Prevalence, incidence and mortality of inflammatory bowel disease in the Netherlands: development and external validation of machine learning models.

Abstract

Background and aims: Large registries are promising tools to study the epidemiology of IBD. We aimed to develop and validate machine learning models to identify IBD cases in administrative data, aiming to determine the prevalence, incidence, and mortality of IBD in the Netherlands.

Methods: We developed machine learning models for administrative data to identify IBD cases and classify them on subtype and incidence year. Models were developed in a population-based cohort and externally validated in a hospital cohort. Models were evaluated on Brier score, area under the receiver operating characteristic curve (AUC), calibration, and accuracy. The best models were used to determine the epidemiology of IBD in the Netherlands between 2013 and 2020.

Results: For identifying IBD cases the random forest model was best (AUC: 0.97, 95% CI [0.96; 0.97]). The gradient boosted trees model for subtype was best (accuracy: 0.95, 95% CI [0.94; 0.95]) as was the random forest model for incidence year (0.88, 95% CI [0.86; 0.89]). The prevalence of IBD in the Netherlands was 577.6 (95% CI [566.7; 586.2]) per 100,000 on December 31, 2020, with varying prevalence across the Netherlands. Incidence of IBD was 20.1 (95% CI [18.0; 22.3]) per 100,000 in 2020 and stable over time. Mortality rates of IBD patients rose over time and were 11.6 (95% CI [10.5; 11.8]) per 1,000 in 2020 as compared to 9.5 in the general population.

Conclusion: IBD cases can be accurately identified using administrative data. The prevalence of IBD in the Netherlands is increasing slower than expected, suggesting a trend towards the epidemiological stage of Prevalence Equilibrium.