Abstract

Background: Cognitive impairment is a common early feature of multiple sclerosis (MS). This decline may be linked to microstructural damage in juxtacortical U-fibers, which are highly vulnerable to early demyelination and essential for connectivity between adjacent cortical gyri. While these changes are undetectable in conventional MRI protocols, we employed a diffusion MRI and Machine Learning (ML) approach to identify subtle alterations in these short-range fibers. Objective: We designed a proof of concept study to investigate the structural integrity of juxtacortical U-fibers and its correlation with cognitive impairment in MS patients by using ML as a robust analytical framework for understanding brain-behavior relationships. Methods: Thirty-five cognitively preserved healthy controls and fifty-eight MS patients, including cognitively preserved and cognitively impaired individuals, were included. The dataset comprised fractional anisotropy (FA) maps extracted from 100 U-fiber regions. Working memory was assessed using the Paced Auditory Serial Addition Test (PASAT), expressed as Z-scores adjusted for age and education. Sequential forward selection (SFS) and fifteen regression models were trained to account for PASAT Z-scores. Model performance was evaluated using 5-fold nested cross-validation with mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R2). Results: Tree-based ensemble models showed superior performance in estimating the relationship between FA measures and working memory (CatBoost: MSE = 0.256 +/- 0.068; MAE = 0.421 +/- 0.068; R2 = 0.488 +/- 0.128). SFS identified sixteen FA features, with left superior temporal and right inferior parietal regions emerging as the most informative features. Conclusion: ML analysis reveals that Diffusion MRI-derived short-range association U-fiber FA features can estimate working memory performance with clinically meaningful accuracy in MS, supporting their potential to detect early juxtacortical dysconnectivity, serving as a biomarker for early cognitive monitoring.