Year: 2026
Type: preprint
Abstract: Intensive care unit (ICU) monitoring faces a critical challenge: translating continuous physiological data into actionable insights for real-time clinical decisions. We developed STREAM (State Trajectory Representation & Evolution-Aware Monitoring), which applies optimal transport theory to identify distinct physiological states from routine ICU data and maps individual patients onto state progressions. Using the multicenter eICU Collaborative Research Database (N=158,294), STREAM identified five reproducible states. Patients whose measurements differ substantially from their assigned state (state outliers) showed 9-fold higher mortality (45.7%) versus patients who remained inside state boundaries (5.1%). State-based features predicted mortality with AUROC 0.863 at 8 hours and 0.903 at 72 hours, with excellent calibration error score (0.002). External validation on MIMIC-IV (N=84,517) demonstrated robust performance (AUROC 0.798 and 0.857, respectively), with state outliers exhibited a 7.1-fold higher mortality risk (41.1% vs. 5.8%). Importantly, STREAM connects this risk stratification to the underlying clinical measures defining each physiological state, providing accurate mortality predictions and interpretable insight.
Source: medRxiv
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7127644497