Abstract
Static ICU risk models fail to track rapid hour-to-hour physiological changes in critically ill patients. To address this, we introduce ICUWorld-QR, a residual latent world model designed for hourly ICU trajectory forecasting. The model combines a history-conditioned encoder, a transition module driven by weak observational context (FiO2 and ICU hour), and a decoder that predicts residual corrections relative to a frozen linear baseline. Crucially, we identify a pitfall in residual architectures with strong linear priors: the prior removes the gradient incentive to build a meaningful latent state, causing it to collapse into a constant. We demonstrate that enforcing a batch-wise standard-deviation hinge constraint prevents this collapse, restoring both forecasting performance and sepsis ranking, whereas a JEPA-style objective does neither. On the PhysioNet Challenge 2019 dataset (standardized units), ICUWorld-QR improves upon last-value persistence by 7.5% overall (masked MAE 0.333 vs. 0.360) and by 15% on hard hours (top quartile of next-hour vital change). Furthermore, its learned latent state ranks next-hour sepsis onset (AUROC 0.773), and this readout holds when we train on one of the two Challenge hospital systems and test on the other, at 0.693 and 0.657. Quantile prediction heads on the latent produce wider intervals on volatile hours, whereas a residual ridge emits one fixed width for all hours. Finally, zero-shot evaluation across 204 eICU hospitals confirms that the frozen model outperforms persistence and matches a Challenge-fitted ridge over three seeds. Fine-tuning yields modest, seed-stable adaptation gains not attributable to renormalization, though it offers no advantage over the frozen initialization when target hospitals are scarce.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MWM_119.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MWM_119_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=N4QYWqE3vx
BibTex
@InProceedings{TanArt_ICUWorldQR_MICCAISAT2026,
author = { Taneja, Arti AND Omar, Hasmila A. AND Vasikarla, Shantaram AND Ranjan, Priya},
title = { { ICUWorld-QR: A Residual Latent World Model for Hourly ICU Physiology } },
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}
}
