A0807
Title: Uncertainty assessment for reinforcement learning for health care in the era of AI
Authors: Junwei Lu - Harvard T.H. Chan School of Public Health (United States) [presenting]
Abstract: Reinforcement learning has emerged as a key framework for clinical decision-making, yet challenges remain in handling dynamic contexts and dependent outcomes. A novel statistical framework that enables both optimal decision-making and rigorous uncertainty assessment in RL is presented, ensuring robust model evaluation. An efficient decision strategy with theoretical guarantees is introduced, and its effectiveness is demonstrated in ranking large language models for medical applications. The critical need for multi-institutional clinical decision-making in personalized medicine is addressed by proposing a multi-site Markov decision process that accommodates both shared and site-specific effects. The federated policy optimization algorithm enhances communication efficiency while preserving privacy, offering practical solutions for real-world healthcare settings such as sepsis management. Through theoretical insights and empirical validation, the importance of uncertainty assessment in RL is highlighted for healthcare, bridging AI advancements with practical, data-driven medical decision-making.
