Federated Learning for Healthcare Data Privacy: A Case Study in Multi-Hospital Collaboration

Federated learning (FL) was assessed as a privacy-safe option to centralized models for predicting hospital readmissions, utilizing 15,200 anonymized patient records from three hospitals (ABC = 5,800; FGH = 4,700; XYZ = 4,700). Cohorts showed variability in average age (55.9-58.7 years), female representation (50–53%), diabetes rates (20-25%), hypertension rates (42-48%), and heart failure incidences (7-9%), with an average duration of stay between 4.9–5.5 days. Performance comparisons (Table 7) indicated that the Multilayer Perceptron (MLP) attained the best AUROC (0.83) and F1-score (0.71), whereas Federated Averaging (FedAvg) was nearly comparable (AUROC = 0.82, F1 = 0.70), reflecting slight decreases of ΔAUROC = –0.01 and ΔF1 = –0.01. Models restricted to local data showed lower performance, with AUROC values ranging from 0.75 to 0.78, while federated learning enhanced per-hospital AUROC by 0.04 to 0.06 (Table 8). The analysis of efficiency (Table 9) showed that FedAvg reached convergence in 45 epochs across 50 communication rounds, utilizing an average bandwidth of 38 MB and requiring 48 minutes for training, which is just 6 minutes more than the centralized MLP, while providing a 14% absolute decrease in the success of membership inference attacks (22% → 8%, Table 10). Error analysis (Table 11) revealed misclassification trends associated with established risk factors: age over 70 with COPD at ABC, a diabetes-hypertension combo at FGH, and more than 10 prescribed medications combined with several previous admissions at XYZ. FL achieved similar predictive accuracy while fully adhering to HIPAA and GDPR/NDPR regulations, proving its effectiveness for healthcare analytics across multiple institutions. These findings offer a numerical approach for implementing FL in practical medical settings, achieving less than 10% privacy risk alongside a maximum of 6 minutes extra training time for reliable, cooperative forecasting of hospital readmissions.