Prediction of post-surgical adverse outcomes using procedural data

Modern surgical procedures have saved lives; however, the occurrence of post-surgical complications is still a concern. This study investigates the relationships between patients' surgical data and the post-surgical adverse outcomes of death, cardiac injury, renal injury, and respiratory failure. We analyze these relationships through the statistical modelling of predictor variables from patients' preoperative data and intraoperative physiological time series data, as well as the intraoperative time series of procedures performed by anesthesiologists and surgeons. In order to include this procedural data we create novel predictors to represent each patient's time series in models. Three different statistical models are described: Random Forest (RF), Generalized Linear Models (GLM), and L1 Regularized Logistic Regression (L1). We evaluate the model results using Receiver Operating Characteristic (ROC) curves and their corresponding area under the curve (AUC) values. The GLM model performs the best for death with an AUC value of .887, and the RF model performs best for cardiac injury, renal injury and respiratory failure, with AUC values of .857, .873 and .831, respectively. For all adverse outcomes, the L1 models minimized false negatives at a threshold of .01. This paper describes the results of these models and identifies the strongest predictors of each post-surgical adverse outcome. Additionally, we create a Mixed Effects (ME) model to isolate interactions between patient types and other significant predictors. This model suggests significant interaction effects between the placement of an arterial line and patient types when predicting renal injury and respiratory failure.