Postoperative monitoring in Crohn’s disease: A multimodal machine learning model for predicting early complications and endoscopic recurrence

Early postoperative complications and endoscopic recurrence (ER) remain important determinants of postoperative prognosis after surgery for Crohn’s disease (CD). This study aimed to develop and validate multimodal prediction models for estimating these outcomes. This retrospective study included 159 patients with CD who underwent intestinal resection. Patients were randomly allocated to training and testing sets at a 7:3 ratio. Candidate predictors were selected in the training set using univariate logistic regression, LASSO regression, and multivariable logistic regression. Conventional logistic regression and seven machine-learning algorithms were subsequently developed and compared. Internal model performance was assessed using 10-times repeated 10-fold cross-validation and 1,000 bootstrap resamples for optimism correction, followed by evaluation in the held-out testing set. Model discrimination, calibration, potential clinical net benefit, and interpretability were assessed using receiver operating characteristic analysis, calibration analysis, decision curve analysis, and Shapley Additive Explanations (SHAP), respectively. Early postoperative complications and ER occurred in 32.08% and 27.67% of patients, respectively. Sarcopenia, sMaRIA, neutrophil-to-lymphocyte ratio, and the number of anastomoses were independently associated with postoperative complications, for which SVM achieved an out-of-fold AUC of 0.847, an optimism-corrected AUC of 0.855, and a testing AUC of 0.873. Perianal disease and granuloma formation were associated with increased ER risk, whereas longer disease duration showed an inverse association. For ER prediction, GBM achieved corresponding AUCs of 0.805, 0.828, and 0.795. Both models showed acceptable calibration and positive net-benefit profiles on decision-curve analysis. Multimodal models integrating clinical, inflammatory, imaging, body-composition, surgical, and pathological features showed promising internal predictive performance for postoperative risk stratification in CD. External multicenter validation is required before routine clinical implementation.