Abstract
Medical world models should move beyond static risk scores by representing a patient’s evolving state, modeling how that state changes, and exposing uncertainty when observations are incomplete. We present MedWorld-Lite, a compact framework for medical world modeling from irregular intensive-care data. The framework combines eight routinely observed physiological variables with explicit observation masks and time-since-last-measurement channels, personalizes the latent patient state using admission context, evolves that state with a recurrent transition, and decodes probabilistic future physiology together with an auxiliary mortality risk. We design the temporal benchmark on the PhysioNet/Computing in Cardiology Challenge 2012 Set A using 24 h of history and 6/12/24 h future rollouts. Because raw sequence bytes were not available in the execution environment used to prepare this submission, we report only experiments that were actually run: a 4,000-patient processed Set-A state-validation study. Across five stratified splits, adding personalization to physiological state variables increased AUROC from 0.753 ± 0.022 to 0.794 ± 0.016; the complete state representation reached 0.795±0.018. On a held-out split, a random forest achieved 0.857 AUROC (95% bootstrap CI 0.823–0.889) and 0.469 AUPRC. Hiding 50% of available physiological state features reduced logistic-regression AUROC from 0.820 to 0.638, demonstrating substantial sensitivity to observation loss. The study therefore contributes a transparent, reproducible foundation for lightweight ICU medical world modeling without reporting unexecuted temporal results.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MWM_116.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=lmeLRZOj2A
BibTex
@InProceedings{FahNaf_MedWorldLite_MICCAISAT2026,
author = { Fahad, Nafiz},
title = { { MedWorld-Lite: A Missingness-Aware Personalized Framework for ICU Medical World Modeling } },
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}
}