Federated Reinforcement Learning for Privacy-Preserving Sepsis Patient Treatment Model
Songmi Oh, Yun‐Ho Choi, Ho-Taek Joo, Kyung-Joong Kim · ACM Transactions on Intelligent Systems and Technology · 2025
Reinforcement learning (RL) for developing a patient treatment model using electronic health records has been actively studied. Although constructing the models requires considerable actual patient treatment records, the establishment of large databases poses challenges due to strict privacy regulations. Therefore, federated RL (FRL), which can train an RL model without sharing data between institutions, is being introduced. This study proposes an FRL framework where local institutions collaborate to make optimal RL models without data sharing or raw data leakage. We constructed FRL models for personalized sepsis treatment models and evaluated their performances in realistic scenarios. The reliability of the FRL framework was evaluated on basic, skewed, imbalanced, and realistic data distribution using two clinical benchmark datasets, the Medical Information Mart for Intensive Care III database, and the eICU Collaborative Research Database v2.0. The performances of FRL models were comparable to those learned from the ideal setting, where all institutions agree to share their datasets to train a global treatment model. Furthermore, the FRL framework showed generalization performance on unseen data during training and showed the applicability of various federated learning algorithms. Through practical experiments using clinical data, we demonstrated the real-world applicability of the FRL framework.