Harnessing machine learning to non-invasively evaluate mucosal healing in children with Crohn's disease.

objectiveMucosal healing (MH) is the therapeutic goal of Crohn’s disease (CD), closely related to the prognosis and quality of life of patients. The purpose of this study was to establish the machine learning models, to non-invasively evaluate the MH of children with CD, and use interpretable analysis to assist clinical decision-making.

methodsThis retrospective study included children diagnosed with CD and treated at least one year at Beijing Children's Hospital from January 2016 to June 2025. Recursive feature elimination was used to identify 15 clinical features. Five machine learning models, Multi-Layer Perceptron (MLP), Logistic Regression, Random Forest, Extreme Gradient Boosting and FT-Transformer were evaluated using the accuracy, precision, recall, F1 score and the area under the receiver operating characteristic curve. SHapley Additive exPlanations (SHAP) analysis was used to assess the contribution of features.

resultsA total of 75 patients were enrolled, of whom 38.67% achieved MH after 1 year of treatment. The MLP model achieved the highest evaluation accuracy, with the AUROC of 0.954. SHAP analysis revealed that high levels of fecal calprotectin after one year of treatment, high initial C-reactive protein levels, the presence of perianal disease, and the presence of intestinal strictures or perforations were the four most remarkable clinical features, all of which were negatively associated with achieving MH.

conclusionsthe four most significant clinical features negatively associated with achieving mh in children with cd were high one-year fecal calprotectin, high baseline c-reactive protein, the presence of perianal disease, and the presence of intestinal strictures or perforations. the machine learning models were established and validated, showing high evaluation accuracy, which was conducive to assisting clinical evaluation and decision-making for children with cd.