Artificial intelligence tools are being deployed in clinical settings in Kenya without a validated governance framework. A diagnostic algorithm trained on population data from high-income countries and deployed in a Kenyan facility is not a neutral tool. It is a tool whose embedded assumptions about normal physiological ranges, disease prevalence, and demographic distribution reflect the population on which it was trained — a population that is likely to be significantly different from the one it is now classifying (Obermeyer et al. 2019).
The Gap in the Literature
The governance literature on AI in healthcare has developed rapidly in OECD contexts, producing the EU AI Act, WHO guidance on AI in health (World Health Organization 2021), and a growing body of scholarship on algorithmic bias in clinical decision support. What is substantially absent from this literature is a governance framework calibrated to the specific institutional, infrastructural, and population conditions of East African health systems — including the absence of locally validated training datasets, the reality of community-level rather than hospital-based primary care, and the particular vulnerability of persons with disabilities to diagnostic tools that have rarely been validated on disabled populations (Whittlestone et al. 2019). This article addresses that gap directly.
The Validation Deficit
Kenya's National Artificial Intelligence Strategy 2023–2027 identifies healthcare as a priority application domain for artificial intelligence and signals an intention to develop governance standards (Republic of Kenya, Ministry of ICT 2023). That signal is necessary. What remains absent is the specific mechanism by which a clinical AI tool would be validated against Kenyan clinical presentations before deployment — a pre-deployment validation infrastructure equivalent to what India has begun to build.
The Data Protection Act of 2019 requires that health data be processed on specified lawful bases and with explicit consent (Republic of Kenya 2019, s. 30). The Office of the Data Protection Commissioner has issued guidance on the processing of health data (ODPC 2022). Neither instrument currently addresses the secondary use of clinical data for AI training or the question of how a model trained on such data is validated before clinical deployment. The gap between what the Data Protection Act requires for data collection and what responsible AI deployment in healthcare requires for model validation is the precise governance lacuna this article identifies.
A tool trained on populations it was not designed to serve will systematically misclassify those it was deployed to help. In healthcare, misclassification is not a metric. It is a clinical outcome.
India's BODH Platform: A Transferable Precedent with Honest Limitations
India offers the most directly applicable precedent from the Global South. The BODH platform, developed by IIT Kanpur in collaboration with India's National Health Authority, functions as an independent benchmarking layer between AI developers and clinical deployment. Tools must be tested against real-world Indian clinical data on the platform before deployment within the Ayushman Bharat Digital Mission ecosystem (National Health Authority India 2024). Developers do not self-certify. A designated technical body runs the test.
The transferable principle is not the specific platform but the structural choice it represents: validation infrastructure as public goods infrastructure, built once and used by every subsequent tool, rather than negotiated bilaterally between each vendor and each purchasing facility. The limitation of the Indian model is equally instructive. BODH is designed for tools operating within the ABDM ecosystem and does not yet address community health worker tools or disability-specific diagnostic applications — precisely the domains most relevant to Kenya's conditions.
Three Steps Available Now
First, KEMRI, in coordination with the Kenya Medical Practitioners and Dentists Council, can be mandated to operate a national clinical AI validation registry — modelled on India's BODH approach — requiring any AI diagnostic or triage tool deployed in a public facility to demonstrate performance on a representative Kenyan clinical dataset before procurement approval.
Second, the Ministry of Health can issue a procurement circular requiring every AI-enabled clinical tool tender to disclose its training population, validation methodology, and disability-inclusive performance data as a condition of bid eligibility — a paperwork requirement, not a new institution.
Third, the ODPC can extend its existing Guidance Note on the Processing of Health Data to explicitly address secondary use of clinical data for AI training, closing the identified legislative gap without waiting for primary legislation.