An Explainable Multimodal Model for Assessing Mucosal Healing in Small Bowel Crohn's Disease: A Multicenter Study with Prospective Validation.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, The First Affiliated Hospital of Anhui Medical University, Hefei, Anhui Province, 230031, China.
- Department of Radiology, The First People's Hospital of Wuhu, Wuhu, Anhui Province, China.
- Department of Radiology, Division of Life Sciences and Medicine, The First Affiliated Hospital of USTC, University of Science and Technology of China, Hefei, Anhui, 230036, China.
- Department of Radiology, Division of Life Sciences and Medicine, The First Affiliated Hospital of USTC, University of Science and Technology of China, Hefei, Anhui, 230001, China. [email protected].
- Department of Radiology, The First Affiliated Hospital of Anhui Medical University, Hefei, Anhui Province, 230031, China. [email protected].
Abstract
Accurate assessment of mucosal healing (MH) in small bowel Crohn's disease (SBCD) remains challenging because double-balloon enteroscopy (DBE), the reference standard, is invasive and unsuitable for repeated monitoring. This study was aimed at developing and validating an explainable multimodal model integrating CT enterography (CTE)-based radiomics with clinical data for post-biologic MH assessment in SBCD. This multicenter study included 206 retrospective patients (training: n = 114; internal validation: n = 50; external test: n = 42) and 20 prospective patients. Clinical predictors were identified by logistic regression, with C-reactive protein (CRP) as the only independent predictor. CTE radiomics features were processed by Pearson correlation and SelectKBest, then compared across six feature-selection methods and nine classifiers to determine the optimal radiomics pipeline. Clinical models used CRP alone, while multimodal models combined CRP with selected radiomics features. Performance was evaluated using AUC, calibration, DCA, and DeLong testing. The SHapley Additive exPlanations (SHAP) provided interpretability. MH was achieved in 52.6%, 54.0%, and 21.4% of patients across cohorts. The LASSO-XGBoost radiomics model achieved AUCs of 0.854, 0.821, and 0.808 across retrospective cohorts and 0.824 in the prospective cohort. The XGBoost multimodal model achieved AUCs of 0.860 and 0.842 in internal and external validation, significantly outperforming the clinical model (P < 0.05). SHAP identified CRP and two radiomics features as top predictors. The explainable multimodal model integrating CTE radiomics and CRP enables accurate, non-invasive MH assessment with prospective-validated generalizability, supporting its potential as a clinical decision-support tool.