Interpretable machine learning model for predicting postoperative complications in esophageal squamous cell carcinoma...

Postoperative complications remain common after neoadjuvant therapy and esophagectomy for esophageal squamous cell carcinoma (ESCC), and existing risk scores have limited bedside utility. We developed and validated an interpretable machine-learning model to predict morbidity.

We retrospectively included ESCC patients who underwent neoadjuvant therapy and curative esophagectomy from 2018 to 2022. Cases from 2018 to 2021 were the training cohort; cases from 2022 were the validation cohort. The endpoint was Clavien-Dindo grade ≥ II complications within 90 days. Multimodal preoperative and intraoperative variables trained several machine-learning algorithms. Performance was evaluated by cross-validation and independent validation. Interpretability used SHapley Additive exPlanations (SHAP), and top features yielded a simplified model.

In total, 239 patients were analyzed in two temporally distinct cohorts (training set, n = 174; validation set, n = 65). XGBoost achieved the best overall performance, with accuracy 0.82, AUC 0.86, and precision 0.78. SHAP analysis identified prognostic nutritional index, smoking status, carcinoembryonic antigen, age, and lymphocyte-to-monocyte ratio as the most influential predictors. A simplified five-variable model preserved predictive performance and enabled clinically actionable risk stratification: in the low-risk group, 9.1 % experienced complications, compared with 23.8 % in the medium-risk group and 68.2 % in the high-risk group. Locked cutpoints-PNI 44.07, CEA 2.98 ng/mL, age 64.00 years, and LMR 3.22-were applied without modification in the validation cohort and yielded clear stratification of incidence across threshold-defined groups.

An interpretable machine-learning model based on routinely available clinical variables accurately predicts postoperative complications in ESCC after neoadjuvant therapy. The simplified model enables clinically meaningful risk stratification and may support personalized perioperative management.