Expert Augmented Prediction of Circulatory and Respiratory Instability from High Resolution Vital Signs

Timely detection of circulatory and respiratory instability (CRI) is critical in intensive care units (ICUs), yet existing early warning systems often rely on single-parameter indices that underutilize continuous vital-sign data or on delayed, difficult-to-interpret multimodal clinical data. Leveraging routinely collected high-frequency vital-sign monitoring, we developed an interpretable, expert-augmented early warning system based on 1-second-resolution heart rate, blood pressure, respiratory rate, and oxygen saturation data. Machine‑learning models were trained on 627,958 h of continuous vital‑sign data from 1702 ICU patients at the First Affiliated Hospital of Sun Yat‑sen University and externally validated in the MIMIC‑III cohort. Models incorporating trend-based, reference, and statistical features derived from vital-sign trajectories achieved strong predictive performance (AUROC > 0.8) in both internal and external validation, outperforming conventional single-parameter indices and achieving performance comparable to models incorporating laboratory and demographic variables. Increasing temporal resolution improved predictive accuracy, with trend-based features contributing most strongly to model predictions. To improve clinical interpretability, tree-based models were transformed into physiologically meaningful decision rules and refined through expert-augmented learning, resulting in the Expert-Augmented Early Warning System (EAEWS). EAEWS generated accurate, low-frequency alerts with transparent explanations aligned with bedside monitoring, and may provide a scalable framework for real-time CRI detection in ICUs.