LAUSR

Purpose Hypotension is a risk factor for adverse perioperative outcomes. Preoperative transthoracic echocardiography has been extended for preoperative risk assessment before noncardiac surgery. This study aimed to develop a machine learning model to predict postinduction hypotension risk using preoperative echocardiographic data and compared it with conventional statistic models. We also aimed to identify preoperative echocardiographic factors that cause postinduction hypotension. Methods In this retrospective observational study, we extracted data from electronic health records of patients aged >18 years who underwent general anesthesia at a single tertiary care center between April 2014 and September 2019. Multiple supervised machine learning classification techniques were used, with postinduction hypotension (mean arterial pressure <55 mmHg from intubation to the start of the procedure) as the primary outcome and 95 transthoracic echocardiography measurements as factors influencing the primary outcome. Based on the mean cross-validation performance, we used 10-fold cross-validation with the training set (70%) to select the optimal hyperparameters and architecture, assessed ten times using a separate test set (30%). Results Of 1,956 patients, 670 (34%) had postinduction hypotension. The area under the receiver operating characteristic curve using the deep neural network was 0.72 (95% confidence interval (CI) = 0.67–0.76), gradient boosting machine was 0.54 (95% CI = 0.51–0.59), linear discriminant analysis was 0.56 (95% CI = 0.51–0.61), and logistic regression was 0.56 (95% CI = 0.51–0.61). Variables of high importance included the ascending aorta diameter, transmitral flow A wave, heart rate, pulmonary venous flow S wave, tricuspid regurgitation pressure gradient, inferior vena cava expiratory diameter, fractional shortening, left ventricular mass index, and end-systolic volume. Conclusion We have created developing models that can predict postinduction hypotension using preoperative echocardiographic data, thereby demonstrating the feasibility of using machine learning models of preoperative echocardiographic data for produce higher accuracy than the conventional model.