Abstract
Medical world models aim to learn latent representations of patient physiology
and transition dynamics that predict future states under temporal evolution
and clinical interventions. However, their reliability is challenged by two
major issues: distribution shift caused by fragmented multi-site clinical
data, and confidence misalignment during multi-step forecasting. We present
\textbf{CalTwin}, a unified regularisation framework that combines
Fisher-Information-based parameter regularisation for mitigating
fragment-level shift with a Confidence Misalignment Penalty for improving
prediction reliability. The proposed formulation adapts these principles to a
latent transition predictor of a medical world model, where Fisher
regularisation is applied to the transition likelihood rather than
classification objectives, and confidence calibration is performed through a
shared clinical prediction head. CalTwin is evaluated on multi-hospital ICU
time-series data using the PhysioNet 2019 Sepsis Challenge, with sequential
hospital fragments for training and an unseen hospital for out-of-distribution
evaluation. CalTwin reduces OOD next-step latent-state mean squared error by
9.1\% compared with an unregularised baseline, with Fisher regularisation
providing the primary contribution. Confidence calibration improvements are
modest, with expected calibration error reduced by 0.7\% using CalTwin. An
additional evaluation on eICU-CRD demonstrates that the contribution of each
component depends on the characteristics of the clinical distribution shift.
These results highlight the potential of combining shift-robust learning and
confidence-aware modelling for medical world models while motivating further
validation on larger-scale multimodal clinical datasets, including medical
imaging.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MWM_122.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MWM_122_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=WdlWc3QG3x
BibTex
@InProceedings{KhaBeh_CalTwin_MICCAISAT2026,
author = { Khan, Behraj AND Syed, Tahir Qasim AND Ahmad, Shabir AND Vigneron, Vincent Martin AND Bukhari, Syed Ahmad Chan},
title = { { CalTwin: Calibrated Shift-Robust Medical World Models via Fisher-Information Regularization } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}