Structural and functional connectomes in people with multiple sclerosis
One of the challenges in multiple sclerosis is that lesion volume does not correlate with symptom severity. Advanced techniques such as diffusion and functional MRI allow imaging of the brain's connectivity networks, which may provide better insight as to brain-behavior relationships in impairment and compensation in multiple sclerosis. We aim to build machine learning models based on structural and functional connectomes to classify a) healthy controls versus people with multiple sclerosis and
... tiple sclerosis and b) impaired versus not impaired people with multiple sclerosis. We also aim to identify the most important imaging modality for both classification tasks, and, finally, to investigate which brain regions' connectome measures contribute most to the classification. Fifteen healthy controls (age=43.6 ± 8.6, 53% female) and 76 people with multiple sclerosis (age: 45.2 ± 11.4 years, 65% female, disease duration: 12.2 ± 7.2 years) were included. Twenty-three people with multiple sclerosis were considered impaired, with an Expanded Disability Status Scale of 2 or higher. Subjects underwent MRI scans that included anatomical, diffusion and resting-state functional MRI. Random Forest models were constructed using structural and static/dynamic functional connectome measures independently; single modality models were then combined for an ensemble prediction. The accuracy of the models was assessed by the area under the receiver operating curve. Models that included structural connectomes significantly outperformed others when classifying healthy controls and people with multiple sclerosis, having a median accuracy of 0.86 (p-value<0.05, corrected). Models that included dynamic functional connectome metrics significantly outperformed others when distinguishing people with multiple sclerosis by impairment level, having a median accuracy of 0.63 (p-value<0.05, corrected). Structural connectivity between subcortical, somatomotor, and visual networks was most damaged by multiple sclerosis. For the classification of patients with multiple sclerosis into impairment severity groups, the most discriminatory metric was dwell time in a dynamic functional connectome state characterized by strong connectivity between and among somatomotor and visual networks. These results suggest that damage to the structural connectome, particularly in the subcortical, visual, and somatomotor networks, is a hallmark of multiple sclerosis, and, furthermore, that increased functional coordination between these same regions may be related to severity of motor disability in multiple sclerosis. The use of multi-modal connectome imaging has the potential to shed light on mechanisms of disease and compensation in multiple sclerosis, thus enabling more accurate prognoses and possibly the development of novel therapeutics.