Improving malaria prediction in Ghana requires data from across its health system, yet Ghana’s Data Protection Act (Act 843) restricts inter-institutional data sharing, and many facilities decline to transfer patient records regardless of legal permission. Federated learning (FL) offers a solution: each site trains a local model and shares only weight updates, not raw patient data. Whether FL holds up under Ghana’s 10-fold regional prevalence difference had not been tested. Using a controlled simulation framework, we partitioned Ghana Demographic and Health Survey and Malaria Indicator Survey data (2016–2022, n = 10,287 children aged 6–59 months) into five regional clients and evaluated FedAvg and FedProx under three scenarios: uniform distribution (S1), real-world prevalence variation from 2.9% to 30.1% (S2), and heterogeneity combined with 5–20% missing data per client (S3). When data was uniformly distributed, FedAvg matched centralized logistic regression (AUC-PR 0.8852 vs. 0.8854; p = 0.057). The performance drop under 10-fold heterogeneity was only 2.21%, far below the 20–55% degradation seen in vision benchmarks. When regional heterogeneity compounded with missing data quality issues, FedProx outperformed FedAvg (AUC-PR 0.8725 vs. 0.8684; Cohen’s d = 1.257). Federated models retained 97.8–98.5% of centralized AUC-PR without sharing patient data. FL is a feasible privacy-preserving approach for malaria predictions in diverse sub-Saharan African health systems, and the selection of algorithms has a significant impact on the actual performance. These benchmarks are based on a full clinical symptom panel at point of care; with the DHS-native features alone, AUC-PR is 0.34–0.37, compared to 0.87–0.89 with the full feature set (S1 Text, sensitivity analysis). Prior to deployment, prospective validation with facility collected records is required. A critical fairness gap exists: Greater Accra’s 2.9% prevalence produced an approximately 50% false negative rate, meaning half of urban malaria cases would be missed. Prevalence-aware aggregation is required.
Kovor et al. (Fri,) studied this question.