AI Framework Optimizes Sepsis Treatment with Digital Twins

Sepsis remains a major cause of mortality, and determining optimal treatment strategies is a complex and often debated challenge. Existing reinforcement learning (RL) approaches for sepsis treatment typically learn fixed policies, which lack the adaptability needed for dynamic clinical scenarios. This new research introduces EHR-MPC, a novel framework designed to overcome these limitations. EHR-MPC operates by first training a "patient digital twin" using a generative electronic health record (EHR) model. This digital twin learns to predict how a patient's condition will evolve under various medical interventions. Crucially, this learning of patient dynamics is decoupled from the actual treatment optimization process. During inference, the framework employs model predictive control (MPC) to plan optimal treatments. This involves simulating multiple future clinical trajectories using the digital twin and selecting the intervention that best aligns with desired clinical objectives. Evaluated on a multicenter ICU sepsis cohort, EHR-MPC achieved comparable off-policy performance and superior simulation performance compared to RL baselines, establishing a flexible framework for real-time, adaptive decision-making in critical care.