Scaling Responsible AI in African Health Systems

Levi Cheptora

Wed, 17 Dec 2025

Abstract

Artificial intelligence (AI) is increasingly embedded in health systems across Sub-Saharan Africa (SSA), from tuberculosis (TB) screening with computer-aided detection to large-scale digital triage. While the potential for impact is substantial, scaling responsibly requires attention to governance, safety, equity, data protection, capacity building, and sustainability. This secondary research paper synthesizes recent guidance from the World Health Organization (WHO), the African Union (AU), and national regulators, alongside case examples (e.g., AI-assisted TB radiography and digital triage in Rwanda) to derive pragmatic lessons for implementers and policymakers. We find that (a) continental and national governance frameworks are converging on rights-preserving, risk-based regulation; (b) AI systems can achieve clinically meaningful performance in priority areas such as TB screening, but require robust external validation and post-market surveillance; (c) sustainability and vendor risk must be managed proactively to avoid service shocks; (d) data protection regimes (e.g., Nigeria NDPA 2023, Kenya ODPC health data guidance, South Africa POPIA) provide enforceable obligations applicable to health AI; and (e) capacity networks (e.g., DHIS2/HISP, Masakhane, Deep Learning Indaba, DS-I Africa) are critical to localization and equity. We conclude with a practical checklist for “responsible-by-design” scale-up in SSA health systems. ScienceDirect+4World Health Organization+4World Health Organization+4

Keywords: Responsible AI, health informatics, Sub-Saharan Africa, governance, data protection, TB screening, health equity

Introduction

The acceleration of AI-enabled tools in global health has coincided with new policy guardrails. WHO’s 2021 ethics report and 2023 regulatory considerations articulate principles and pathways for safe, effective, and equitable AI in health, emphasizing transparency, external validation, human oversight, and lifecycle risk management. World Health Organization+1 In parallel, regional policy has advanced: the African Union’s 2024 Continental AI Strategy endorses an Africa-centric, rights-preserving approach to AI deployment, with clear relevance for health programs. African Union

In SSA, AI deployments are most visible where clinical gaps are largest and data are relatively structured—e.g., TB screening with AI-assisted chest X-ray (CXR) and algorithmic triage/chat-based navigation. Evidence and experience from these frontlines offer concrete lessons for scaling responsibly.

Methods (Secondary Research)

We conducted a narrative synthesis of (a) multilateral guidance (WHO/AU), (b) peer-reviewed studies on AI for TB screening, (c) national data-protection and sectoral guidance (Nigeria, Kenya, South Africa, Ghana), and (d) ecosystem/capacity sources (DHIS2/HISP, Masakhane, Deep Learning Indaba, DS-I Africa). Evidence sources (2021–2025) were prioritized for recency and policy salience; where appropriate, we include case-based reporting on digital triage deployments and their failure modes. Government of South Africa+6World Health Organization+6World Health Organization+6

Findings

1) Governance is maturing: clear principles and risk-based regulation

WHO’s 2023 Regulatory Considerations on AI for Health outlines expectations across the product lifecycle—intended use definition, data governance, documentation, clinical evaluation, post-market monitoring, and change management—while urging context-appropriate regulatory capacity building. World Health Organization The AU’s 2024 Continental AI Strategy complements these with continental priorities around ethics, inclusion, capability building, and cooperation—relevant for cross-border health data sharing and regional procurement. African Union

Implication: Ministries of Health (MoHs) and regulators can adopt a risk-proportionate approach (e.g., higher scrutiny for diagnostic/therapeutic AI; lighter touch for workflow optimization) and align procurement with WHO’s documentation and evaluation expectations. World Health Organization

2) Clinical performance can be strong—but depends on external validation and continuous monitoring

CAD (computer-aided detection) for TB on CXR has WHO endorsement as an alternative to human readers for screening/triage in adults since 2021, with an implementation guide updated in 2024–2025. Systematic reviews and recent external validations in South Africa report CAD systems approaching radiologist-level performance for TB triage, although thresholds require calibration to prevalence and program goals. ScienceDirect+3World Health Organization+3The Lancet+3

Implication: Programs should (a) perform local external validation, (b) tune decision thresholds for their epidemiology and workflow, and (c) establish post-market surveillance for dataset shift and version changes—consistent with WHO guidance. World Health Organization

3) Sustainability and vendor risk are central to “responsible scale”

Rwanda offers a cautionary tale. Babyl (the local arm of Babylon Health) pioneered at-scale AI-enabled digital triage and teleconsultation under a government partnership, reportedly handling thousands of daily consultations. In 2023, following parent-company bankruptcy, Babyl wound down Rwandan operations, creating a service discontinuity and prompting sector reflection on contingency planning and public digital infrastructure. DCCC+1

Implication: Health systems should require (a) escrow or open interfaces for critical components, (b) data portability and exit plans, (c) SLAs tied to continuity, and (d) preference for solutions that integrate with national data backbones (e.g., DHIS2-based platforms; Rwanda’s new Health Intelligence Center). DHIS2+1

4) Enforceable data protection frameworks apply to health AI

SSA features a growing patchwork of enforceable data-protection law with direct implications for health AI:

• Nigeria: The Nigeria Data Protection Act (NDPA) 2023 established the Data Protection Commission and a legal framework for personal data processing, including sensitive health data—enforceable with penalties. KPMG Assets+1
• Kenya: The Office of the Data Protection Commissioner (ODPC) issued a 2023/2024 Guidance Note on the Processing of Health Data, detailing safeguards and organizational/technical measures for health data processing. odpc.go.ke
• South Africa: POPIA (effective 2020/2021) governs health information with conditions of lawful processing, security safeguards, and breach notification. Government of South Africa+1
• Ghana: The Data Protection Act (Act 843, 2012) and an active Data Protection Commission set obligations for controllers, relevant to health programs and cross-border transfers. NITA+1

Implication: AI implementers must conduct data protection impact assessments (DPIAs/AIPAs), define lawful bases (e.g., public interest in public health), minimize data, and design for portability/subject rights.

5) Equity requires investment in local language, dataset diversity, and cultural context

Bias remains a salient risk in computer vision and clinical NLP. Dermatology AI studies continue to show under-representation and reduced performance on darker skin tones, while new work highlights hue-related biases beyond light–dark scales—risks directly relevant to African populations. ScienceDirect+1 At the same time, grassroots capacity efforts are closing gaps: Masakhane advances African-language NLP resources; Deep Learning Indaba and DS-I Africa are scaling researcher capacity; Data Science Africa convenes applied training and collaborations. Data Science Africa+3Masakhane+3Deep Learning Indaba+3

Implication: “Responsibility” must include funded data diversification (e.g., African skin-of-color datasets), support for local language NLP, and participatory development with clinicians and communities.

Case Snapshots

A. AI-assisted CXR for TB screening

• Policy status: WHO recommends CAD as an alternative to human readers for screening and triage in people ≥15 years (since 2021), with updated implementation guidance (2024–2025). World Health Organization
• Evidence: Multicountry analyses and South African validations report high AUC and operational utility when thresholds are locally calibrated; programs must plan for QA, periodic re-validation, and version transparency. The Lancet+1
• Lesson: Pair adoption with governance: pre-deployment validation, documented intended use, calibration plans, and post-market monitoring consistent with WHO’s regulatory lifecycle. World Health Organization

B. Digital triage at national scale in Rwanda

• Scale & service: Prior to winding down in 2023, Babyl reportedly served >2.5 million users and thousands of consultations per day. DCCC
• Shock & response: Parent-company bankruptcy precipitated service cessation, underscoring the need for vendor-neutral architectures and contingency plans. Rwanda’s 2025 Health Intelligence Center signals a pivot toward sovereign data/analytics capacity. Africa Press Arabic+1
• Lesson: Sustainability is a safety issue—design exit ramps and ensure critical services are resilient to vendor failure.

Discussion

SSA experiences illustrate that “responsible” at scale is multi-dimensional: technical performance is necessary but insufficient without regulatory fit, data protection compliance, equity, and sustainability. Continental and national policy advances reduce ambiguity—the AU strategy clarifies principles and coordination opportunities; WHO provides detailed expectations for regulators and implementers. African Union+1

Operationally, TB CAD shows how AI can close diagnostic gaps when embedded in programmatic pathways and governed through lifecycle monitoring. Conversely, the Rwanda triage case highlights that vendor solvency and data portability are not peripheral—they are core to continuity and trust. The Lancet+2ScienceDirect+2

Equity requires active correction of data and model blind spots (skin tone, language, culture). Community-driven capacity networks (Masakhane, Indaba, DS-I Africa) are not merely “nice to have”; they are structural enablers for localization, oversight, and sustainability. Masakhane+2Deep Learning Indaba+2

Recommendations: A Responsible-by-Design Checklist for SSA Health AI

1. Anchor to policy & regulation
• Map the intended use to WHO regulatory expectations and national law; align documentation with WHO’s lifecycle (intended use, data governance, evaluation, post-market). World Health Organization
• Where relevant, consult AU strategy to harmonize cross-border data and procurement approaches. African Union
2. Clinical validation & monitoring
• Perform local external validation and calibration; pre-specify performance targets and safety nets; monitor for dataset shift and version drift. (TB CAD provides a template). World Health Organization+1
3. Data protection by design
• Run DPIAs/AIPAs; define lawful bases; minimize, pseudonymize, and secure data; plan for data subject rights and cross-border transfer compliance (e.g., NDPA, POPIA, Kenya ODPC guidance, Ghana Act 843). NITA+3KPMG Assets+3Government of South Africa+3
4. Equity & inclusion
• Budget for dataset diversification (e.g., skin-of-color images, African-language corpora); engage communities; measure and report subgroup performance. ScienceDirect
5. Sustainability & vendor neutrality
• Require source transparency (model/version docs), open standards/APIs, data portability, escrow/contingency plans, and SLAs tied to continuity; prefer integration with national backbones (DHIS2/HISP; national health intelligence platforms). DHIS2+1
6. Capacity & local ownership
• Leverage regional networks (Masakhane for NLP, Deep Learning Indaba, DS-I Africa, DSA) for training, evaluation partnerships, and co-development. Data Science Africa+3Masakhane+3Deep Learning Indaba+3

Limitations

This synthesis relies on publicly available guidance and studies; national implementations vary, and unpublished program evaluations may alter specific conclusions. TB CAD evidence is stronger than for some other clinical AI domains (e.g., dermatology in primary care), where equity risks remain active areas of research. The Lancet+1

Conclusion

Scaling responsible AI for health in SSA is feasible when anchored in risk-based regulation, rigorous local validation, enforceable data protection, equity investments, and sustainability planning. Regional capacity ecosystems and national digital backbones provide the scaffolding for durable, locally owned AI systems that measurably improve care while safeguarding rights. The lessons from TB screening and Rwanda’s digital triage illustrate both the upside and the operational guardrails needed for responsible scale. World Health Organization+2World Health Organization+2

References

African Union. (2024). Continental Artificial Intelligence Strategy. Addis Ababa: AU. African Union

HISP Centre / DHIS2. (n.d.). HISP network overview. University of Oslo. DHIS2

Kenya Office of the Data Protection Commissioner (ODPC). (2024). Guidance Note on the Processing of Health Data. Nairobi: ODPC. odpc.go.ke

Nigeria Federal Government. (2023). Nigeria Data Protection Act (NDPA). Abuja. (See also KPMG explanatory memorandum). KPMG Assets

Rwanda Ministry of Health. (2025). New Health Intelligence Center to drive real-time, evidence-based decisions. Kigali. moh.gov.rw

South Africa Government. (2020–2021). Protection of Personal Information Act (POPIA) commencement and compliance. Pretoria. Government of South Africa+1

The Lancet Digital Health. (2024). Computer-aided detection of tuberculosis from chest radiographs. The Lancet

WHO. (2021). Ethics and governance of artificial intelligence for health. Geneva: World Health Organization. World Health Organization

WHO. (2023). Regulatory considerations on artificial intelligence for health. Geneva: World Health Organization. World Health Organization

WHO. (2024/2025). Use of computer-aided detection software for tuberculosis screening: Implementation guide. Geneva: WHO Global TB Programme. World Health Organization

Digital Connected Care Coalition. (2024). Scaling digital healthcare solutions in Rwanda: Lessons learned from Babyl. DCCC

Badrie, S. (2025). Skin tone bias in AI dermatology tools. RCSI Student Medical Journal. RCSIsmj