Machine-Learning Models Improve Magnetic Resonance Enterography based Assessment of Mucosal Healing in Pediatric Crohn's Disease

Itai Guez, Gili Focht, Mary-Louise C. Greer, Ruth Cytter-Kuint, Li-tal Pratt, Denise Castro, Dan Turner, Anne M. Griffiths, and Moti Freiman

Abstract

Mucosal healing is considered a major Crohn's Disease (CD) treatment goal. Simple Endoscopic Score of Crohn's Disease (SES-CD) assessed by ileocolonoscopy is a standard quantitative endoscopic score for mucosal healing. The most frequent disease location is the terminal ileum (TI); however the TI is narrow and intubation failure is common, occurring in 20-25% of pediatric CD ileocolonoscopies. Our aim is to assess the added value of machine-learning models in imputing TI SES-CD from Magnetic Resonance Enterography (MRE) data.

Materials and Methods

This is a substudy of the ImageKids study in which 240 pediatric patients with CD (22 centers, age 14.2 ± 2.5 years) underwent a baseline ileocolonoscopy scored by SES-CD, followed by an MRE examination within 14 days. We used a non-linear machine-learning random-forest (RF) model and a multiple linear regression model (MLR) with two sets (in total we produced 4 models) of biomarkers as input: Magnetic Resonance Index of Activity (MaRIA) biomarkers (wall thickness, ulcers, edema, relative contrast enhancement (RCE)) and the Pediatric Inflammatory Crohn's MRE Index (PICMI) biomarkers (wall thickness, diffusion weighted imaging, ulcers, edema, comb sign) - the latter without gadolinium-based contrast injection or post-contrast T1-weighted sequences. We assessed the added-value of machine-learning models in imputing SES-CD values by comparing RF-MaRIA vs. MLR-MaRIA, and for RF-PICMI vs. MLR-PICMI. We determined which set of biomarkers (PICMI vs. MaRIA) provides more accurate predictions. The mean squared error (MSE) from the reference endoscopic SES-CD was used as the measure of accuracy. Statistical analysis was performed with the Wilcoxon non-parametric test in order to determine whether the median of validation MSE differed between two given models.

Conclusion

Machine-learning non-linear models are more accurate than linear regression models in predicting TI SES-CD when using the same MRE-based biomarkers. Models based on PICMI biomarkers are more accurate than models based on MaRIA biomarkers in predicting TI SES-CD using same data from Imagekids.

Itai presenting his research for the ESPGHAN 2021

The Imagekids study was supported by a grant from AbbVie.