Multimodal prediction of infliximab initiation within 6 months after initiation of conventional therapy in pediatric...

Early prediction of treatment course in pediatric Crohn's disease (CD) remains limited despite the importance of timely biologic therapy. We developed and externally validated a multimodal model integrating colonoscopy images and clinical data to predict infliximab initiation within 6 months after conventional therapy.

We collected data from children newly diagnosed with CD at three hospitals: Seoul National University Bundang Hospital for internal validation, Kyung Hee University Medical Center, and Soonchunhyang University Bucheon Hospital for external validation. Each patient contributed 7 endoscopic images and 20 clinical variables. An attention-based multiple-instance learning framework aggregated endoscopic images into a patient-level representation and integrated them with clinical features. The imaging encoder used a ResNet-50 backbone pretrained with momentum contrastive learning on GastroNet-5M, and clinical features were encoded using a multilayer perceptron. Performance was assessed using the area under the receiver operating characteristic curve (AUROC), balanced accuracy, precision, recall, and F1 score. Explainable AI was evaluated using attention maps (image data) and SHAP values (tabular data).

Seventy-seven children were included for model development and internal validation, with external validation performed in two independent cohorts (n = 10 and n = 20). The multimodal model achieved an AUROC of 0.839 and balanced accuracy of 0.732 in the discovery cohort. External-validation AUROCs were 0.857 and 0.688. SHapley Additive exPlanation analysis identified immunoglobulin G, body mass index, prealbumin, hematocrit, and age as key predictors. Attention maps highlighted inflamed mucosa and ulcerative lesions.

A multimodal model integrating colonoscopy images with clinical data showed promising performance for predicting infliximab initiation within 6 months after conventional therapy in pediatric CD. Multimodal information may support early risk stratification and identify patients who could benefit from earlier biologic intervention.