CalTwin Enhances Medical World Models for Shift Robustness.

Medical world models, designed to learn patient or organ physiology and forecast its evolution under interventions, are crucial for tasks like diagnosis and digital-twin treatment planning. However, their clinical deployment faces two major reliability issues: covariate shift, where training data fragmentation across hospitals or scanners leads to differing feature distributions, and confidence misalignment, where multi-step forecasts are often overconfident, especially in high-risk scenarios. This research introduces CalTwin, a lightweight regularization objective that offers a unified solution to both problems. CalTwin integrates a Fisher-Information-based shift penalty, adapted from prior work on fragmented covariate-shift remediation, with a Confidence Misalignment Penalty, adapted from calibrated vision-language classification. This combined objective is applied to a GRU-based medical world model's latent transition predictor. Evaluated on the PhysioNet 2019 Sepsis Challenge, treating different hospital systems as sequential training fragments and an unseen system as an out-of-distribution test, CalTwin demonstrated significant improvements. It reduced out-of-distribution next-step latent-state Mean Squared Error (MSE) by 9.1% compared to a baseline without penalties, with the FIM penalty contributing 7.0%. While the ECE reduction from the Confidence Misalignment Penalty was smaller, the overall approach enhances the robustness and calibration of medical world models.