Work produced through AIxBio Africa's fellowships, courses, and wider research activities.
AIxBio Africa Outputs
Work emerging from our programmes
This library brings together fellowship research, course Capstone outputs, tools, reports, and other work produced through AIxBio Africa. Output types are labelled clearly so that project work is not presented as peer-reviewed research unless it is actually published as such.
Fellowship
Fellowship Publications
Research and project outputs produced through the AIxBio Africa Research Fellowship.
PDFFeaturedPilot Cohort 2026
Evaluating Confidence and Policy Collapse in Multimodal Agents: Towards Robust Auditing Standards for African Healthcare
Tegan Jegede
Research Question
How reliably does a multimodal agent's confidence correspond to its correctness across healthcare domains, and how does that relationship change with degraded medical images or regionally relevant clinical questions?
"How well does an AI system know when it is wrong? This project investigated whether a multimodal AI model could reliably estimate the correctness of its healthcare answers, respond appropriately to degraded medical images, and abstain when the available evidence was insufficient.
I evaluated LLaVA-1.5-7B using AfriMed-QA, a Pan-African clinical question-answering benchmark, alongside SLAKE and PathVQA medical-image datasets. The final study comprised 400 evaluations: 100 AfriMed-QA clinical questions and 100 radiology and pathology questions, each tested using clean, blurred, and darkened images. I measured accuracy, confidence calibration, high-confidence errors, abstention, and prediction stability.
The model achieved 33% accuracy on AfriMed-QA while reporting 61.7% average confidence, revealing substantial overconfidence. Thirteen percent of its AfriMed-QA responses were incorrect despite confidence of at least 80%. Across the visual benchmarks, combined accuracy declined from 64% on clean images to 60% under blur and 59% under darkness. The model never selected the available “insufficient evidence” option in any of the 300 visual evaluations.
This work matters in an African context because medical AI is often evaluated using datasets that insufficiently represent African institutions, clinical knowledge, and healthcare conditions. The findings demonstrate the need for locally relevant evaluation, calibrated uncertainty, and effective human referral before multimodal AI systems are considered for high-stakes healthcare use."
Uncertainty QuantificationMultimodal AI SafetyHealthcare AI AuditingModel Calibration
PDFFeaturedPilot Cohort 2026
Minimum Risk Management Checks for AI Decision Support Tools in Primary and Community Health Care
Gideon Abako
Research Question
What minimum risk management checks should health institutions apply across the deployment lifecycle of AI decision support tools used in primary and community health care?
"This project investigated what minimum risk checks African health institutions should apply before clinical AI decision support tools are used in primary care. The work responds to a practical gap. Global guidance from WHO, the African Union, NIST and medical device regulators provide principles but frontline institutions still need a way to review AI tools before they influence care.
The project developed a life cycle risk management framework for clinical AI decision support in African primary care. The framework was built through a review of AI for health governance, digital health implementation guidance, health AI evaluation literature and empirical evidence from primary care and frontline settings. It has five stages of procurement and selection, pre-deployment assessment, deployment and adoption, monitoring and risk review and post deployment learning and decommissioning.
The main finding is that risk is produced by the interaction between the AI system, health worker, patient population, clinical workflow, local guidelines, data quality, infrastructure, vendor arrangements and institutional accountability. The output is a practitioner framework and checklist that help district health teams, facility managers, programme teams and procurement actors ask better questions before adopting or scaling AI decision support in primary care.
"
AI GovernanceHealth AIAI Risk ManagementPrimary Health Care
PDFFeaturedPilot Cohort 2026
When Answering Is Not Enough
Frances Chinaza Agba
Research Question
Does AI behave differently when culturally grounded Nigerian public-health scenarios are presented in Nigerian languages (Pidgin, Yoruba, Igbo) vs. standard English?
When Answering Is Not Enough examines whether three frontier-model configurations can safely answer legitimate public-health questions across English, Nigerian Pidgin, Yorùbá and Igbo. Across 1,080 responses to direct and culturally contextualised prompts, no response was classified as an unnecessary refusal, yet 79 failed after engagement through misunderstanding, wrong-language delivery, incomplete guidance or materially unsafe advice. By identifying failures that refusal-only testing would miss, the project provides an auditable, provenance-sensitive method for evaluating multilingual safety across the full response pathway.
AI SafetyMultilingual AIPublic HealthCultural Calibration
GitHub projectFeaturedPilot Cohort 2026
Biosignals and Behavioral Prediction
Aayush Gandhi
Research Question
Behaviour and decisions leave traces in the body: in the heart, the skin, the pupil, the hormones, and the brain. Each trace can be picked up with cheap, already-deployed hardware (a wrist strap, a phone camera, a saliva strip, a low-cost EEG cap). On its own, each trace is a weak predictor. The safety question this project studies is what happens when an AI system is given several of them at once
"I investigated whether a documented neuroforecasting phenomenon, that neural activity, particularly in the nucleus accumbens, predicts population-level outcomes more accurately than individual self-report, holds up under rigorous, leakage-safe conditions across multiple data modalities. Prior work in this area (Genevsky and colleagues) demonstrated the effect using single modalities in domains like crowdfunding and public health campaigns, but forecasting pipelines in this space are vulnerable to a specific failure mode: models that appear to predict behavior but are actually just learning to recognize individuals across trials. I built a framework combining fMRI, EEG, facial video, and behavioral self-report that enforces subject-grouped cross-validation, nested hyperparameter tuning, and permutation testing to rule this out, and uses synthetic data as a positive control to confirm the pipeline recovers known effects by construction before touching real data. The framework is validated on real fMRI (NARPS) and EEG (DEAP) datasets with a 66-test suite specifically targeting leakage.
This matters in an African context because population-level forecasting for microlending uptake, public health campaign effectiveness, and agricultural technology adoption, exactly the domains where this neuroforecasting effect has been shown, currently depends on large, expensive self-report surveys. A validated, leakage-resistant pipeline is a prerequisite for eventually using cheaper signals like EEG for this kind of forecasting in resource-constrained research settings, where survey-based data collection at scale is often infeasible. However, this work also has the potential to be negatively used, enabling targeting messaging designed to influence behavioral actions through neuro-manipulation."