Clinical implementation of a machine learning system to detect deteriorating patients reduces time to response and intervention

Abstract

Introduction: Acute physiological deterioration is a major contributor to in-hospital morbidity and mortality. Early detection and intervention of deteriorating patients is key to improving patient outcomes. Prior research has demonstrated the effectiveness of Early Warning Systems and other algorithmic approaches in automatically identifying these patients from passively monitoring vital signs. Methods: In this work, we conduct a prospective pilot study of clinical deployment of the Mayo Clinic Bedside Patient Rescue (BPR) system using an escalating alerting logic enabled by machine learning. Among four units where the BPR system was deployed, time to response and time to intervention for deteriorating patients were significantly reduced relative to matched control units. Results: In pilot units, time to response decreased by 35.4% (from 63.2 minutes to 40.8 minutes) and time to intervention decreased by 48.5% (from 106.3 minutes to 55.9 minutes). No significant differences were observed in counterbalance metrics of mortality, ICU transfer rate, and Rapid Response Team activation rate. Furthermore, the automated alerting system was well-received by clinicians participating in the pilot study, as assessed by survey. Discussion: These results demonstrate a successful clinical deployment of a practice-changing machine learning alert system with demonstrable impact on improving patient care.