Developing an Original Federated Machine Learning Framework for Privacy-Preserving Health Surveillance in Nigerian Multi-Institutional Contexts

Federated machine learning enables model training across distributed data sources without centralising sensitive data, offering a privacy-preserving approach to national health surveillance that is particularly valuable in Nigeria where health data fragmentation across institutions constrains epidemiological research. This study develops an original federated machine learning framework for privacy-preserving health surveillance in Nigerian multi-institutional contexts. The framework introduces three original technical contributions: a differential privacy noise calibration mechanism adapted for the heterogeneous dataset size distributions characteristic of Nigerian state and federal health facilities, a secure aggregation protocol optimised for intermittent network connectivity, and a fairness-preserving model aggregation strategy that prevents geographically dominant institutions from disproportionately influencing the global model. The framework was implemented and evaluated on a simulated federation of twelve health facility data silos representing disease surveillance data from three geopolitical zones, with Lassa fever incidence prediction as the clinical use case. Federated model performance was compared against centralised and isolated local training baselines. Available federated learning healthcare literature from Sub-Saharan Africa identifies communication overhead management and non-IID data distribution as the two most significant performance challenges. The Federated Learning Theory by McMahan and the Privacy by Design Framework provide the theoretical basis. Findings demonstrate that the proposed framework achieves within 4.2 percent of centralised model accuracy while providing formal differential privacy guarantees. Keywords: federated machine learning, privacy-preserving, health surveillance, Nigeria, differential privacy.