Personalized Risk Prediction for Prevention and Early Detection of Postoperative Failure to Rescue

Abstract In the Hospital of the Future hospitalization will be reserved almost exclusively for patients with severe acute illness, staff numbers will be reduced, and hospitals will be built around smart environments that facilitate consistent delivery of effective, equitable, and error-free care focused on patient-centered rather than provider- centered outcomes. This is particularly relevant to the surgical population. While ambulatory surgical centers are the fastest growing providers, more than 51 million inpatients procedures are performed annually in hospitals in the US and inpatient surgery centers are taking care of sicker and older patients. While intraoperative mortality is rare due to improvements in surgical techniques, anesthesia management, and intraoperative monitoring, global postoperative mortality remains the third leading cause of death among American People. Recent studies have shown that while the incidence of postoperative major complications after major surgery is similar between hospitals (~25%), the postoperative mortality following postoperative major complications from one hospital to the other can be up to 2.5-fold higher. This suggests that reducing variations in mortality following major surgery will require strategies to improve the ability of high-mortality hospitals to manage postoperative major complications and decrease failure-to-rescue. One of the solutions identified is to leverage Health Information Technologies. The goal of this proposal is to use machine learning approaches to develop, validate, and test real-time postoperative risk prediction tools based on multi-modal data sources using electronic health record data, high-fidelity physiological waveform features, and genomic data to identify patients who are at risk of developing postoperative major complications after surgery. Using extensive electronic health record derived annotation augmented with high-fidelity physiological waveform features and genomic data and applying state-of-the-art machine learning approaches, common patterns in subjects destined to develop postoperative major complications and those at very low risk of developing postoperative major complications after surgery will be characterized and quantified. These inputs will then be used in simulated real-time bedside management to iteratively design a prototype clinical decision support tool. This clinical decision support tool will be used at discharge from the post anesthesia care unit to identify surgical patients who will benefit from continuous remote monitoring and early warning system on the ward to prevent postoperative failure to rescue. The feasibility and acceptability of this approach will then be assessed in a small-scale prospective, longitudinal pilot evaluation in sequential 10-weeks, 13-weeks, 10-weeks phases at UCLA to help design a future, large-scale clinical trial.