Multivariate pattern classification of gray matter pathology in multiple sclerosis

    loading  Checking for direct PDF access through Ovid

Abstract

Univariate analyses have identified gray matter (GM) alterations in different groups of MS patients. While these methods detect differences on the basis of the single voxel or cluster, multivariate methods like support vector machines (SVM) identify the complex neuroanatomical patterns of GM differences. Using multivariate linear SVM analysis and leave-one-out cross-validation, we aimed at identifying neuroanatomical GM patterns relevant for individual classification of MS patients.

We used SVM to separate GM segmentations of T1-weighted three-dimensional magnetic resonance (MR) imaging scans within different age- and sex-matched groups of MS patients with either early (n = 17) or late MS (n = 17) (contrast I), low (n = 20) or high (n = 20) white matter lesion load (contrast II), and benign MS (BMS, n = 13) or non-benign MS (NBMS, n = 13) (contrast III) scanned on a single 1.5 T MR scanner.

GM patterns most relevant for individual separation of MS patients comprised cortical areas of all the cerebral lobes as well as deep GM structures, including the thalamus and caudate. The patterns detected were sufficiently informative to separate individuals of the respective groups with high sensitivity and specificity in 85% (contrast I), 83% (contrast II) and 77% (contrast III) of cases.

The study demonstrates that neuroanatomical spatial patterns of GM segmentations contain information sufficient for correct classification of MS patients at the single case level, thus making multivariate SVM analysis a promising clinical application.

Highlights

□ First study using multivariate image analysis for classification of grey matter in MS. □ Up to 92% sensitivity and 90% classification accuracy in individual's unseen data. □ Advantage of multivariate analysis in allowing inferences at the individual's level.

Related Topics

    loading  Loading Related Articles