Development and validation of a Bayesian network-based surgical risk-prediction tool for patients with small-bowel stricturing Crohn’s disease

Background: Patients with small-bowel stricturing Crohn’s disease (sbsCD) usually have a higher risk of intestinal surgical resection. We aimed to develop a machine-learning model for predicting the 1-year surgery risk in these patients. Methods: This study included 520 retrospectively enrolled patients with sbsCD (training cohort, n = 416; testing cohort, n = 104) from January 2018 to May 2021 and 126 prospectively enrolled patients in the validation cohort from July 2021 to December 2023 across four centers for inflammatory bowel disease in China. Clinical and radiological features were assessed by using logistic regression analyses to identify independent surgery risk factors. Six predictive machine-learning models for 1-year surgical risk were developed and the model performance was comprehensively evaluated by constructing receiver-operating characteristic (ROC) curves and comparing area under the curve (AUC) values. Four simplified Bayesian network (BN)-based risk matrices were constructed for clinical practice. Results: There were 158 (24.4%) Crohn’s disease (CD)-related surgeries during the 1-year follow-up. Eight selected predictors of surgery included penetrating lesions, nonuse of biologics, nonuse of corticosteroids, a CD obstructive score of ≥3, endoscopic strictures, anemia, radiologic luminal narrowing, and prestenotic dilation. Among the six models evaluated, the Tree-Augmented Naïve Bayes (TAN) model demonstrated optimal performance, with a mean AUC of 0.878. A further prospective validation cohort verified the efficacy of the model, with 87.5% specificity, 76.7% sensitivity, and 84.9% accuracy for predicting 1-year surgery. Four simplified BN-based risk matrices were constructed for practical use. An online prediction tool is available at http://prebn.site/. Conclusion: We developed and validated a TAN-based BN model incorporating clinical and radiological features to accurately predict the 1-year surgical risk for clinical application in patients with sbsCD, thereby providing a promising tool for decision-making.