Year: 2023
Type: article
Study design: Observational Study
Abstract: Intensive care units (ICUs) generate over 1,500 data points per patient per day from continuous physiological monitors,ventilators, infusion pumps, and laboratory systems, yet 86% of clinically actionable deterioration events are preceded byobservable warning signs that go unrecognised in the data stream due to alarm fatigue, information overload, and thecognitive limitations of human monitoring. Real-time intelligent monitoring systems that synthesise multiparameter datastreams into actionable clinical alerts promise to reduce missed deterioration events and improve ICU outcomes. Thisstudy developed and evaluated five monitoring architectures -- threshold-based alarm systems (current standard),Modified Early Warning Score (MEWS), single-parameter trend analysis, multi-parameter statistical process control(SPC), and a proposed AI-integrated clinical surveillance platform (AICSP) combining streaming time-series anomalydetection, multi-organ failure risk prediction, and contextual alert suppression -- in a prospective observational studyacross 3 ICUs (n = 842 patients, 14,286 patient-days). AICSP detected 92.4% of clinical deterioration events (cardiacarrest, unplanned intubation, vasopressor escalation, emergency surgery) with a median lead time of 4.8 hours beforeclinical recognition, while reducing total alarm burden by 68.4% compared to threshold-based systems (from 187 +- 42alarms/bed/day to 59 +- 18; p < 0.001). The positive predictive value of AICSP alerts was 42.6% (vs 6.8% for thresholdalarms and 18.4% for MEWS; p < 0.001), meaning that fewer than 3 AICSP alerts were needed to identify each truedeterioration event compared to 15 threshold alarms. Nursing response time to AICSP alerts was 3.2 +- 1.4 minutesversus 8.6 +- 4.2 minutes for threshold alarms (p < 0.001), reflecting increased clinician trust in the filtered,contextualised alert stream. These results establish AI-integrated clinical surveillance as a transformative approach toICU monitoring that simultaneously improves detection sensitivity, reduces alarm fatigue, and accelerates clinicalresponse.
Authors: Marco Horvath, Eva Novak, Andreas Rossi
Language: English
Subfield: Surgery
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W7203548529