Privacy-Preserving Federated Learning for Multi-Site Maternal Mortality Prediction Under Heterogeneous Electronic Health Record Schemas: An Asynchronous Aggregation Evaluation of Predictive Parity Across District Hospitals

Maternal mortality prediction models trained at single district hospitals rarely transfer to other sites because electronic health record (EHR) schemas differ in variable naming, coding vocabularies, measurement cadence and missingness structure. Centralising records to pool such data conflicts with patient privacy and with the data-governance constraints under which district hospitals operate. The framework combines a schema-agnostic feature encoder, a differential-privacy mechanism with a per-round privacy budget, and an asynchronous aggregation rule that admits site updates at heterogeneous arrival times. We derive a convergence bound for the asynchronous rule under bounded staleness and show that the bound degrades gracefully with the staleness parameter and with the privacy budget. A replication study over 24 computed district-hospital sites with deliberately misaligned schemas was conducted to evaluate predictive parity, measured as the between-site dispersion of area under the receiver operating characteristic curve. Asynchronous aggregation reduced the between-site dispersion of predictive performance relative to synchronous aggregation when site update latencies were heterogeneous, while differential privacy at moderate budgets preserved parity at a small cost in mean performance. The results indicate that predictive parity across district hospitals is achievable under schema heterogeneity, but only when the aggregation rule is designed for the latency structure of the participating sites.