Abstract:
This study presents a privacy-preserving learning model designed for cross-border telemedicine in East Africa that keeps raw patient records in country while hospitals collaborate on model quality. The core of this approach is to keep sensitive patient records localized within each country, with hospitals training models locally and only sharing model updates. Using synthetic EHRs split across seven hospitals in Kenya, Tanzania, and Uganda, we compare centralized training, standard federated learning, and federated learning with differential privacy. Federated learning improves utility while maintaining data localization, with accuracy rising by about 0.0665, recall for the positive class improving by about 0.1193, and F1 increasing by about 0.0657 relative to centralized training. Adding differential privacy made the system more resilient to attacks. The success rate of model-inversion attacks dropped from 0.696 in the centralized training scenario to 0.686 with standard FL and further to 0.638 with FL + DP. This represents an absolute reduction of 0.058, or about 8.4 percent, in attack success. Membership-inference leakage has an AUC of around 0.50. The trade-off is tunable utility at a chosen privacy budget, for example accuracy near 0.530 at ε = 0.30. The originality is practical, we pair federated learning with an attack simulator and an ε register that turns privacy into an auditable setting hospitals can manage during cross-border care.
