Machine learning model for predicting a high comprehensive complication index following rectal cancer surgery.

Abstract

Postoperative complications following rectal cancer surgery can significantly affect patient's health and prognosis. It has been reported that the comprehensive complication index (CCI) is a more sensitive assessment tool for severe complications than the Clavien-Dindo classification (CDC) system. This study aims to construct a predictive model for high postoperative CCI using machine learning methods to guide clinical practice. A total of 1029 patients with mid and low rectal cancer who underwent rectal resection were included. Preoperative, intraoperative clinicopathological characteristics and pelvic measurement data were collected. Five predictive models were constructed using machine learning methods, including Random Forest (RF), LightGBM, Logistic Regression (LR), Naive Bayes Model (NBM) and XGBoost, and their performances were compared. Finally, the Shapley Additive exPlanations (SHAP) was used to visually interpret the predictive variables of the best model. Six predictive variables, including surgical time, interspinous distance, pelvic depth, age, diabetes, and tumor distance, were included in the model construction. Among the five models, LightGBM was the optimal model, with an AUC of 0.746 in the training set, 0.760 in the testing set and 0.709 in the validation set. It had the best DCA curve for most thresholds, indicating excellent performance in predicting high CCI. This study developed a predictive model for assessing the risk of high CCI following anterior resection for rectal cancer. It could provide personalized treatment strategies for patients at high risk of severe complications, improves patient prognosis, and promotes its use through an online web tool ( https://mypredict.shinyapps.io/CDC_CCI/ ).