Machine learning improves classification of serious adverse events compared to threshold-based continuous vital sign monitoring alerts

Abstract Continuous Vital Sign Monitoring (CVSM) enables early detection of patient deteriorations, yet current alert algorithms are prone to false positives from missing data and artefacts, causing alert fatigue, particularly in low-staffed hospital settings such as general wards. Machine learning (ML) may solve this by capturing vital sign trajectories. This study compared four ML analyses of seven vital signs against a National Early Warning Score threshold-severity algorithm. Data came from 2325 patients monitored during major surgery or acute medical admission, including continuous and semi-continuous vital signs, physician-curated Serious Adverse Events (SAE), and patient metadata. At a fixed false positive rate matching the threshold system, the 24-hour ML model achieved a true positive rate of 0.96 versus 0.50. At a matched true positive rate, the false positive rate was 0.06 versus 0.84. The ML approach out-performed threshold-based alerts for classifying SAE intervals at 24 h (AUROC = 0.81) and 8 h (AUROC = 0.71). Precision-recall and calibration analyses showed moderate precision for the 24-hour model and low precision at 8 h under class imbalance. Performance differed across two same-region cohorts, thus geographically independent validation remains necessary. These findings indicate ML-based vital sign monitoring could improve SAE detection over current threshold-based systems.