TITLE:
Federated Learning for Privacy-Preserving Antidepressant Safety Surveillance across Multi-Site Bipolar Disorder Registries
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Federated Learning, Privacy-Preserving, Antidepressant Safety, Bipolar Disorder, FedAvg, FedProx, Differential Privacy, Pharmacovigilance, GDPR, Multi-Site Registry
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.9,
September
17,
2026
ABSTRACT: Antidepressant safety surveillance in bipolar disorder requires large, multi-site patient populations to achieve reliable pharmacovigilance signal detection yet the raw patient data necessary to train centralized machine learning models are subject to strict privacy regulations (GDPR, HIPAA) that prohibit cross-site data sharing in most jurisdictions. Federated learning (FL) offers a principled solution: locally trained models share only gradient updates or model weights with a central aggregation server, enabling collaborative learning across institutions without any raw data leaving the originating site. No federated learning framework has been applied to antidepressant safety surveillance in BD, and the privacy-utility tradeoff of differential privacy augmentation in this clinical context remains uncharacterised. We implemented and evaluated a Federated Averaging (FedAvg) framework across five fully simulated, heterogeneous bipolar-disorder registry sites (total N = 800; Sites A-E represented academic, community, European, primary-care, and specialised bipolar-clinic settings). A multilayer perceptron was trained locally for 15 communication rounds with five local epochs per round. Comparators included local-only models, centralised MLP and XGBoost reference models, FedProx, and differentially private FedAvg (DP-FedAvg). A privacy-utility analysis examined seven nominal privacy settings. The composite outcome was a simulated SSRI-associated safety event within six months of initiation, comprising mood switch, cycle acceleration, or unplanned discontinuation. In the held-out simulated test set, FedAvg achieved AUC = 0.957 (95% CI: 0.920 - 0.987), F1 = 0.842 (CI: 0.750 - 0.925), sensitivity = 0.814, and specificity = 0.936. Its point-estimate AUC was slightly higher than the centralised MLP (0.931) and similar to centralised XGBoost (0.954), FedProx (0.956), and local-only averaging (0.950). These small differences should be interpreted as performance equivalence within uncertainty, not evidence that federation intrinsically outperforms centralised training. FedAvg converged within approximately 12 communication rounds, and the simulated privacy-noise analysis showed limited AUC degradation over the tested range. Within this proof-of-concept simulation, federated training preserved predictive performance without sharing raw records. The findings support evaluation on real multi-site bipolar-disorder registries but do not by themselves establish clinical validity, formal GDPR/HIPAA compliance, or deployability.