Federated Learning for Suicide Risk Prediction across Heterogeneous Hospitals Using Privacy-Preserving Synthetic Data

Accurate suicide risk prediction in clinical practice is hindered by stringent privacy regulations, fragmented data ownership, and pronounced heterogeneity across healthcare institutions in patient demographics, symptom severity, and social determinants of health. To address these challenges, we propose a federated learning (FL) framework for binary suicide-risk stratifi-cation (high-risk vs. lower-risk) that enables collaborative model training across hospitals without sharing raw patient data. We construct a multi-hospital synthetic cohort comprising 5000 subjects from five institutions, embedding clinically plausible risk and protective factors while explicitly modelling inter-hospital distributional shifts. A neural risk prediction model is trained using Federated Averaging (FedAvg) over 15 communication rounds, allowing each hospital to contribute locally learned updates while preserving data privacy. The proposed FL approach achieves a final global accuracy of 0.942 and a global AUC-ROC of 0.9568, closely matching centralized training performance (0.945 accuracy; 0.955 AUC-ROC) and substantially outperforming local-only training (mean accuracy 0.930; mean AUC-ROC 0.8962). Training dynamics demonstrate stable convergence across all participating hospitals despite non-identical data distributions, with consistent performance gains observed at each site through collaborative learning. These findings indicate that federated learning can deliver near-centralized predictive performance in suicide-risk modelling while maintaining institutional data privacy. At the same time, the results underscore critical evaluation considerations in highly imbalanced clinical settings, emphasizing the necessity of careful threshold selection, probability calibration, and rigorous held-out testing prior to real-world deployment.