Privacy-Preserving Federated Learning for Multi- Hospital Risk Prediction: An Illustrative Healthcare Analytics Study

This empirical-style paper presents a simulated study at the intersection of Healthcare Informatics, Computer Science, and Privacy Engineering. The study addresses a current research direction in which digital and data-driven methods are being integrated with domain-specific decision systems. Six synthetic hospitals contributed between 1,900 and 3,900 simulated records each, creating deliberate class imbalance and non-identically distributed predictor profiles. A simulated cross-silo dataset representing six hospitals was created with 18,000 patient records and 24 routinely available predictors including age band, vital-sign summaries, comorbidity counts, medication burden, laboratory trends, and prior utilization. No real patient information was used. A binary 30-day deterioration outcome was simulated at a prevalence of 17.8%. Logistic regression and gradient-boosted trees were tested locally, centrally, and under federated averaging. The privacy-enhanced model added clipped updates, Gaussian noise, and secure aggregation. Performance was assessed using AUROC, F1, Brier score, calibration slope, subgroup recall, and communication rounds. The standard federated model approached the performance of the centralized benchmark while preserving data locality. The privacy-enhanced federated model showed a small AUROC reduction but improved the modeled privacy posture. Local-only models were less stable across smaller hospitals and showed greater subgroup variation. Communication-efficient update compression reduced transmitted model payload by 42% with minimal loss in AUROC. All numerical observations, experimental assignments, outcomes, and performance values are simulated for methodological illustration; no real participants, institutions, farms, factories, hospitals, or city systems were studied. The results are intended to demonstrate a coherent research design, provide testable hypotheses, and show how a future real-world study could report evidence transparently. The paper concludes with implementation constraints, methodological limitations, and priorities for empirical validation.