Abstract

--- - "Cognitive impairment is a common and disabling symptom in patients with relapsing-remitting multiple sclerosis (RRMS), yet early detection remains a clinical challenge. This study presents a machine learning (ML) framework for identifying subcortical white matter biomarkers associated with verbal episodic memory impairment utilizing diffusion-weighted imaging (DWI). A total of 93 participants (58 RRMS patients with cognitive preservation (RRMS-CP) and with cognitive impairment (RRMS-CI), and 35 healthy controls with cognitive preservation (HC-CP) were evaluated using the Auditory Verbal Learning Test and stratified into three cognitive groups. U-fiber diffusivity metrics-including fractional anisotropy, axial diffusivity, radial diffusivity, and mean diffusivity-were extracted from 100 white matter regions. Feature selection via Sequential Forward Selection was combined with multiple classifiers, including Random Forest and Linear Discriminant Analysis. Our models achieved up to 100% accuracy (sensitivity: 98.64%, specificity: 98.17%) in distinguishing between HC-CP and RRMS-CP patients, 96.37% accuracy (sensitivity: 100%, specificity: 97.50%) between HC-CP and RRMS-CI, 97.23% accuracy (sensitivity: 98.75%, specificity: 96.00%) between RRMS-CP and RRMS-CI, and 89.2% accuracy (sensitivity: 91.09%, specificity: 91.58%) across all three groups." - We acknowledge limitations, including potential information leakage in feature selection, limited diffusion directions, and the need for advanced preprocessing pipelines. These limitations will be addressed in future work with larger cohorts and nested cross-validation. The findings demonstrate the potential of interpretable ML for identifying DWI-based biomarkers of cognitive dysfunction in MS.