The conventional statistical approach for analyzing resting state functional MRI (rs-fMRI) data struggles to accurately distinguish between patients with multiple sclerosis (MS) and those with neuromyelitis optic spectrum disorders (NMOSD), highlighting the need for improved diagnostic efficacy. In this study, multilevel functional metrics including resting state functional connectivity, amplitude of low frequency fluctuation (ALFF), and regional homogeneity (ReHo) were calculated and extracted from 116 regions of interest in the anatomical automatic labeling atlas. Subsequently, classifiers were developed using different combinations of these selected features to distinguish between MS and NMOSD. Compared to models based on individual MRI features, support vector machine (SVM) and logistic regression (LR) models that integrated multilevel functional features such as RSFC, ALFF, and ReHo demonstrated the highest levels of performance on the testing cohorts (SVM, AUC = 0.857; LR, AUC = 0.929). Adding structural features of gray matter volume (GMV) data did not notably improve the classification performance of the machine learning models using multilevel rs-fMRI features. Notably, similar trends were observed across different brain templates, with models based on RSFC, ALFF, and ReHo yielding optimal performance. These findings suggest that utilizing multilevel fMRI features can effectively differentiate between MS and NMOSD patients.
Keywords: Machine learning; Multiple sclerosis; Neuromyelitis optic spectrum disorders; Resting state functional magnetic resonance imaging.
© 2025. The Author(s).