Abstract
Early prediction is essential for timely intervention in acute respiratory distress syndrome (ARDS), especially within the first 12 hours, when optimizing ventilation strategy and fluid management can mitigate progression to severe hypoxemia. Chest X-ray (CXR) is non-invasive and routinely acquired in the ICU, and it has potential value for the detection, evaluation, and monitoring of ARDS. However, most prior ARDS-assessment AI approaches focused on modeling the current static status as a discrete classification problem, without explicitly incorporating temporal dynamics or underlying physiologic constraints. In this context, we present ArdsTracker, a physiology-informed (PI) framework designed to characterize the dynamic evolution of key gas-exchange (GE) markers, i.e., the clinical indices of ARDS severity. Specifically, it forecasts future trajectories of PaO₂/FiO₂ (PF ratio) and PaCO₂ from a single baseline CXR. To address irregular temporal sampling across patients and the challenges of long-horizon prediction, we develop a Trajectory-Aware Dual Attention (TADA) module to highlight trajectory-relevant features. Furthermore, to enhance physiologic plausibility, we impose a gas-exchange-guided loss informed by clinical prior knowledge. ArdsTracker was evaluated on 4,744 studies from MIMIC-CXR and MIMIC-IV using 5-fold cross-validation. ArdsTracker non-invasively improves GE marker trajectory regression, reducing MSE and MAE by 6.5% and 4.0%, respectively. It also achieves superior ARDS-related risk stratification across multiple time points, with AUCs at 3, 6, 9, and 12 hours of 0.756, 0.780, 0.774, and 0.802, respectively. Moreover, for more immediate clinical use (within 1–3 hours), point-wise predictions of the PF ratio and PaCO₂ exhibited significant separation between subsequent ARDS and non-ARDS events, further supporting early ICU decision-making. Code will be released publicly.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AMAI_014.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/profile?id=~XIN_WANG45
BibTex
@InProceedings{WanWei_ArdsTracker_MICCAISAT2026,
author = { Wang, Weiru AND Wei, Jianming AND Kraaijveld, Rosa Cornelia Jacoba AND Zhou, Hong-Yu AND Ursprung, Stephan AND Sun, Yue AND Tan, Tao AND Gao, Yuan AND Wang, Xin},
title = { { ArdsTracker: Time-series physiological neural networks for predicting respiratory failure based on chest X-ray } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17273},
month = {pending},
page = {pending}
}