Machine-learning models using preoperative and intraoperative data predicted unstable intraoperative bradycardia with AUCs of 0.81, 0.87, and 0.89 at three different time points.
Observational (n=62,182)
No
Can machine-learning models utilizing preoperative and intraoperative data predict the occurrence of intraoperative bradycardia associated with hypotension during noncardiac procedures?
Machine-learning models leveraging preoperative and real-time intraoperative data can accurately predict the occurrence of unstable bradycardia during noncardiac procedures.
Effect estimate: AUC 0.81, 0.87, and 0.89
BACKGROUND: Predictive analytics systems may improve perioperative care by enhancing preparation for, recognition of, and response to high-risk clinical events. Bradycardia is a fairly common and unpredictable clinical event with many causes; it may be benign or become associated with hypotension requiring aggressive treatment. Our aim was to build models to predict the occurrence of clinically significant intraoperative bradycardia at 3 time points during an operative course by utilizing available preoperative electronic medical record and intraoperative anesthesia information management system data. METHODS: The analyzed data include 62,182 scheduled noncardiac procedures performed at the University of Washington Medical Center between 2012 and 2017. The clinical event was defined as severe bradycardia (heart rate <50 beats per minute) followed by hypotension (mean arterial pressure <55 mm Hg) within a 10-minute window. We developed models to predict the presence of at least 1 event following 3 time points: induction of anesthesia (TP1), start of the procedure (TP2), and 30 minutes after the start of the procedure (TP3). Predictor variables were based on data available before each time point and included preoperative patient and procedure data (TP1), followed by intraoperative minute-to-minute patient monitor, ventilator, intravenous fluid, infusion, and bolus medication data (TP2 and TP3). Machine-learning and logistic regression models were developed, and their predictive abilities were evaluated using the area under the ROC curve (AUC). The contribution of the input variables to the models were evaluated. RESULTS: The number of events was 3498 (5.6%) after TP1, 2404 (3.9%) after TP2, and 1066 (1.7%) after TP3. Heart rate was the strongest predictor for events after TP1. Occurrence of a previous event, mean heart rate, and mean pulse rates before TP2 were the strongest predictor for events after TP2. Occurrence of a previous event, mean heart rate, mean pulse rates before TP2 (and their interaction), and 15-minute slopes in heart rate and blood pressure before TP2 were the strongest predictors for events after TP3. The best performing machine-learning models including all cases produced an AUC of 0.81 (TP1), 0.87 (TP2), and 0.89 (TP3) with positive predictive values of 0.30, 0.29, and 0.15 at 95% specificity, respectively. CONCLUSIONS: We developed models to predict unstable bradycardia leveraging preoperative and real-time intraoperative data. Our study demonstrates how predictive models may be utilized to predict clinical events across multiple time intervals, with a future goal of developing real-time, intraoperative, decision support.
Solomon et al. (Tue,) conducted a observational in Intraoperative bradycardia associated with hypotension (n=62,182). Machine-learning predictive models was evaluated on Severe bradycardia (heart rate <50 beats per minute) followed by hypotension (mean arterial pressure <55 mm Hg) within a 10-minute window (AUC 0.81, 0.87, and 0.89). Machine-learning models using preoperative and intraoperative data predicted unstable intraoperative bradycardia with AUCs of 0.81, 0.87, and 0.89 at three different time points.