Poster Session II – Poster of Distinction II - A213 PREDICTING SUSTAINED REMISSION AND MAXIMAL DISEASE SEVERITY IN PEDIATRIC CROHN’S DISEASE USING MACHINE LEARNING

TL;DR

This study uses machine learning to predict disease outcomes in children with Crohn’s disease, helping identify those who need early aggressive treatment.

Contribution

The study introduces integrated machine learning models combining clinical and microbiome data to predict pediatric Crohn’s disease severity and remission.

Findings

Integrated models outperformed single-modality models in predicting sustained remission, with logistic regression achieving a mean AUC of 0.763.

Microbiome models best predicted maximal disease severity, with Gaussian naïve Bayes reaching a mean AUC of 0.801.

Key predictive features included clinical variables and specific gut microbes like Haemophilus, Clostridium, and Coprococcus.

Abstract

Pediatric Crohn’s disease (CD) is a chronic inflammatory condition affecting the gastrointestinal tract. It displays more heterogeneous disease trajectories and treatment responses than adult-onset cases, posing significant management challenges. While patients following more severe trajectories may benefit from early aggressive treatments, no reliable objective method exists to identify which children will follow a severe trajectory at diagnosis. This prognostic gap leaves risk stratification dependent on subjective clinical judgment, potentially delaying interventions for high-risk patients. Early identification of severe trajectories could transform treatment decisions and improve outcomes through timely aggressive therapy.

This study aims to predict one-year sustained remission and maximal disease severity using machine learning models trained on baseline clinical and microbiome data in a nation-wide cohort of Canadian children with CD.

Using baseline clinical and microbiome data from the Canadian Children IBD Network inception cohort, we developed machine learning models to predict two first-year outcomes: 1) sustained remission vs non-sustained remission, defined as maintaining a post-remission Weighted Pediatric Crohn’s Disease Activity Index (wPCDAI) <12.5 without inflammatory episodes, and 2) maximal disease severity (remission/mild [post-diagnosis wPCDAI <40] vs moderate/severe [wPCDAI ≥40]). Nine algorithms were trained on three data modalities (clinical alone, microbiome alone, and integrated clinical-microbiome) using repeated nested K-fold cross-validation, with minimum redundancy maximal relevance feature selection, Bayesian hyperparameter optimization, and SHAP for model explainability.

For sustained remission prediction, integrated models outperformed microbiome- or clinical-only models, with integrated logistic regression achieving the highest mean AUC (0.763); key features included initial treatment at diagnosis, disease location, and wPCDAI at diagnosis, as well as taxa known to play a role in CD such as Haemophilus and Lachnospiraceae. For maximal disease severity prediction, microbiome models performed best, with Gaussian naïve Bayes reaching a mean AUC of 0.801 and highlighting microbes such as Clostridium and Veillonella as predictors of severe disease, while taxa such as Coprococcus and Romboutsia were associated with milder disease.

Our results demonstrate the potential of integrated machine learning approaches to support clinical decision-making in pediatric Crohn’s disease. By enabling early identification of high-risk patients at diagnosis, this work paves the way for personalized treatment strategies that could improve long-term outcomes in this vulnerable population.

CIHRBCCHRI, Government of British Columbia

Figures