Serum proteins predict incident inflammatory bowel disease 14 years before diagnosis: a machine-learning and Cox modelling in the UK Biobank

Using an analysis pipeline that combines machine learning-based feature selection with classical statistical modelling, we aimed to predict incident inflammatory bowel disease (IBD) and to identify candidate proteins associated with disease development. We used UK Biobank data in an observational cohort study of 52,331 participants with proteomics data and without a history of IBD. The cohort was split into an 80/20 train/test dataset, and three separate extreme gradient boosting models were trained on the training dataset. Using Shapley values, we ranked the predictive importance of 2,912 proteins for incident IBD, Crohn's disease (CD), and ulcerative colitis (UC). The top 20 proteins per outcome served as features in multivariate Cox models predicting 350 incident cases over 14 years. All three proteomics Cox models showed moderate discrimination in the test dataset, with the UC model performing the best (C-index: 0.68), followed by IBD (C-index: 0.65), and CD (C-index: 0.60). All proteomics models outperformed clinical models. Smoking remained the only significant clinical predictor in the combined proteomics + clinical model. 48 unique proteins were identified across the three models, with 16 reaching statistical significance (p< 0.05). Machine learning-based feature selection with statistical survival modelling identifies plasma proteins that predict incident IBD and highlight biologically plausible, potentially targetable pathways in the preclinical phase of IBD.