Analytical Evaluation of Federated Learning for Privacy-Preserving Healthcare Data Analytics in Nigeria

Healthcare data sharing across Nigerian hospitals is constrained by privacy regulations, institutional competition, and technical incompatibility, limiting the scale of data available for AI model training. Federated learning offers a privacy-preserving collaborative training architecture that addresses these constraints without centralising patient data. This study analytically evaluates federated learning for healthcare data analytics in the Nigerian context, using diabetic patient records from four hospitals in Kano and Abuja as a case study. A federated learning framework was implemented using PySyft and TensorFlow Federated, with each hospital simulating a federated participant. Two prediction tasks were modelled: 30-day hospital readmission and HbA1c level classification. Federated model performance was compared against centralised training (as a performance ceiling) and locally trained models (as a baseline). Communication round efficiency and differential privacy cost were analysed. Across 50 federated rounds, the federated model achieved 91 percent of centralised training performance on readmission prediction (AUC-ROC 0.824 vs 0.907). Local models achieved only 0.71 to 0.78 AUC-ROC depending on individual hospital dataset size. Differential privacy noise addition with epsilon of 8 reduced AUC-ROC by 3.2 percentage points. The study fills a research gap in federated learning evaluation for African healthcare settings and demonstrates that federated learning can enable collaborative predictive analytics with acceptable privacy-accuracy trade-offs. Piloting the framework across Abuja teaching hospitals is recommended.

Keywords: federated learning, healthcare analytics, privacy-preserving AI, Nigeria, diabetes prediction