,767 sentences and over 100,000 tokens, annotated across five domains: blogs, Bible, movies, radio broadcasts, and Wikipedia. Each sentence was labeled by three native Yoruba speakers following consistent guidelines. The dataset supports three entity types: Person, Location, and Organization. Inter-annotator agreement scores (Fleiss’ Kappa) were consistently high across all domains, with no disagreements from entity type mismatch. This dataset fills a critical gap in Yoruba NLP and enables research in cross-domain evaluation, domain-adaptive modeling, and the development of baseline and unified NER systems.
Silima, Walter
Machine Learning Approaches to Study Star Formation and Black Hole Accretion in the MeerKAT/MIGHTEE survey Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{SilimaWalter2025,
title = {Machine Learning Approaches to Study Star Formation and Black Hole Accretion in the MeerKAT/MIGHTEE survey},
author = {Walter Silima},
url = {https://docs.google.com/presentation/d/1uHVLsoGUuRCFyOT1W5qXfNJHsSgBUIHkoz3iYdfkMn8/edit?usp=sharing},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Radio synchrotron emission originates from both massive star formation and black hole accretion, two processes that drive galaxy evolution. Therefore, current high-sensitivity and wide-field extragalactic radio continuum surveys require efficient and reliable classification of radio sources dominated by star formation or black hole accretion before utilizing radio continuum for exploring cosmic evolution. In this study, we implement, optimize, and compare five widely used supervised machine-learning (ML) algorithms to classify radio sources detected in the MeerKAT International GHz Tiered Extragalactic Exploration (MIGHTEE)–COSMOS survey as star-forming galaxies (SFGs) and active galactic nuclei (AGN). We utilize conventionally classified SFGs and AGN from MIGHTEE-COSMOS to construct training and test sets for evaluating the ML algorithm’s performance. To select input features for our ML analyses, we incorporate 18 physical parameters of MIGHTEE-detected radio sources. As anticipated, our feature analyses rank the six parameters used in conventional classification as the most effective: the infrared-radio correlation parameter (qIR), the optical compactness morphology parameter (class_star), three combined mid-infrared colors, and stellar mass. By optimizing the ML models with these selected features and testing classifiers across various feature combinations, we find that model performance generally improves as additional features are incorporated. Overall, all five algorithms yield an F1-score (the harmonic mean of precision and recall) 90% even when only 20% of the data is used for training. These findings highlight the strong potential of ML techniques for classifying radio sources in upcoming large radio continuum surveys.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Radio synchrotron emission originates from both massive star formation and black hole accretion, two processes that drive galaxy evolution. Therefore, current high-sensitivity and wide-field extragalactic radio continuum surveys require efficient and reliable classification of radio sources dominated by star formation or black hole accretion before utilizing radio continuum for exploring cosmic evolution. In this study, we implement, optimize, and compare five widely used supervised machine-learning (ML) algorithms to classify radio sources detected in the MeerKAT International GHz Tiered Extragalactic Exploration (MIGHTEE)–COSMOS survey as star-forming galaxies (SFGs) and active galactic nuclei (AGN). We utilize conventionally classified SFGs and AGN from MIGHTEE-COSMOS to construct training and test sets for evaluating the ML algorithm’s performance. To select input features for our ML analyses, we incorporate 18 physical parameters of MIGHTEE-detected radio sources. As anticipated, our feature analyses rank the six parameters used in conventional classification as the most effective: the infrared-radio correlation parameter (qIR), the optical compactness morphology parameter (class_star), three combined mid-infrared colors, and stellar mass. By optimizing the ML models with these selected features and testing classifiers across various feature combinations, we find that model performance generally improves as additional features are incorporated. Overall, all five algorithms yield an F1-score (the harmonic mean of precision and recall) 90% even when only 20% of the data is used for training. These findings highlight the strong potential of ML techniques for classifying radio sources in upcoming large radio continuum surveys.
Ssempeebwa, Phillip
Hierarchical CXR-Net: A Two-Stage Framework for Efficient and Interpretable Chest X-Ray Diagnosis Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{SsempeebwaPhillip2025,
title = {Hierarchical CXR-Net: A Two-Stage Framework for Efficient and Interpretable Chest X-Ray Diagnosis},
author = {Phillip Ssempeebwa},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Phillip%20_Ssempeebwa%20.pdf?generation=1755026612009720&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Chest radiography is one of the most common and vital diagnostic imaging tools globally. However, the interpretation of chest X-rays can be challenging, time-consuming, and subject to inter-observer variability, particularly in resource-limited settings where there is a shortage of expert radiologists. To address this, we present an end-to-end deep learning model designed to provide comprehensive radiological assistance by not only identifying abnormalities but also localizing them and generating structured reports.
Our methodology leverages the large-scale NIH Chest X-ray dataset, comprising over 112,000 images. We trained a DenseNet-121 model for multi-label classification across 14 common thoracic pathologies. Crucially, to overcome the scarcity of location-specific annotations, we employ a weakly-supervised learning approach. Gradient-weighted Class Activation Mapping (Grad-CAM) is used on the trained classification model to generate visual heatmaps that highlight the regions indicative of predicted diseases, providing effective abnormality localization without requiring explicit bounding box labels during training.
The classification model achieved a strong mean Area Under the ROC Curve (AUROC) of 0.796 across all pathologies on a held-out test set, demonstrating robust diagnostic performance. Qualitative results show that the Grad-CAM heatmaps successfully and plausibly highlight relevant pathological regions. The system's final output is a structured, human-readable report that synthesizes the classification and localization findings, mimicking a preliminary radiological summary.
This work demonstrates a viable and effective pipeline for AI-powered radiological assistance. By combining classification, localization, and automated reporting, our system has the potential to enhance diagnostic accuracy, improve workflow efficiency, and serve as a valuable educational and decision-support tool for healthcare professionals.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Chest radiography is one of the most common and vital diagnostic imaging tools globally. However, the interpretation of chest X-rays can be challenging, time-consuming, and subject to inter-observer variability, particularly in resource-limited settings where there is a shortage of expert radiologists. To address this, we present an end-to-end deep learning model designed to provide comprehensive radiological assistance by not only identifying abnormalities but also localizing them and generating structured reports.
Our methodology leverages the large-scale NIH Chest X-ray dataset, comprising over 112,000 images. We trained a DenseNet-121 model for multi-label classification across 14 common thoracic pathologies. Crucially, to overcome the scarcity of location-specific annotations, we employ a weakly-supervised learning approach. Gradient-weighted Class Activation Mapping (Grad-CAM) is used on the trained classification model to generate visual heatmaps that highlight the regions indicative of predicted diseases, providing effective abnormality localization without requiring explicit bounding box labels during training.
The classification model achieved a strong mean Area Under the ROC Curve (AUROC) of 0.796 across all pathologies on a held-out test set, demonstrating robust diagnostic performance. Qualitative results show that the Grad-CAM heatmaps successfully and plausibly highlight relevant pathological regions. The system's final output is a structured, human-readable report that synthesizes the classification and localization findings, mimicking a preliminary radiological summary.
This work demonstrates a viable and effective pipeline for AI-powered radiological assistance. By combining classification, localization, and automated reporting, our system has the potential to enhance diagnostic accuracy, improve workflow efficiency, and serve as a valuable educational and decision-support tool for healthcare professionals.
Our methodology leverages the large-scale NIH Chest X-ray dataset, comprising over 112,000 images. We trained a DenseNet-121 model for multi-label classification across 14 common thoracic pathologies. Crucially, to overcome the scarcity of location-specific annotations, we employ a weakly-supervised learning approach. Gradient-weighted Class Activation Mapping (Grad-CAM) is used on the trained classification model to generate visual heatmaps that highlight the regions indicative of predicted diseases, providing effective abnormality localization without requiring explicit bounding box labels during training.
The classification model achieved a strong mean Area Under the ROC Curve (AUROC) of 0.796 across all pathologies on a held-out test set, demonstrating robust diagnostic performance. Qualitative results show that the Grad-CAM heatmaps successfully and plausibly highlight relevant pathological regions. The system's final output is a structured, human-readable report that synthesizes the classification and localization findings, mimicking a preliminary radiological summary.
This work demonstrates a viable and effective pipeline for AI-powered radiological assistance. By combining classification, localization, and automated reporting, our system has the potential to enhance diagnostic accuracy, improve workflow efficiency, and serve as a valuable educational and decision-support tool for healthcare professionals.
Pidy, Pidy
Enhancing Urban Mobility in Developing Countries: A VANET Architecture with Secure mRSU Routing and Machine Learning Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{PidyPidyleoncetherese2025,
title = {Enhancing Urban Mobility in Developing Countries: A VANET Architecture with Secure mRSU Routing and Machine Learning},
author = {Pidy Pidy},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/leonce%20therese_Pidy%20Pidy.pdf?generation=1755026586816721&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Developing countries struggle to ensure efficient and safe mobility due to urban congestion, inadequate infrastructure, and limited technical resources. This leads to lost productivity, increased pollution, limited access to essential services, and heightened vulnerability to road accidents and safety hazards. This research proposes a comprehensive, context-aware solution leveraging Vehicular Ad-Hoc Networks (VANETs), integrating: Mobile mRSUs deployed on motorcycle-taxis to extend network coverage, An adaptive ACO-based routing protocol with dynamic clustering to optimize traffic efficiency, A machine learning-powered intrusion detection system to secure communications against Blackhole, Grayhole, and Sybil attacks.
Experimental results demonstrate significant improvements in throughput, latency reduction, and resilience against cyber threats. This work represents a major step toward intelligent, context-adapted urban transport systems, specifically designed for the challenges of African cities.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Developing countries struggle to ensure efficient and safe mobility due to urban congestion, inadequate infrastructure, and limited technical resources. This leads to lost productivity, increased pollution, limited access to essential services, and heightened vulnerability to road accidents and safety hazards. This research proposes a comprehensive, context-aware solution leveraging Vehicular Ad-Hoc Networks (VANETs), integrating: Mobile mRSUs deployed on motorcycle-taxis to extend network coverage, An adaptive ACO-based routing protocol with dynamic clustering to optimize traffic efficiency, A machine learning-powered intrusion detection system to secure communications against Blackhole, Grayhole, and Sybil attacks.
Experimental results demonstrate significant improvements in throughput, latency reduction, and resilience against cyber threats. This work represents a major step toward intelligent, context-adapted urban transport systems, specifically designed for the challenges of African cities.
Experimental results demonstrate significant improvements in throughput, latency reduction, and resilience against cyber threats. This work represents a major step toward intelligent, context-adapted urban transport systems, specifically designed for the challenges of African cities.
John, Ngeta
Music Information Retrieval: Teaching Machines to Listen Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{JohnNgeta2025,
title = {Music Information Retrieval: Teaching Machines to Listen},
author = {Ngeta John},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Ngeta_John.pdf?generation=1755026598516963&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Abstract:
Music Information Retrieval (MIR) represents a fundamental challenge in teaching machines to understand music the way humans do. This poster explores the evolution of deep learning approaches to automatic chord recognition, from early CNN architectures achieving 77% accuracy to modern foundation models like MERT reaching 86.9% performance on standard benchmarks. We examine the complete technical pipeline from raw audio to musical understanding: audio preprocessing, chroma feature extraction, and the architectural evolution from CNNs (2012) through LSTM networks (2018) to Transformer-based models (2019) and foundation models (2023). Current systems achieve real-time processing with <100ms latency, enabling applications in music therapy, personalized education, and adaptive entertainment. However, a critical challenge remains: existing MIR systems exhibit significant Western bias, achieving 88% accuracy on Western pop music but <60% on traditional African music. We discuss how foundation models like MERT, with their self-supervised learning capabilities and scaling from 95M to 330M parameters, offer potential pathways to culturally-inclusive music AI. The poster demonstrates how techniques familiar to deep learning practitioners—CNNs for pattern recognition, LSTMs for sequence modeling, and Transformers for attention-based understanding—apply directly to music, creating systems that not only recognize chords but generate new musical content. We conclude with a vision for building MIR systems that celebrate musical diversity and serve all cultures, highlighting opportunities for African researchers to contribute to more inclusive AI development.
Keywords: Music Information Retrieval, Deep Learning, Chord Recognition, Foundation Models, Cultural AI, MERT, Transformers},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Abstract:
Music Information Retrieval (MIR) represents a fundamental challenge in teaching machines to understand music the way humans do. This poster explores the evolution of deep learning approaches to automatic chord recognition, from early CNN architectures achieving 77% accuracy to modern foundation models like MERT reaching 86.9% performance on standard benchmarks. We examine the complete technical pipeline from raw audio to musical understanding: audio preprocessing, chroma feature extraction, and the architectural evolution from CNNs (2012) through LSTM networks (2018) to Transformer-based models (2019) and foundation models (2023). Current systems achieve real-time processing with <100ms latency, enabling applications in music therapy, personalized education, and adaptive entertainment. However, a critical challenge remains: existing MIR systems exhibit significant Western bias, achieving 88% accuracy on Western pop music but <60% on traditional African music. We discuss how foundation models like MERT, with their self-supervised learning capabilities and scaling from 95M to 330M parameters, offer potential pathways to culturally-inclusive music AI. The poster demonstrates how techniques familiar to deep learning practitioners—CNNs for pattern recognition, LSTMs for sequence modeling, and Transformers for attention-based understanding—apply directly to music, creating systems that not only recognize chords but generate new musical content. We conclude with a vision for building MIR systems that celebrate musical diversity and serve all cultures, highlighting opportunities for African researchers to contribute to more inclusive AI development.
Keywords: Music Information Retrieval, Deep Learning, Chord Recognition, Foundation Models, Cultural AI, MERT, Transformers
Music Information Retrieval (MIR) represents a fundamental challenge in teaching machines to understand music the way humans do. This poster explores the evolution of deep learning approaches to automatic chord recognition, from early CNN architectures achieving 77% accuracy to modern foundation models like MERT reaching 86.9% performance on standard benchmarks. We examine the complete technical pipeline from raw audio to musical understanding: audio preprocessing, chroma feature extraction, and the architectural evolution from CNNs (2012) through LSTM networks (2018) to Transformer-based models (2019) and foundation models (2023). Current systems achieve real-time processing with <100ms latency, enabling applications in music therapy, personalized education, and adaptive entertainment. However, a critical challenge remains: existing MIR systems exhibit significant Western bias, achieving 88% accuracy on Western pop music but <60% on traditional African music. We discuss how foundation models like MERT, with their self-supervised learning capabilities and scaling from 95M to 330M parameters, offer potential pathways to culturally-inclusive music AI. The poster demonstrates how techniques familiar to deep learning practitioners—CNNs for pattern recognition, LSTMs for sequence modeling, and Transformers for attention-based understanding—apply directly to music, creating systems that not only recognize chords but generate new musical content. We conclude with a vision for building MIR systems that celebrate musical diversity and serve all cultures, highlighting opportunities for African researchers to contribute to more inclusive AI development.
Keywords: Music Information Retrieval, Deep Learning, Chord Recognition, Foundation Models, Cultural AI, MERT, Transformers
Idakwo, Patricia Ojonoka
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{IdakwoPatriciaOjonoka2025,
title = {Road Traffic Crash Severity Prediction in Low-Resource Contexts: Ensemble Machine Learning and Deep Learning Approaches},
author = {Patricia Ojonoka Idakwo},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Patricia%20Ojonoka%20_Idakwo.pdf?generation=1755026551811226&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {Road traffic crash (RTC) severity prediction is critical for improving post-crash care and reducing fatalities, particularly in low-resource contexts like Nigeria, where structured crash data is scarce and emergency medical services (EMS) face operational constraints. This study presents a data-centric, multi-modal approach to RTC severity prediction, addressing these challenges. A Nigerian RTC dataset comprising 59 features across three data modes: unstructured textual data (mode 1), structured numerical data (mode 2), and a fusion of both (mode 3) was curated from unstructured online crash narratives using Natural Language Processing techniques - named entity recognition, one-hot encoding, and text mining. Owing to class imbalance, we utilized the weighted average F1-score to evaluate the performance of the ensemble machine learning and deep learning models in RTC severity prediction. Across all modes, the LSTM-CNN model with Word2Vec embeddings achieved the best performance on the mode 3 data with a weighted F1-score of 0.755, and on mode 1 data with 0.674, while Gradient Boosting achieved the highest score (0.520) on mode 2 data. These findings highlight the advantage of multi-modal data fusion and hybrid neural networks in enhancing RTC severity prediction for data-driven EMS resource allocation and road safety, supporting efforts to reduce mortality and serious injuries in resource-constrained emergency response systems.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Road traffic crash (RTC) severity prediction is critical for improving post-crash care and reducing fatalities, particularly in low-resource contexts like Nigeria, where structured crash data is scarce and emergency medical services (EMS) face operational constraints. This study presents a data-centric, multi-modal approach to RTC severity prediction, addressing these challenges. A Nigerian RTC dataset comprising 59 features across three data modes: unstructured textual data (mode 1), structured numerical data (mode 2), and a fusion of both (mode 3) was curated from unstructured online crash narratives using Natural Language Processing techniques - named entity recognition, one-hot encoding, and text mining. Owing to class imbalance, we utilized the weighted average F1-score to evaluate the performance of the ensemble machine learning and deep learning models in RTC severity prediction. Across all modes, the LSTM-CNN model with Word2Vec embeddings achieved the best performance on the mode 3 data with a weighted F1-score of 0.755, and on mode 1 data with 0.674, while Gradient Boosting achieved the highest score (0.520) on mode 2 data. These findings highlight the advantage of multi-modal data fusion and hybrid neural networks in enhancing RTC severity prediction for data-driven EMS resource allocation and road safety, supporting efforts to reduce mortality and serious injuries in resource-constrained emergency response systems.
Oketta, Peter
Enhancing Bean Crop Disease Diagnosis with Vision-Language Models: A Multitask Approach using PaliGemma Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{OkettaPeter2025,
title = {Enhancing Bean Crop Disease Diagnosis with Vision-Language Models: A Multitask Approach using PaliGemma},
author = {Peter Oketta},
url = {https://drive.google.com/file/d/1U5QvziNXzJTqB2zoPoIcTI759S5jYxxt/view?usp=sharing},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Baker, Rameeze
[No title] Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{BakerRameeze2025,
title = {[No title]},
author = {Rameeze Baker},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Rameeze_Baker.pdf?generation=1755026593751224&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {For too long, compliance has been a barrier instead of a bridge. Small businesses, traders, and entrepreneurs face complex regulations, limited financial access, and lack of trust in the system. This keeps them locked out of opportunities that could help them grow. The reality? Millions of hardworking business owners are stuck navigating rules that weren’t designed for them, without the tools to succeed. S2P redefines Compliance as a force for growth, not limitation.
Our Goals:
1. Turning Trust into Currency
2. Unleashing The Power of Crowdsourced Accountability
3. Bridging the Gap Between Informal & Formal Economies
4. Future-Proofing Against Compliance & Sustainability Shifts
S2P; an economic enabler, a trust engine, and a bridge between fragmented markets.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
For too long, compliance has been a barrier instead of a bridge. Small businesses, traders, and entrepreneurs face complex regulations, limited financial access, and lack of trust in the system. This keeps them locked out of opportunities that could help them grow. The reality? Millions of hardworking business owners are stuck navigating rules that weren’t designed for them, without the tools to succeed. S2P redefines Compliance as a force for growth, not limitation.
Our Goals:
1. Turning Trust into Currency
2. Unleashing The Power of Crowdsourced Accountability
3. Bridging the Gap Between Informal & Formal Economies
4. Future-Proofing Against Compliance & Sustainability Shifts
S2P; an economic enabler, a trust engine, and a bridge between fragmented markets.
Our Goals:
1. Turning Trust into Currency
2. Unleashing The Power of Crowdsourced Accountability
3. Bridging the Gap Between Informal & Formal Economies
4. Future-Proofing Against Compliance & Sustainability Shifts
S2P; an economic enabler, a trust engine, and a bridge between fragmented markets.
LUPYANI, REBECCA
Leveraging AI for Visual impairment: A Linguistically Inclusive Model for Zambian Learners Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{LUPYANIREBECCA2025,
title = {Leveraging AI for Visual impairment: A Linguistically Inclusive Model for Zambian Learners},
author = {REBECCA LUPYANI},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/REBECCA%20_LUPYANI%20.pdf?generation=1755026574730734&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
abstract = {In today’s digital era, emerging technologies are being harnessed to enhance and support various sectors, including inclusive education. Recognizing the imperative of educational equity, a number of international and national policies have been instituted to promote inclusive learning environments. One such policy is Article 24 of the United Nations Convention on the Rights of Persons with Disabilities (UNCRPD), adopted in 2006. It affirms the right of persons with disabilities to inclusive education at all levels and mandates that states ensure reasonable accommodations, individualized support, and accessibility within mainstream education systems. This aligns closely with the Envision 2030 Agenda, which articulates the 17 Sustainable Development Goals (SDGs) aimed at creating a more inclusive and equitable world for all, particularly persons with disabilities. Anchored in the principle of “leaving no one behind,” the agenda underscores inclusive education as a critical driver of sustainable development.
Despite global advocacy, the realization of SDG 4, which seeks to ensure inclusive and equitable quality education for all remains elusive for learners who are blind or visually impaired across Africa. These learners continue to face entrenched barriers, including inaccessible learning materials, limited availability and affordability of assistive technologies, and a shortage of trained educators, all of which hinder their full participation in educational systems . However, recent advances in Artificial Intelligence (AI) and Machine Learning (ML) have introduced promising new possibilities for assistive technologies. Innovations such as AI-powered screen readers, object recognition tools, and voice-enabled navigation systems are increasingly being developed to support independent learning and mobility for individuals with visual impairments .
Within African contexts, emerging studies reveal both potential and persisting challenges. In Nigeria, for instance, only about 36% of visually impaired adults are aware of existing AI- powered assistive technologies, and fewer than 18% possess the skills to use them effectively . In Kenya, AI-powered tools such as smart canes, screen readers, and applications like ‘Seeing AI’ and ‘Be My Eyes’ have contributed to enhanced mobility and digital inclusion. Nonetheless, inequities in internet access, affordability, and public awareness continue to impede widespread adoption.
In Zambia, a study by Ndume (2025) that investigated the use of assistive devices for blind and visually impaired learners, revealed that 65.6% of visually impaired students relied on no-tech tools, namely, braille slates, styluses, abacuses, 25% used low-tech devices, while only 9.4% accessed high-tech assistive technologies such as computers, tablets, smartphones, and screen reader software like JAWS. Although AI-powered assistive technologies have gained traction globally, their potential remains largely untapped in Zambia. A critical gap exists not only in the adoption of AI-based tools but also in the capacity of these tools to function in Zambian local languages, which is crucial for inclusive learning.
Therefore, this study seeks to design and develop an AI- Powered wearable assistive device for the blind and visually impaired, that is designed to provide learners with an awareness of their surrounding by providing facial recognition, object identification, screen description and providing auditory feedback using Bemba which is one of Zambia’s widely spoken local languages.},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
In today’s digital era, emerging technologies are being harnessed to enhance and support various sectors, including inclusive education. Recognizing the imperative of educational equity, a number of international and national policies have been instituted to promote inclusive learning environments. One such policy is Article 24 of the United Nations Convention on the Rights of Persons with Disabilities (UNCRPD), adopted in 2006. It affirms the right of persons with disabilities to inclusive education at all levels and mandates that states ensure reasonable accommodations, individualized support, and accessibility within mainstream education systems. This aligns closely with the Envision 2030 Agenda, which articulates the 17 Sustainable Development Goals (SDGs) aimed at creating a more inclusive and equitable world for all, particularly persons with disabilities. Anchored in the principle of “leaving no one behind,” the agenda underscores inclusive education as a critical driver of sustainable development.
Despite global advocacy, the realization of SDG 4, which seeks to ensure inclusive and equitable quality education for all remains elusive for learners who are blind or visually impaired across Africa. These learners continue to face entrenched barriers, including inaccessible learning materials, limited availability and affordability of assistive technologies, and a shortage of trained educators, all of which hinder their full participation in educational systems . However, recent advances in Artificial Intelligence (AI) and Machine Learning (ML) have introduced promising new possibilities for assistive technologies. Innovations such as AI-powered screen readers, object recognition tools, and voice-enabled navigation systems are increasingly being developed to support independent learning and mobility for individuals with visual impairments .
Within African contexts, emerging studies reveal both potential and persisting challenges. In Nigeria, for instance, only about 36% of visually impaired adults are aware of existing AI- powered assistive technologies, and fewer than 18% possess the skills to use them effectively . In Kenya, AI-powered tools such as smart canes, screen readers, and applications like ‘Seeing AI’ and ‘Be My Eyes’ have contributed to enhanced mobility and digital inclusion. Nonetheless, inequities in internet access, affordability, and public awareness continue to impede widespread adoption.
In Zambia, a study by Ndume (2025) that investigated the use of assistive devices for blind and visually impaired learners, revealed that 65.6% of visually impaired students relied on no-tech tools, namely, braille slates, styluses, abacuses, 25% used low-tech devices, while only 9.4% accessed high-tech assistive technologies such as computers, tablets, smartphones, and screen reader software like JAWS. Although AI-powered assistive technologies have gained traction globally, their potential remains largely untapped in Zambia. A critical gap exists not only in the adoption of AI-based tools but also in the capacity of these tools to function in Zambian local languages, which is crucial for inclusive learning.
Therefore, this study seeks to design and develop an AI- Powered wearable assistive device for the blind and visually impaired, that is designed to provide learners with an awareness of their surrounding by providing facial recognition, object identification, screen description and providing auditory feedback using Bemba which is one of Zambia’s widely spoken local languages.
Despite global advocacy, the realization of SDG 4, which seeks to ensure inclusive and equitable quality education for all remains elusive for learners who are blind or visually impaired across Africa. These learners continue to face entrenched barriers, including inaccessible learning materials, limited availability and affordability of assistive technologies, and a shortage of trained educators, all of which hinder their full participation in educational systems . However, recent advances in Artificial Intelligence (AI) and Machine Learning (ML) have introduced promising new possibilities for assistive technologies. Innovations such as AI-powered screen readers, object recognition tools, and voice-enabled navigation systems are increasingly being developed to support independent learning and mobility for individuals with visual impairments .
Within African contexts, emerging studies reveal both potential and persisting challenges. In Nigeria, for instance, only about 36% of visually impaired adults are aware of existing AI- powered assistive technologies, and fewer than 18% possess the skills to use them effectively . In Kenya, AI-powered tools such as smart canes, screen readers, and applications like ‘Seeing AI’ and ‘Be My Eyes’ have contributed to enhanced mobility and digital inclusion. Nonetheless, inequities in internet access, affordability, and public awareness continue to impede widespread adoption.
In Zambia, a study by Ndume (2025) that investigated the use of assistive devices for blind and visually impaired learners, revealed that 65.6% of visually impaired students relied on no-tech tools, namely, braille slates, styluses, abacuses, 25% used low-tech devices, while only 9.4% accessed high-tech assistive technologies such as computers, tablets, smartphones, and screen reader software like JAWS. Although AI-powered assistive technologies have gained traction globally, their potential remains largely untapped in Zambia. A critical gap exists not only in the adoption of AI-based tools but also in the capacity of these tools to function in Zambian local languages, which is crucial for inclusive learning.
Therefore, this study seeks to design and develop an AI- Powered wearable assistive device for the blind and visually impaired, that is designed to provide learners with an awareness of their surrounding by providing facial recognition, object identification, screen description and providing auditory feedback using Bemba which is one of Zambia’s widely spoken local languages.
N’guessan, Regis
The Reflexive Integrated Information Unit: A Differentiable Primitive for Artificial Consciousness Presentation
Poster presented at the Deep Learning Indaba 2025, Kigali, Rwanda, 01.08.2025, (Non-archival).
@misc{N’guessanRegis2025,
title = {The Reflexive Integrated Information Unit: A Differentiable Primitive for Artificial Consciousness},
author = {Regis N’guessan},
url = {https://storage.googleapis.com/download/storage/v1/b/indaba-2025-posters/o/Regis%20_N’guessan%20.pdf?generation=1755026590101730&alt=media},
year = {2025},
date = {2025-08-01},
address = {Kigali, Rwanda},
howpublished = {Poster presented at the Deep Learning Indaba 2025},
note = {Non-archival},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}