EHR-MPC uses generative digital twins for sepsis treatment

Existing reinforcement learning approaches for sepsis treatment learn fixed policies, limiting adaptability when clinical objectives shift during patient care. The EHR-MPC framework, proposed by Pickard et al., decouples learning patient dynamics from treatment optimization. It trains a generative electronic health record model that acts as a patient digital twin, predicting clinical trajectories under various interventions. At inference time, it applies model predictive control (MPC) to simulate thousands of possible treatment paths and selects the optimal one in real time.

Evaluated on a multicenter ICU sepsis cohort spanning 8 hospitals in the Mass General Brigham health system, EHR-MPC achieved comparable off-policy performance to RL baselines and improved on-policy simulation performance. This approach transforms sepsis treatment from a fixed policy problem into an inference-time control problem, enabling clinicians to adapt plans as new data arrives. The framework is generalizable to other clinical decision-making tasks where generative models can simulate patient responses.