Federated Learning for Privacy-Preserving Psychiatric Decision Support: A Simulation Proof-of-Concept for Multi-Institutional Collaborative Risk Prediction

Psychiatric decision support systems hold promise for improving clinical outcomes, yet their development is hindered by data privacy regulations and institutional silos that prevent aggregation of sensitive patient information across healthcare facilities. This proof-of-concept simulation demonstrates that privacy-preserving federated learning can match centralized training performance under synthetic non-IID conditions; real-world validation on operational electronic health record data is required before clinical or regulatory conclusions can be drawn, enabling collaborative training of psychiatric readmission prediction models without centralizing raw patient data. Five simulated hospitals with non-independent and identically distributed data participated in federated training of neural network models over 20 communication rounds. We compared standard Federated Averaging (FedAvg) with differentially private federated learning (DP-FL, ε = 1.0) against a centralized baseline. The federated model achieved mean AUC-ROC of 0.800 (95% CI: 0.795 - 0.805), statistically equivalent to the centralized approach (AUC = 0.802, p = 0.42) while preserving data locality. DP-FL maintained strong performance (AUC = 0.806) with formal privacy guarantees. Per-hospital performance varied substantially (AUC range: 0.761 - 0.822), reflecting real-world data heterogeneity. Feature importance analysis identified medication adherence, PHQ-9 depression scores, and prior hospitalizations as top predictors. Communication costs were reduced 500-fold compared to raw data centralization. This federated learning framework demonstrates that privacy-preserving collaborative machine learning can achieve centralized-level predictive accuracy for psychiatric risk stratification while maintaining institutional data sovereignty and regulatory compliance.Subject AreasPsychiatry & Psychology