An Interpretable Machine Learning Model to Predict Cortical Atrophy in Multiple Sclerosis

To date, the relationship between central hallmarks of multiple sclerosis (MS), such as white matter (WM)/cortical demyelinated lesions and cortical gray matter atrophy, remains unclear. We investigated the interplay between cortical atrophy and individual lesion-type patterns that have recently emerged as new radiological markers of MS disease progression. We employed a machine learning model to predict mean cortical thinning in whole-brain and single hemispheres in 150 cortical regions using demographic and lesion-related characteristics, evaluated via an ultrahigh field (7 Tesla) MRI. We found that (i) volume and rimless (i.e., without a rim of iron-laden immune cells) WM lesions, patient age, and volume of intracortical lesions have the most predictive power; (ii) WM lesions are more important for prediction when their load is small, while cortical lesion load becomes more important as it increases; (iii) WM lesions play a greater role in the progression of atrophy during the latest stages of the disease. Our results highlight the intricacy of MS pathology across the whole brain. In turn, this calls for multivariate statistical analyses and mechanistic modeling techniques to understand the etiopathogenesis of lesions.

Automated summary

The relationship between central hallmarks of multiple sclerosis (MS), such as white matter (WM)/cortical demyelinated lesions and cortical gray matter atrophy, remains unclear. A machine learning model was employed to predict mean cortical thinning in different brain regions using demographic and lesion-related characteristics. The study found that volume and rimless WM lesions, patient age, and volume of intracortical lesions have the most predictive power.