Explainable random forest-SHAP framework for Montreal phenotype-stratified prediction of one-year complications in
Kun Xia1,2,3,4, Ying Shi5, Xiaohan Huang1
1Department of Gastroenterology, The National Key Clinical Specialty, Zhongshan Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, Fujian, China.
Background:
Early identification of phenotype-stratified, short-term complication risk remains a critical unmet need in Crohn's disease (CD). Existing models focus on single outcomes and lack interpretability, limiting clinical use. We developed a clinically interpretable machine-learning framework to enable Montreal phenotype-stratified prediction of three one-year complication categories: bowel resection (Task 1), perianal complications (Task 2), and abdominal complications (Task 3).
Methods:
Data from 370 patients across two clinical centers were pooled for model development and evaluated using stratified 10-fold cross-validation. Within each fold, data preprocessing, LASSO feature selection, SMOTE-Tomek resampling, model optimization, calibration, and threshold determination were restricted to the training data. Seven machine-learning algorithms were compared in terms of discrimination, calibration, with decision-curve analysis used to explore potential net benefit. SHapley Additive exPlanations (SHAP) were used for model interpretation, and nomograms, heatmaps, and a web-based calculator were generated as exploratory visualizations of model-derived risk estimates rather than as clinical decision-support tools. Phenotype-stratified and non-linear associations between CDAI and complication risk were further analyzed within the Montreal framework.
Results:
Within one year after discharge, bowel resection, perianal complications, and abdominal complications occurred in 70 (18.9%), 142 (38.4%), and 84 (22.7%) patients, respectively. Random Forest showed the most consistent overall performance, achieving an AUROC of 0.86 (95% CI, 0.82-0.90) for bowel resection, 0.71 (95% CI, 0.64-0.78) for perianal complications, and 0.73 (95% CI, 0.69-0.78) for abdominal complications. SHAP identified distinct predictors across outcomes: disease behavior, nutritional support therapy, and disease duration for bowel resection; prior perianal complications, extraintestinal manifestations, and CDAI for perianal complications; and prior abdominal complications, nutritional support therapy, and Montreal L3 phenotype for abdominal complications. Receipt of NST was interpreted as a clinical marker of nutritional depletion and greater disease burden rather than as a causal risk factor. Montreal-based stratification revealed marked heterogeneity, with isolated small-bowel disease associated with higher risks of bowel resection and abdominal complications, and isolated colonic disease with the highest burden of perianal complications.
Conclusion:
The interpretable RF-SHAP framework showed potential for Montreal phenotype-stratified prediction of one-year complication risk in CD and inform future evaluation of risk stratification and phenotype-informed management.