Development and validation of an interpretable machine learning model for predicting postoperative complications
Z Jingtao1,2, Y Jun3, C Shaoqin4
1Department of Gastrointestinal Surgery 2 Section, the First Affiliated Hospital, Fujian Medical University, Fuzhou, 350005, China.
Objective:
This study aimed to develop and externally validate an interpretable machine learning (ML) model for predicting postoperative complications after transanal total mesorectal excision (taTME).
Methods:
We conducted a multicenter case-control study of 245 consecutive patients undergoing taTME at three centers in China. Patients from one center formed the development cohort (n = 185); two independent centers provided an external validation cohort (n = 60). The primary endpoint was any 30-day complication (Clavien-Dindo). Candidate variables were screened with least absolute shrinkage and selection operator (LASSO) and entered into multivariable logistic regression to define features. In total, 11 ML algorithms were trained with tenfold cross-validation; the final model was selected by area under the receiver operating characteristic (ROC) curve (AUC). Discrimination, calibration, decision-curve analysis (DCA), and clinical-impact curves were assessed. Model interpretability was examined with Shapley additive explanations (SHAP).
Results:
Overall, 27.8% of patients experienced postoperative complications. Six predictors-postoperative C-reactive protein (CRP), preoperative total protein (TP), smoking history, age group (> 80 years), neoadjuvant therapy, and ASA score-were retained. The AdaBoost classifier performed best (development AUC 0.822; 95% CI, 0.751-0.892). External validation demonstrated higher discrimination (AUC 0.872; 95% CI, 0.771-0.972), with accuracy 83.3%, sensitivity 56.25%, and specificity 94.7%. Calibration and DCA supported clinical utility, and clinical-impact curves indicated favorable threshold behavior. The SHAP analysis identified postoperative CRP as the most influential predictor and provided transparent, individualized risk explanations.
Conclusions:
We developed and externally validated a robust and interpretable ML model to predict complications following taTME. By leveraging routinely available clinical variables and providing transparent predictive explanations, this model can aid clinicians in identifying high-risk patients for targeted interventions, thereby potentially improving perioperative outcomes.
