Interpretable machine learning models for early prediction of Crohn’s disease and ulcerative colitis in a nationwide health system

BACKGROUND & AIMS: Delayed diagnosis of inflammatory bowel disease (IBD) is associated with worse outcomes. Existing prediction studies often rely on small cohorts or specialized biomarkers that limit scalability. We evaluated whether machine learning models based on routinely collected clinical data can predict future Crohn's disease (CD) and ulcerative colitis (UC) years before diagnosis in a large population-based cohort. METHODS: We conducted a retrospective study using electronic medical records from Maccabi Health Services (Israel), including 81 960 individuals (21 695 with IBD). Longitudinal laboratory tests, body mass index, demographics, and comorbidities were used to train Explainable Boosting Machine (EBM) models. Two frameworks were examined: (1) diagnosis-anchored prediction 0.5-2 years before diagnosis, and (2) age-based screening, estimating risk at predefined ages with future prediction windows. Performance was assessed using area under the receiver operating characteristic curve (AUROC) and relative risk (RR) in high-risk subgroups. RESULTS: In diagnosis-anchored models, AUROC for CD ranged from 0.71 to 0.75 up to 2 years before diagnosis and from 0.62 to 0.64 for UC. In age-based screening, the best performance was observed at age 25 with a 1-11-year prediction window (AUROC 0.72 for CD; 0.70 for UC). Individuals in the top decile of predicted risk had an approximately 3-fold higher risk of future IBD, with RR remaining >2.5 for CD and >2.0 for UC for shorter windows. Kaplan-Meier analyses showed clear separation between high- and low-risk groups (P < .001). CONCLUSIONS: EBM models using routine clinical data can identify individuals at increased future IBD risk, supporting scalable, interpretable population-level risk stratification and earlier evaluation.