Privacy-Preserving Generative AI for Secure Healthcare Data Sharing in Kenya

Kenya's healthcare system holds enormous untapped potential for AI-driven clinical research, but the data that would power it is trapped behind silos, fragmented county systems, and a relatively new privacy law that most existing research tools were never designed to satisfy. This paper proposes a framework that combines Wasserstein Generative Adversarial Networks with differential privacy to generate synthetic healthcare datasets that are statistically realistic, clinically useful, and provably private.

Designed to operate within Kenya's Data Protection Act 2019 and aligned with Africa CDC's emerging health data-sharing guidelines, the system enables secure exchange of electronic health records across county hospitals and national research institutions without exposing patient identities. New utility metrics, including Jensen-Shannon divergence and predictive task validation on malaria severity prediction and maternal mortality risk scoring, demonstrate that synthetic records can support the same clinical AI research as real patient data.

The framework is built for Kenya's infrastructure reality: lightweight enough to deploy on commodity hospital hardware, containerized for integration with the Kenya Health Information System (KHIS), and designed around the 47-county devolved structure. The output is an open-source tool that health researchers across East Africa can use without waiting months for the complex data-sharing agreements that currently block clinical AI development.

Working with health data in Kenya means sitting inside a frustrating standoff. On one side, you have researchers and AI teams who need patient records to build models that could genuinely save lives. On the other, you have hospitals and county health offices holding records they cannot legally share, even internally across counties, because Kenya's Data Protection Act 2019 classifies health information as sensitive personal data with strict handling and transfer restrictions.

The result is paralysis. Kenya's KHIS platform collects aggregate statistics from across the country, but the granular patient-level data needed to train a malaria severity predictor or a maternal mortality risk model sits in paper files and disconnected EMR systems that cannot communicate with each other. The Social Health Authority (SHA), which replaced NHIF in 2024 as Kenya's national health insurance scheme, is building new digital infrastructure under the Universal Health Coverage agenda, but the data pipeline for AI research is still years away from usable.

Traditional anonymization techniques make the problem worse rather than better. K-anonymity degrades data utility and fails on small population clusters, which are common in Kenya's county-level datasets. A person's county, age bracket, and primary diagnosis can be enough for re-identification in a sub-county of 80,000 people. Standard anonymization gives a false sense of compliance while delivering neither genuine privacy nor usable data.

This project applies Wasserstein GANs combined with differential privacy to do exactly that, grounded in Kenya's regulatory environment and validated against the health challenges that matter most here: malaria, HIV, maternal mortality, and the rising burden of non-communicable diseases in urban and peri-urban areas.

Four phases over 24 months, each grounded in Kenya's institutional structure and infrastructure constraints. The sequencing matters: ethics clearance from three Kenyan bodies, not one, is required before data access, and that timeline is realistic only if started early.

Kenya is not the obvious choice for a healthcare AI privacy paper. The EHR datasets are smaller than MIMIC-III. Digital record-keeping is still partial across county facilities. The Data Protection Act 2019 is less than six years old and enforcement is still being established. On paper, it looks like a harder problem than running the same research in a high-income country with mature hospital data infrastructure.

That is exactly the point. Privacy-preserving healthcare AI research is almost entirely designed for high-income contexts with large hospital networks, stable EHR systems, and well-resourced compliance teams. The techniques that work in those environments often break or become impractical in a 47-county devolved system where some facilities still run paper registers and connectivity cuts out during rainy season. Retrofitting a framework designed for MIMIC-III onto Kenyan data does not work cleanly, and the gap between what the research literature assumes and what Kenya's health system actually looks like is significant.

Kenya's specific conditions also create privacy research problems that do not exist in Western datasets. Small county-level population clusters make re-identification risk materially higher, which means standard epsilon values need recalibration for this context. Swahili-English code-switching in clinical notes requires different text preprocessing before synthesis. Missing data patterns from paper-to-digital migration demand different imputation strategies. These are not edge cases to handle later. They are the core research problem.

Kenya also has the institutional base to make this work. KEMRI is one of Africa's strongest health research bodies. The KHIS platform on DHIS2 is nationally deployed and actively maintained. SHA is building new claims infrastructure with digital-first architecture. The Data Protection Act 2019 gives the ODPC authority to define compliant synthetic data practices. The ingredients exist. The question is whether the technical tooling exists that is designed for this context rather than imported from elsewhere and hoped to fit.

The wider beneficiaries extend beyond Kenya. Uganda, Tanzania, Rwanda, and Ethiopia all use DHIS2. Most face the same data access standoff. A DP-GAN framework with a DHIS2 adapter and governance documentation written for the African Union regulatory context is directly transferable across the region in a way that a framework designed for HIPAA compliance is not.