Development Of a Federated Learning Model for Fair Healthcare Data Analysis in Selected African Countries

Abstract

Federated learning (FL) offers a privacy-preserving approach to collaborative machine learning by enabling model training across decentralised health datasets without sharing sensitive individual-level data. This study investigated the fairness and reliability of FL for predicting antenatal care (ANC) utilisation in low-resource African settings, where intermittent connectivity, client dropout, limited computational capacity, and heterogeneous data may affect model performance. Demographic and Health Survey microdata from Nigeria (2024, n = 13,594) and Kenya (2022, n = 10,380), comprising 23,974 observations and 15 predictor features after pre-processing, were used to develop a horizontal FL simulation. Data was partitioned across four simulated municipal clients according to country and urban-rural strata to represent heterogeneous, non-identically distributed conditions. Federated Averaging (FedAvg) and coordinate-wise median aggregation were evaluated under clean conditions and three injected failure conditions such as client dropout, stale updates, and Byzantine corruption, while non-IID characteristics were represented through the client partitioning strategy. Logistic Regression and XGBoost served as centralised baselines. Model explanations were assessed using SHAP, with explanation stability measured using Top-K Jaccard similarity and Kendall’s tau rank correlation, while fairness was evaluated using demographic parity difference and equal opportunity difference across wealth quintiles and urban-rural strata. XGBoost achieved the strongest centralised performance (AUC = 0.7449; F1 = 0.7419). FedAvg achieved a slightly higher AUC of 0.7549 under clean federated conditions. Moreover, Byzantine corruption reduced FedAvg performance to near-random prediction (AUC = 0.4988), whereas coordinate-wise median maintained an AUC of 0.7116. Stale updates produced the greatest fairness degradation, while coordinate-wise median generally demonstrated lower wealth-based disparities than FedAvg. Explanation stability was also limited across all conditions, with Top-K Jaccard similarity below 0.82 and Kendall’s tau below 0.47 relative to the centralised benchmark. Education level and wealth quintile consistently emerged as the strongest predictors of ANC utilisation. The findings demonstrate that FL is viable for privacy-preserving maternal health prediction but requires robust aggregation, stale-update mitigation, and joint assessment of fairness and explanation stability. The observed socioeconomic disparities also highlight the need for complementary policy interventions beyond model design.