Machine learning for predicting Crohn's disease from routine blood tests years before diagnosis: results from the epi-IIRN cohort.

Abstract

Objectives: In this nationwide study, we used the epi-Israeli Inflammatory Bowel Disease (IBD) Research Nucleus (IIRN) validated cohort to explore the utility of routine blood tests as markers predicting IBD occurrence years before diagnosis.

Methods: We included all health maintenance organization (HMO)-insured IBD patients in Israel diagnosed during 2005-2020 to identify discriminative results of blood tests performed up to 15 years before diagnosis. Each patient was individually matched to two non-IBD controls. Means were compared using Welch's t-test with false discovery rate correction to account for multiple comparisons. A machine-learning model was developed using the most significant blood tests to predict future Crohn's disease (CD).

Results: Pre-diagnosis results from 84 blood tests were collected for 8630 CD and 6791 ulcerative colitis (UC) patients, including 1162 children with CD and 580 with UC, and their matched controls. Among adults with CD, 29 tests differed consistently from controls earlier than 1 year pre-diagnosis; three showed consistent differences more than 10 years pre-diagnosis. For children, 17 tests differed consistently more than 1 year pre-diagnosis. No tests significantly differed between UC cases and controls. The machine-learning model predicted CD in adults with an area under the curve (AUC) of 0.70 1 year pre-diagnosis and 0.61 7 years pre-diagnosis.

Conclusion: We were able to detect changes in routinely collected blood tests long before CD diagnosis and to predict future CD using a machine-learning model, which may be used for developing screening and prediction models for prevention strategies.