Deep recurrent neural networks for predicting intraoperative and postoperative outcomes and trends

Abstract

Intraoperative vital signals convey a wealth of complex temporal information that can provide significant insights into a patient's physiological status during the surgery, as well as outcomes after the surgery. Our study involves the use of a deep recurrent neural network architecture to predict patient's outcomes after the surgery, as well as to predict the immediate changes in the intraoperative signals during the surgery. More specifically, we will use a Long Short-Term Memory (LSTM) model which is a gated deep recurrent neural network architecture. We have performed two experiments on a large intraoperative dataset of 12,036 surgeries containing information on 7 intraoperative signals including body temperature, respiratory rate, heart rate, diastolic blood pressure, systolic blood pressure, fraction of inspired O 2 and end-tidal CO 2 . We first evaluated the capability of LSTM in predicting the immediate changes in intraoperative signals, and then we evaluated its performance on predicting each patient's length of stay outcome. Our experiments show the effectiveness of LSTM with promising results on both tasks compared to the traditional models.