The support vector machine (SVM) is a supervised classification method [1]. SVM "learns" how to distinguish data that belong to one of two possible groups (i.e.: healthy subjects, patient) using a training dataset, than "builds" a prediction model that is used to assign new data (testing dataset) into one group or the other. The present study focuses on the application of nested SVM to prediction of Alzheimer's disease (AD) on the basis of very little number of features (1 to 10) extracted from structural and diffusions MR scans.

Subjects and data acquisition

We recruited 40 patients (25F/15M) diagnosed with probable AD (age=69.5±6.5; MMSE=19.1±4.3, range: 11-28) and 28 (12F/16M) healthy subjects (HS; age=66.4±7.0; MMSE=28.8±1.6, range=25-30), age and gender matched to the AD group. All subjects underwent an MRI acquisition at 3.0T including: (1) T1-weighted (MDEFT) scan (TR=1338ms, TE=2.4ms, Matrix=256x224, n. slices=176, thick=1mm) and (2) Diffusion Weighted (DW) twice-refocused spin echo echo-planar imaging (SE EPI; TR=7s, TE=85ms, b factor=1000s/mm², isotropic resolution=2.3mm³), collecting seven images with no diffusion weighting (b=0) and 61 images with diffusion gradients applied along 61 non-collinear directions.

Image analysis

The MDEFTs were first processed with SPM8 to yield maps of gray matter (GM), white matter (WM) and CSF volume in native space. Brain tissues volumes (GMvol, WMvol, CSFvol) were calculated for each subject. To account for subjects' head size differences, GMvol and WMvol were expressed as percent of the total intracranial volume, and yielded the GM fraction (GMf) and WM fraction (WMf). DW images were processed (using FSL and CAMINO) to compute fractional anisotropy (FA), mean diffusivity (MD), radial diffusivity (RAD) and axial diffusivity (AXD). The FA maps were warped to the T1 images (used for the tissues segmentation). This way we could compute for each subject the mean values of FA, MD, RAD, AXD in GM and WM respectively.

SVM analysis

The 10 MR features obtained from the image analysis (GMf, WMf; FA, MD, RAD, AXD in GM and WM) were used to compute 1023 different n-features SVM (n=1,2,...,10) classifiers, one for each possible combination (1 by 1, 2 by 2, …, 9 by 9, all 10) of the features (brute force approach). In particular, we used non-linear SVMs with gaussian (RBF) kernel, using custom-made Matlab script, exploiting the libSVM library [2]. For each of the 1023 classifiers, the parameters of the SVM model (the soft margin constant C and the width of the gaussian kernel γ) were tuned through a grid search and leave-one-out (LOO) nested cross-validation (CV) [3]. For each classifier, the optimized values of C and γ were then used to create the optimized SVM model, whose classification accuracy (i.e., proportion of AD and HS subjects correctly classified), sensitivity (i.e., the proportion of AD patients correctly classified) and specificity (i.e., the proportion of HS correctly classified) were computed.

FIG.1 summarizes the main results of the study. Surprisingly, with SVM exploiting only one features (1-feature SVM), the best accuracy (83.82%) and specificity (78.57%) were obtained with one of the WM features, namely with AXD (FIG.2). Similarly, the best sensitivity (92.50%) was obtained with MD of WM (FIG.2). These classification performances were improved by SVM exploiting more features (n-features SVM, n>1): in particular the overall best accuracy (89.71%) was obtained with two different SVMs: the 4-features SVM exploiting the GM diffusion measures and WMf and the 9-features SVM including all measures but FA of WM, even if the specificity and sensitivity of the two SVMs were different (FIG.3).In the current study, we investigated the classification between HS and patients with AD, using multimodal (structural and diffusion) MR data as input to SVM classifiers. The 1-feature SVM performances (FIG.2) indicate that, regarding overall brain measures, the diffusion properties of WM provide a better discrimination between AD and HS. In particular the false positive rate, as highlighted by the specificity of AXD in WM (78.57%), is largely better than that obtained with GM measures. Using brute force approach we explored all the possible n-features SVMs and we found that best classification performance (ACC=89.71%) was obtained combining diffusion and structural features coming from both GM and WM (FIG.3). This finding indicates that, as expected, structural and diffusion MR properties of brain provide complementary information on the dementia status. However, it can be noticed that the best sensitivity (97.50%) was obtained with 3-features SVM including only GM measures (FIG.3), but outweighted by a very poor specificity (57.14%), indicating that aging is a confounding factor mainly affecting GM measures.