B104-14 Stream: AI Algorithm for Real-time Monitoring of Intensive Care Unit (ICU) Patients

Rationale ICU monitoring systems face a critical gap: translating continuous physiological data into actionable insights that guide real-time clinical decisions. Existing approaches rely on static thresholds or severity scores that fail to capture dynamic disease progression patterns. These systems provide limited insight into why patients deteriorate, or which interventions might be most effective. To address this gap, we developed STREAM (State Trajectory Representation & Evolution-Aware Monitoring), which models each patient as a point in a multidimensional physiological space and applies geometric analysis of routine ICU data to track deterioration and inform interventions. Methods STREAM analyzes 26 routinely collected clinical features including vital signs and laboratory values. Using optimal transport theory, STREAM applies a voting algorithm to construct population-level states from clinical features and maps each patient to their nearest state. For mortality prediction, we trained a gradient boosting model (XGBoost) on STREAM-derived features. We evaluated STREAM using publicly available patient data, analyzing trajectories over the first 10 days of ICU stay. We used data from the multi-center Collaborative Research Database (eICU-CRD; N = 158,294) to develop our model and from MIMIC-IV (N = 84,517) for external validation. Results STREAM identified five reproducible states with distinct clinical features (Figure 1A). Patients who spent less than 10% of their ICU stay within their expected state (n = 5,717) had a high mortality of 45.7% (Figure 1B). This is 9 times higher than those who remained within their assigned states (5.1%; n = 126,242). Our mortality prediction model demonstrated strong performance, achieving an AUROC of 0.840 using patient data acquired at 8 hours (one time point) and improving to 0.892 at 72 hours (six time points). Calibration was excellent, with a minimal calibration error of 0.002. External validation using MIMIC-IV maintained similar performance (AUROC 0.796 and 0.853, respectively). Feature importance analysis revealed that specific laboratory values and vital signs drive each state transition, providing interpretable clinical explanations. Conclusions STREAM enables real-time, transparent monitoring of patient trajectories across population-defined physiological states. By linking physiological dynamics with outcome prediction, STREAM enables earlier recognition of deterioration and provides interpretable, data-driven explanations to support bedside decision-making. STREAM identifies the specific laboratory values and vital signs driving each state transition. Its strong discrimination, excellent calibration, and reproducibility across large multicenter datasets underscore the method’s robustness and clinical scalability.