Identification of faulty DTI-based sub-networks in autism using network regularized SVM

Neural imaging studies of autism spectrum disorders (ASD) have consistently demonstrated deficits in connectivity. In this paper, we propose a new network regularized support vector machines (SVM) method to identify the faulty subnetworks associated with ASD using diffusion tensor imaging (DTI). After constructing the brain connectivity network of each subject using DTI, the SVM-recursive feature elimination (RFE) algorithm is adopted to identify the faulty sub-networks in order to distinguish ASD from typical developing (TD) controls. Since connections in the network are not independent of each other, their topological proximities are incorporated as the network regularization of SVM-RFE by using the graph Laplacian to obtain robust sub-networks. Experiments on both simulated and clinical datasets showed a better performance of the proposed method in faulty sub-network identification, compared with the traditional SVM-RFE method.