Predicting mortality risk associated with serious treatable surgical complications at the University of Virginia health system

Abstract

This study focuses on predicting the risk of occurrence of serious but treatable complications and subsequent risk of mortality using a patient's preoperative conditions. Serious treatable complications include deep vein thrombosis/ pulmonary embolism, pneumonia, sepsis, and shock/cardiac arrest. These complications, if not identified and treated in time, can cause lengthened hospital stays, morbidity, and in some cases, mortality. We have modeled the risk of developing complications, and mortality due to complications, using a hierarchical prediction approach. In the first level of the hierarchy, extreme gradient boosted trees with cost sensitive weighting was used to model the risk of each complication and to identify the factors responsible for each type of complication. In the second level, similar statistical methods were used but on a smaller population set of patients, specifically those who developed one or more complications, to predict the risk of mortality. In our population of 32,202 patients, 963 developed one of the complications of interest, and of those with complications 174 died. Our predictions for sepsis, pneumonia, cardiac shock, and deep vein thrombosis/ pulmonary embolism, resulted in mean AUC values of 0.815, 0.935, 0.854, and 0.879 respectively. When making mortality predictions we achieved a mean AUC of 0.921. A propensity score analysis of patients that were predicted to be low risk but actually developed a complication was also performed. The framework proposed in this study provides hospitals with a way to more closely examine patient data regarding quality metrics by enabling them to identify patient born risks before surgical procedures are performed.