TRACE Models Daily Clinical Trajectories for Inpatient Risk Prediction
Minjun Kim and Jong Hak Moon submitted TRACE to arXiv on October 19, 2026, introducing a daily-scale framework for predicting in-hospital mortality, prolonged length of stay, and readmission. TRACE organizes EHR data, chest X-rays, radiology reports, and discharge summaries into temporal trajectories while allowing unavailable modalities to contribute no tokens. The paper reports the highest AUROC across all nine dataset-task settings on MIMIC-IV, eICU, and MIMIC-III, with the highest accuracy in eight settings.
The paper proposes a distinct way to model multimodal clinical evidence as day-level trajectories rather than only fusing modalities or compensating for missing data, and evaluates it across nearly 180,000 hospital stays.
Computer Science > Artificial Intelligence
arXiv:2610.25765 (cs)
[Submitted on 19 Oct 2026]
Title:TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction
Authors:Minjun Kim, Jong Hak Moon
View a PDF of the paper titled TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction, by Minjun Kim and Jong Hak Moon