PREDICTING PATIENT INSTABILITY NONINVASIVELY FOR NURSING CARE (PPINNC-3) Project Summary Timely recognition and forecasting of cardiorespiratory instability (CRI) in hospitalized patients in step-down units (SDU) has clear implications for strategies to reduce preventable morbidity and mortality. Continuous noninvasive vital signs (VS) monitoring is widely used to facilitate nurse detection of actionable events requiring a diagnostic or therapeutic response, yet failure-to-rescue rates in US hospitals, defined as death due to complications, remain high. Traditional VS monitoring alarms are largely based on numeric threshold exceedance, translating to very low true positive rates and adversely leading to alarm fatigue and reactive nursing care. We and others have demonstrated that decompensation evolves over time and featurization, trending, and phenotyping of multi-channel VS time series for building relevant models for forecasting CRI is feasible. Over the past two funding periods, we have built the largest multi-site database of EHR-linked, high fidelity VS monitoring data from SDU patients known to us. Using this multi-expert, multi-tier ground truth annotated database we have begun to build clinically relevant Machine Learning models to discriminate artifactual anomaly from real CRI with high accuracy, as well as to classify mild vs. serious CRI and forecast cases that require escalation of care or up transfer to intensive care. We now aim to move these extensive efforts to fruition and refine and build these models in the workflow for prospective validation and clinical deployment. The specific aims of this renewal application are: 1) build and deploy a real-time data streaming architecture at bedside for an intelligent alerting system, including the iterative design and usability testing of a clinician-facing graphical user interface; 2) build and externally validate a multi-layered alerting system forecasting CRI; and 3) perform a prospective validation of the intelligent alerting system, including silent deployment and evaluation at SDUs at UPITT and UCSF followed by prospective field testing at a 16-bed SDU at UPITT. The final deliverable is an intelligent alerting system for detection and mitigation of CRI in SDU patients of sufficient readiness to be deployed in a multicenter human effectiveness trial as a next step. Building such a clinically relevant system in clinical workflow to predict patient instability has important implications for reducing preventable morbidity and mortality, eliminating alarm fatigue, improving patient safety, nursing care logistics (monitoring frequency, case load and mixture, staff allocation) and care delivery systems (triage, bed allocation, prevention of adverse events).