Different models for prediction of radical cystectomy postoperative complications and care pathways

Abstract

Background: Radical cystectomy for bladder cancer has one of the highest rates of morbidity among urologic surgery, but the ability to predict postoperative complications remains poor. Our study objective was to create machine learning models to predict complications and factors leading to extended length of hospital stay and discharge to a higher level of care after radical cystectomy. Methods: Using the American College of Surgeons National Surgical Quality Improvement Program, peri-operative adverse outcome variables for patients undergoing elective radical cystectomy for bladder cancer from 2005 to 2016 were extracted. Variables assessed include occurrence of minor, infectious, serious, or any adverse events, extended length of hospital stay, and discharge to higher-level care. To develop predictive models of radical cystectomy complications, we fit generalized additive model (GAM), least absolute shrinkage and selection operator (LASSO) logistic, neural network, and random forest models to training data using various candidate predictor variables. Each model was evaluated on the test data using receiver operating characteristic curves. Results: A total of 7557 patients were identified who met the inclusion criteria, and 2221 complications occurred. LASSO logistic models demonstrated the highest area under curve for predicting any complications (0.63), discharge to a higher level of care (0.75), extended length of stay (0.68), and infectious (0.62) adverse events. This was comparable with random forest in predicting minor (0.60) and serious (0.63) adverse events. Conclusions: Our models perform modestly in predicting radical cystectomy complications, highlighting both the complex cystectomy process and the limitations of large healthcare datasets. Identifying the most important variable leading to each type of adverse event may allow for further strategies to model cystectomy complications and target optimization of modifiable variables pre-operative to reduce postoperative adverse events.