Development and Temporal Validation of a Multimodal Machine Learning Model for Classification of Recent Flare Activity in Inflammatory Bowel Disease: A Retrospective Cohort Study

INTRODUCTION: The aim of this study was to develop and temporally validate a multimodal machine learning model for classifying clinically relevant flare-related activity during the 90 days preceding the index visit in patients with inflammatory bowel disease (IBD).
METHODS: This retrospective cohort study included 345 adults with IBD in clinical remission or with mild disease activity who underwent routine ileocolonoscopy at a tertiary referral center between December 2025 and March 2026. Demographic, clinical, laboratory, endoscopic, histopathological, and patient-reported data were extracted from electronic medical records for the 90-day pre-index period. The primary outcome was a composite of treatment intensification due to symptomatic worsening, IBD-related hospitalization, or IBD-related surgery during that period. Least absolute shrink-age and selection operator regression was used for feature selection in a development cohort (n=242). Three algorithms – regularized logistic regression, random forest, and eXtreme Gradient Boosting (XGBoost) – were evaluated in a temporally distinct validation cohort (n=103) using area under the receiver operating characteristic curve (AUROC), calibration metrics, and decision curve analysis.
RESULTS: Flare-related activity was identified in 76 of 345 patients (22.0%). In the validation cohort, the multimodal XGBoost model showed the best discriminative performance, with an AUROC of 0.84 (95% confidence interval [CI], 0.77–0.91), sensitivity of 78.0%, specificity of 73.8%, positive predictive value of 55.3%, and negative predictive value of 89.1%. The multimodal model outperformed unimodal models (clinical-only AUROC 0.66, laboratory-only 0.71, endoscopic-only 0.74, and histology-only 0.69) and a baseline clinical model (AUROC 0.73). Calibration was good (Brier score 0.15), and decision curve analysis demonstrated net benefit across clinically relevant thresholds.
DISCUSSION AND CONCLUSION: A multimodal model integrating routinely collected data demonstrated good performance in classifying recent flare-related activity in IBD and outperformed single-modality approaches in a temporally distinct validation cohort. External validation is required before clinical implementation.

Keywords: Biomarkers, decision support systems, endoscopy, inflammatory bowel diseases, machine learning.