TITLE:
Federated Learning for Treatment Response Prediction in Bipolar Disorder: A Simulation-Based Institutional Study
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Federated Learning, Bipolar Disorder, Treatment Response Prediction, Neural Networks, Healthcare Privacy, Feature Importance, Synthetic Data
JOURNAL NAME:
Open Access Library Journal,
Vol.12 No.8,
August
18,
2025
ABSTRACT: Bipolar disorder is a complex psychiatric condition characterized by high variability in treatment response, posing a major challenge for clinicians striving to personalize care. Traditional machine learning models often require centralized data access, which is typically restricted in healthcare settings due to privacy regulations and institutional barriers. To address this, we propose a Federated Learning (FL) framework that enables collaborative model training across multiple institutions without the need to share sensitive patient-level data. In this study, we generated synthetic datasets representing five distinct hospitals, each comprising 1000 virtual bipolar disorder patients. These datasets included a range of features encompassing demographic characteristics (e.g., age, gender), clinical history (e.g., illness duration, episode counts), and comorbid conditions. A fully connected neural network was trained using a federated approach over 15 communication rounds. Each institution performed local training, followed by weight averaging to update a shared global model. Our global model achieved an overall test accuracy of 73.0% and an area under the receiver operating characteristic curve (AUC) of 0.84, demonstrating robust performance across diverse simulated institutional settings. Performance improved steadily over training rounds, with the highest-performing site reaching 84.5% accuracy. Permutation-based feature importance analysis revealed that illness duration and baseline mood were the most predictive features of treatment response. These results support the feasibility of federated learning for psychiatric prediction tasks, offering a promising path toward building generalizable, privacy-preserving models in real-world healthcare environments. Future work will extend this framework to real electronic health records and explore fairness-aware training to address institutional bias and population heterogeneity.