Deep-learning model for predicting 30-day postoperative mortality

Abstract Background Postoperative mortality occurs in 1–2% of patients undergoing major inpatient surgery. The currently available prediction tools using summaries of intraoperative data are limited by their inability to reflect shifting risk associated with intraoperative physiological perturbations. We sought to compare similar benchmarks to a deep-learning algorithm predicting postoperative 30-day mortality. Methods We constructed a multipath convolutional neural network model using patient characteristics, co-morbid conditions, preoperative laboratory values, and intraoperative numerical data from patients undergoing surgery with tracheal intubation at a single medical centre. Data for 60 min prior to a randomly selected time point were utilised. Model performance was compared with a deep neural network, a random forest, a support vector machine, and a logistic regression using predetermined summary statistics of intraoperative data. Results Of 95 907 patients, 941 (1%) died within 30 days. The multipath convolutional neural network predicted postoperative 30-day mortality with an area under the receiver operating characteristic curve of 0.867 (95% confidence interval [CI]: 0.835–0.899). This was higher than that for the deep neural network (0.825; 95% CI: 0.790–0.860), random forest (0.848; 95% CI: 0.815–0.882), support vector machine (0.836; 95% CI: 0.802–870), and logistic regression (0.837; 95% CI: 0.803–0.871). Conclusions A deep-learning time-series model improves prediction compared with models with simple summaries of intraoperative data. We have created a model that can be used in real time to detect dynamic changes in a patient's risk for postoperative mortality.