A158 APPLYING MACHINE LEARNING FOR PREDICTING TREATMENT RESPONSE TO VEDOLIZUMAB IN PEDIATRIC IBD BY SERUM METABOLOMICS

TL;DR

This study uses machine learning and serum metabolites to predict how children with IBD will respond to vedolizumab treatment.

Contribution

The study introduces a machine learning model using metabolomic data to predict treatment response in pediatric IBD patients.

Findings

A machine learning model predicted treatment response with an AUC of 0.95 for Crohn's disease patients at week 14.

For ulcerative colitis patients at week 30, the model achieved an AUC of 0.81.

Key metabolites like N-Acetyl-Aspartic acid and Isoleucine showed high predictive importance.

Abstract

Vedolizumab (VDZ) is effective to induce remission in children with Crohn disease (CD) and ulcerative colitis (UC), but effectiveness varies.

Metabolites produced by interactions between intestinal microbiota and host metabolic processes can be useful to identify metabolome signatures that may preferentially favor response to a specific therapeutic class. Therefore, metabolomic studies can potentially inform precision medicine in Inflammatory Bowel Diseases (IBD).

This study aimed to apply machine learning to assess metabolites as potential predictors for forecasting the response to VDZ treatment.

VedoKids is a multicenter, prospective, observational cohort study, designed to report the effectiveness and safety of VDZ. Children aged 0-18 years, diagnosed with IBD, who initiated VDZ treatment at any stage of their condition, were subjected to comprehensive assessments at the onset and subsequently at 2, 6, 14, 30, 54 weeks and thereafter. Detailed demographic, clinical, and safety information was meticulously recorded in a prospective manner throughout the study period. Metabolomic profiling was conducted in serum at three time points: baseline, 14 weeks, and 30 weeks. Metabolites were identified using a quantitative metabolomics approach utilizing DI/LC-MS/MS technology for the analysis of serum samples.

We constructed a learning algorithm to predict treatment response to identify the most relevant serum metabolites subset. The algorithm consisted of a Random Fores (RF) model and maximum relevance minimum redundancy (mRMR) feature selection. Response was defined as decrease of ≥20 points in PUCAI or >20 points in wPCDAI and clinical remission as PUCAI<10 or wPCDAI<12.5.

We were able to train different RF models using clinical, pre-treatment, and serum metabolite data. Results for the experiment performed with CD patients at week 14 shows an AUC = 0.95 with N-Acetyl-Aspartic acid, Fumaric acid, and Glutamine scoring among the features with highest predictive importance. Results for the experiment performed with UC patients at week 30 shows an AUC = 0.81 with Isoleucine, Malic acid, and N-Acetyl-Glutamic acid scoring among the features with highest predictive importance.

Our results suggest that it is possible to produce predictor capable of predicting response to VDZ treatment using clinical and metabolome data in children with IBD. Importantly, some of the identified metabolites have been previously associated with IBD pathophysiology.

CIHR