Using machine learning to predict clinical remission with exclusive enteral nutrition in pediatric Crohn disease

Exclusive enteral nutrition (EEN) is a first-line treatment for induction of remission in luminal pediatric Crohn disease (pCD). However, as the efficacy of EEN varies from patient to patient, there is a need to distinguish between responders and nonresponding patients. This study had two aims. First, to develop a model to predict EEN-induced clinical remission (weighted pediatric CD activity index [wPCDA] ≤ 12.5) using baseline clinical information. Second, to develop a model to predict corticosteroid-free sustained clinical remission post-EEN induction (wPCDA ≤ 12.5, for ≥36 weeks after EEN).

We applied machine learning to clinical and laboratory data from a prospectively followed cohort of pCD patients who received EEN as their first treatment for CD (n = 308). This learning algorithm used feature selection and k-fold (internal) cross-validation to systematically find the model with the best combination of features and hyperparameter settings. To estimate the quality of the learned model, we used k-fold (external) cross-validation.

Clinical, laboratory, and treatment data were compiled into two different datasets: EEN clinical remission at the end of EEN treatment (mean of 60 days; n = 114) and corticosteroid-free sustained clinical remission post-EEN induction (n = 206). Our resulting models were effective, with external area under the curves of 0.65  $\pm$  0.015 and 0.60  $\pm$  0.018. Moreover, a permutation label test showed that our learning process was stable and significantly different from chance, at p-values of 0.002 and 0.01, respectively.

Our models, based on accessible clinical features, were able to effectively predict EEN success above chance. This supports the plausibility of building clinical tools to assist precision therapy for pCD patients.