Machine learning prediction of anastomotic leak after low anterior resection: Nationwide database analysis

Predicting anastomotic leak preoperatively in patients undergoing low anterior resection for rectal cancer remains a significant challenge. This study aims to develop and validate a predictive model for anastomotic leaks using a machine learning algorithm. Data were collected on patients who underwent low anterior resection from 2015 to 2021 from the National Clinical Database in Japan. The patients were divided into 2 cohorts: a derivation cohort and a validation cohort. The derivation cohort included patients who underwent surgery between January 2015 and December 2019, while the validation cohort included those from January 2020 to December 2021. Three models were developed: logistic regression, logistic regression with least absolute shrinkage and selection operator (Lasso regression), and eXtreme gradient boosting model. We calculated the area under the receiver operating characteristic curve (AUROC) and compared it with the logistic regression model using the DeLong test. A total of 119,818 eligible patients were identified. The incidence of anastomotic leaks was 9.6% in the derivation cohort and 8.4% in the validation cohort, respectively. The predictive ability for the validation cohort using logistic regression (AUROC 0.6324, 95% confidence interval [CI] 0.6220-0.6427, reference) was similar to that of Lasso regression (AUROC 0.6333, 95% CI 0.6229-0.6436, P = .13) and eXtreme gradient boosting (AUROC 0.6333, 95% CI 0.6230-0.6437, P = .41). Machine learning prediction model for anastomotic leak using preoperative information routinely inputted in the National Clinical Database, showed suboptimal prediction ability. It wound be possible to share the fact with patients that preoperative prediction of anastomotic leak is difficult.