Background: Previously we identified that patients with initial diagnosis of Ulcerative Colitis (UC) but subsequently had a diagnosis change to Crohn’s Disease (CD) (UC2CD) are a unique IBD subgroup with distinct serological, genetic and clinical features, including non-response to anti-TNF treatments and high risk of Extraintestinal Manifestations (EIM). We developed a UC2CD prediction model using a Machine Learning (ML) algorithm based on serological, clinical and lab test values. This initial finding was made based on observational study in a real-life clinical cohort at CSMC. Now four years later with continuous follow up we have validated the performance of the UC2CD ML model in patients who at that stage had a diagnosis of UC and have since had their diagnosis switched to CD. For clinical utility, we developed a clinical score for UC2CD and examined its performance in the MIRIAD cohort.
Methods: EHR data including history of diagnosis change, lab test values and clinical phenotypes were extracted from the MIRIAD biobank. IBD patients with initial diagnosis of UC at the 2019 data freeze but developed into CD thereafter were defined as new UC2CD cases. Frequency of new UC2CD cases in the top and bottom 25% of the UC2CD ML prediction were compared using logistic regression. For clinical utility, we constructed the UC2CD clinical score based on a sum of five major components in the ML prediction model that included IBD serologies (anti-CBir-1, OmpC, IgA-ASCA), serum globulin, and erythrocyte sedimentation rate (ESR), as listed in table 1. Performance of the UC2CD clinical score in distinguishing UC2CD from ‘established’ UC subjects was evaluated in all subjects with initial UC diagnosis in the MIRIAD Biobank.
Results: In 2,629 IBD patients with an initial diagnosis of UC in MIRIAD, 332 were UC2CD cases in the 2019 data freeze and 139 more UC2CD cases were identified during follow up after 2019. 61(10.26%) from the top 25% of the ML prediction and 28 (3.43%) from the bottom 25% of the ML prediction were UC2CD, corresponding to an OR of 3.34 (P = 1.85E-7). 491 patients with initial diagnosis of UC in MIRIAD cohort had all five components (anti-CBir, IgA-ASCA, OmpC, globulin and ESR) of the UC2CD clinical score. Figure 1 shows the proportion of UC2CD subjects according to different UC2CD clinical scores. For initial UC patients with UC2CD clinical score ≥ 3, 46.3% had developed CD during follow up.
Conclusion: With an additional 4 years of follow up of UC patient cohort in the MIRIAD Biobank, we confirmed the validity of our ML prediction model. We also developed a UC2CD clinical score based on 5 components for potential clinical implementation. A simple scoring system may identify IBD subjects at high risk to switch diagnoses to this refractory phenotype who may require more intensive monitoring.

Carol Landers

Appearances