Unsupervised Pattern Analysis to Differentiate Multiple Sclerosis Phenotypes Using Principal Component Analysis on Various MRI Sequences

Abstract

Background: Multiple sclerosis (MS) has two main phenotypes: relapse-remitting MS (RRMS) and progressive MS (PMS), distinguished by disability profiles and treatment response. Differentiating them using conventional MRI is challenging. Objective: This study explores the use of scaled subprofile modelling using principal component analysis (SSM/PCA) on MRI data to distinguish between MS phenotypes. Methods: MRI scans were performed on patients with RRMS (n = 30) and patients with PMS (n = 20), using the standard sequences T1w, T2w, T2w-FLAIR, and the myelin-sensitive sequences magnetisation transfer (MT) ratio (MTR), quantitative MT (qMT), inhomogeneous MT ratio (ihMTR), and quantitative inhomogeneous MT (qihMT). Results: SSM/PCA analysis of qihMT images best differentiated PMS from RRMS, with the highest specificity (87%) and positive predictive value (PPV) (83%), but a lower sensitivity (67%) and negative predictive value (NPV) (72%). Conversely, T1w data analysis showed the highest sensitivity (93%) and NPV (89%), with a lower PPV (67%) and specificity (53%). Phenotype classification agreement between T1w and qihMT was observed in 57% of patients. In the subset with concordant classifications, the sensitivity, specificity, PPV, and NPV were 100%, 88%, 90%, and 100%, respectively. Conclusions: SSM/PCA on MRI data revealed distinctive patterns for MS phenotypes. Optimal discrimination occurred with qihMT and T1w sequences, with qihMT identifying PMS and T1w identifying RRMS. When qihMT and T1w analyses align, MS phenotype prediction improves.