Learning longitudinal patterns and identifying subtypes of pediatric Crohn disease treated with infliximab via trajectory cluster analysis of electronic health records

Background The current classification of pediatric CD is mainly based on cross-sectional data. The objective of this study is to identify subgroups of pediatric CD through trajectory cluster analysis of disease activity using data from electronic health records.

Methods We conducted a retrospective study of pediatric CD patients who had been treated with infliximab. The evolution of disease over time was described using trajectory analysis of longitudinal data of C-Reactive Protein (CRP). Patterns of disease evolution were extracted through functional principal components analysis and subgroups were identified based on those patterns using the Gaussian mixture model. We compared patient characteristics, a biomarker for disease activity, received treatments, and long-term surgical outcomes across subgroups.

Results We identified four subgroups of pediatric CD patients with differential relapse-and-remission risk profiles. They had significantly different disease phenotype (p < 0.001), CRP (p < 0.001) and calprotectin (p = 0.037) at diagnosis, with increasing percentage of inflammatory phenotype and declining CRP and fecal calprotectin levels from Subgroup 1 through 4. The risk of colorectal surgery within 10 years after diagnosis was significantly different between groups (p < 0.001). We did not find statistical significance in gender or age at diagnosis across subgroups, but the BMI z-score was slightly smaller in subgroup 1 (p =0.055).

Conclusions Readily available longitudinal data from electronic health records can be leveraged to provide a deeper characterization of pediatric Crohn disease. The identified subgroups captured novel forms of variation in pediatric Crohn disease that were not explained by baseline measurements and treatment information.

Summary The current classification of pediatric Crohn disease mainly relies on cross-sectional data, e.g., the Paris classification. However, the phenotypic classification may evolve over time after diagnosis. Our study utilized longitudinal measures from the electronic health records and stratified pediatric Crohn disease patients with differential relapse-and-remission risk profiles based on patterns of disease evolution. We found trajectories of well-maintained low disease activity were associated with less severe disease at baseline, early initiation of infliximab treatment, and lower risk of surgery within 10 years of diagnosis, but the difference was not fully explained by phenotype at diagnosis.