A machine learning model for predicting early adverse outcomes after surgery in patients with early-onset rectal cancer: a two-center retrospective cohort study

Abstract

Patients often present at advanced stages of early-onset rectal cancer (EORC), which is becoming more common. Within a year following rectal resection, some patients experience early adverse outcomes (EAO), such as severe complications, recurrence, or death; however, there is currently no validated postoperative risk-stratification tool for this population. Current prognostic models concentrate on individual endpoints rather than the composite adversity spectrum that is pertinent to younger patients because they were mostly established in older cohorts. This work sought to create and externally validate a model for detecting EORC patients at greater risk of EAO using a stacking ensemble machine learning approach with SHAP interpretability in order to guide individualized postoperative surveillance and adjuvant therapy decisions. In this two-center retrospective cohort study, 648 EORC patients from Center A (2020–2025) constituted the development cohort and 102 patients from Center B (2024–2025) the external validation cohort. Eleven individual machine learning models and a two-layer stacking ensemble (combining three base learners with logistic regression meta-learner) were trained. Model performance was assessed by area under the receiver operating characteristic curve (AUC), calibration, and decision curve analysis. Boruta feature selection identified eleven important predictors, with pathological T stage (pT), lymphovascular invasion (LVI), and pathological TNM stage (pTNM) contributing most to model predictions. The stacking ensemble outperformed all 11 individual models, achieving AUCs of 0.91 (95% CI 0.88–0.93) in the training set, 0.89 (95% CI 0.85–0.94) in the internal validation set, and 0.83 (95% CI 0.76–0.91) in the external validation cohort, with a Brier score of 0.1104 and sustained net benefit on decision curve analysis. An interactive web calculator was deployed for individualized risk assessment. An interpretable stacking ensemble model integrating SHAP analysis and a web-based calculator showed promising predictive performance for EAO following EORC surgery. Designed for postoperative use once pathology data are available, the model provides a comprehensive risk estimate to guide surveillance intensity and interdisciplinary adjuvant therapy decision-making.