Accurate discrimination of alcoholic patients using a multivariate SVM approach of mGluR5 PET

Objectives: For analyzing neuroimaging data, the most commonly used analysis approach is based on mass univariate statistical methods, where statistical parametric maps are obtained considering all voxels as independent from one another. In multivariate pattern recognition methods, in particular using machine learning, predictive models for group classification can be devised, where the joint combination of all voxels defines the statistical models which can compute predictions on new individual data sets. Apart from diagnostic classification, multivariate techniques can also address specific questions such as prognosis from longitudinal datasets. Recent evidence supports a central role for the metabotropic glutamate receptors subtype 5 (mGluR5) in alcohol drinking behavior and dependence, encouraging future studies to assess its potential predictive value for relapse. As a first step, we constructed the voxel-based discriminative cerebral mGluR5 pattern for patients with alcohol dependence using a multivariate support vector machine (SVM) approach to define the spatial discriminative features and evaluate the classification performance thereof. Methods: Dynamic 90 minute [18F]FPEB PET scans were acquired in 16 patients with a DSM-IV diagnosis of alcohol dependence (ALC: age = 46 ± 8 years, reported lifetime alcohol use = 26.5 ± 12.9 years) and 32 age-matched healthy controls (HC: age = 45 ± 13 years), on a HiRez Biograph 16 slice PET/CT camera. Voxel-wise mGluR5 availability was quantified by the [18F]FPEB total distribution volume. A multivariate mGluR5 brain pattern was identified by the weight vector output using a linear classifier derived from binary classification between the ALC and HC groups, using a SVM with a soft margin and linear kernel. To balance the group of ALC patients, the HC group was divided into two folds, and the SVM performance were obtained as the average over the two HC folds. Performance of the classifier (sensitivity and specificity) was evaluated by leave-one-out (LOO) cross validation for each group, and parameter optimization was performed by inner loop cross validation. Results: The discriminative multivariate pattern between ALC and HC was characterized by a decreased in mGluR5 availability in a widespread corticosubcortical network including anterior and posterior cingulate cortex, caudate nucleus, middle frontal gyrus, superior parietal gyrus, superior temporal gyrus and cerebellum (Figure 1). For the multivariate SVM-bas