Dynamic prediction of advanced colorectal neoplasia in inflammatory bowel disease

Background & Aims Colonoscopic surveillance is recommended in patients with colonic inflammatory bowel disease (IBD) given their increased risk of colorectal cancer (CRC). We aimed to develop and validate a dynamic prediction model for the occurrence of advanced colorectal neoplasia (aCRN, including high-grade dysplasia and CRC) in IBD. Methods We pooled data from six existing cohort studies from Canada, the Netherlands, the UK, and the USA. Patients with IBD and an indication for CRC surveillance were included if they underwent at least one follow-up procedure. Exclusion criteria included prior aCRN, prior colectomy, or an unclear indication for surveillance. Predictor variables were selected based on literature. A dynamic prediction model was developed using a landmarking approach based on Cox proportional hazard modelling. Model performance was assessed with Harrell’s concordance-statistic (discrimination) and by calibration curves. Generalizability across surveillance cohorts was evaluated by internal-external cross-validation. Results The surveillance cohorts comprised 3,731 patients, enrolled and followed-up in the time period 1973-2021, with a median follow-up of 5.7 years (26,336 patient-years of follow-up); 146 individuals were diagnosed with aCRN. The model contained eight predictors, with a cross-validation median c-statistic of 0.74 and 0.75 for a 5- and 10-year prediction window, respectively. Calibration plots showed good calibration. Internal-external cross-validation results showed medium discrimination and reasonable to good calibration. Conclusion The new prediction model showed good discrimination and calibration, however, generalizability results varied. Future research should focus on formal external validation and relate predicted aCRN risks to surveillance intervals prior to clinical application.