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

What To Know

A large retrospective study used routine electronic health record data (labs, BMI, demographics, comorbidities) from Maccabi Health Services to train Explainable Boosting Machine models that predict future Crohn’s disease and ulcerative colitis risk months to years before diagnosis.

Models performed better for Crohn’s (AUROC ~0.71–0.75 up to 2 years before diagnosis) than for UC (AUROC ~0.62–0.64) and identified a high-risk decile with roughly 2–3× higher future IBD risk.

The researchers tested two approaches: prediction anchored to time before diagnosis (0.5–2 years) and age-based screening (estimating risk at set ages with multi-year windows). Age-based screening at 25 years showed the strongest performance (AUROC ~0.70–0.72 for 1–11 year windows).

The models are described as interpretable (EBM), which may help clinicians understand contributing features. This work suggests that routinely collected clinical data can support scalable, population-level risk stratification to flag people for earlier evaluation, but it is a retrospective cohort study rather than a prospective trial or clinical deployment.