975: LOWERING ALARM BURDEN BY THE USE OF ARTIFICIAL INTELLIGENCE

Introduction: Continuous monitoring is an essential part of critical care and specifically for telemedicine based critical care coverage (Tele-ICU). Most monitoring systems provide warnings intended to alert when a patient’s condition deviates from a predetermined range. When clinicians experience high exposure to alarms (alarm burden), alarm desensitization occurs (Alarm fatigue), leading to missed alarms or delayed response. Alarm fatigue has been increasingly recognized as an important patient safety issue also leading to significant burnout. We have previously developed two novel AI based algorithms that predict respiratory and hemodynamic deterioration with high performance. The aim of the present work was to compare the alarm burden resulting from critical alerts generated by usual Tele-ICU care monitors and our AI based algorithms. Methods: Two separate prospectively designated cohorts (n=6,541 and 6,536 stays) were randomly selected out of 72,650 unique stays of patients admitted to one of 7 ICUs across the UMass Memorial Health Care system from 7-2006 to 9-2017. Performance of the AI based algorithm alerts (AIA) in predicting significant respiratory and hemodynamic deterioration was compared to that of usual Tele-ICU care alerts (UCA). Results: 1306 significant hemodynamic deterioration events and 548 significant respiratory deterioration events occurred across both cohorts. Our AI based algorithm predicted hemodynamic deterioration event with an AUC of 0.96 and 0.97 a median lead time of more than 3.5 hours. Respiratory deterioration prediction by the model had a median lead time of almost 4 hours with an AUC of 0.95 and 0.96 for both cohorts. AIA were more than 20-fold less frequent than UCA alerts, had significantly higher precision and accuracy as well a significantly longer lead time. As a result AIAs were more than 50 fold less frequent than UCAs, leading to a significantly lower alarm burden. Conclusions: Our artificial intelligence based algorithms were able to predict significant deterioration events hours before they occurred with significantly favorable performance as compared to usual care alerts. Besides the great clinical value that will result from earlier intervention the significant reduction in alert burden may assist in reducing alert fatigue and clinician burnout.