To develop advanced methods for the analysis of resting state (RS) functional connectivity (FC) to classify patients with multiple sclerosis (MS) according to disease phenotype.

RS FC graph theoretical analysis is helping to shed light into brain functional reorganization in MS. The application of advanced network-based methods exploiting machine learning on RS FC data may produce important clues to support MS patients’ classification.

RS fMRI scans were obtained from 46 healthy controls (HC) and 113 MS patients (62 relapsing-remitting [RR] and 51 progressive [P]MS). Dominant sets clustering was used to group covariance-based RS FC matrices into clusters of subjects sharing some similarities in their network configuration. Support vector machines (SVMs) were then used to classify disease phenotypes exploiting a representation of networks based on their geodesic distance from cluster means. Finally, a sensitivity analysis on the trained classifier was used to identify clusters and connections more relevant for classification.

The described machine learning tool was able to classify RRMS patients from HC with an accuracy of 72.5%, PMS patients from HC with an accuracy of 84.1% and PMS from RRMS patients with an accuracy of 76%. With the sensitivity analysis on trained SVMs we found that increased connectivity within the basal ganglia sub-network (particularly involving the bilateral thalami) and decreased RS FC within the temporal sub-network contributed to an accurate classification of both RRMS and PMS patients from HC. Moreover, decreased RS FC within the occipital and parietal sub-networks contributed to differentiate PMS patients from HC.

A combination of different machine learning principles allowed to detect specific RS FC configurations of our study subjects, which allowed to classify MS patients with different clinical phenotypes from HC with a good accuracy. Distinct sub-networks abnormalities contributed to an accurate classification of different phenotypes.