Classification between Norm and Pathology Based on Spectral Features of Network Brain Structures

This study aims at tackling the problem of brain network classification using machine learning algorithms based on the spectra of the networks' matrices. Three approaches are discussed. First, linear and tree-based models are run on the vectors of sorted eigenvalues of the adjacency matrix, the Laplacian matrix and the normalized Laplacian. Second, SVM classifier with kernels based on information divergence between the eigenvalue distributions is trained. Third, the SVM classifier is fed with a kernel that uses a metric arising as a solution to a transportation problem. The information divergence approach gives the most promising results in the classification of autism spectrum disorder versus typical development and of the carriers versus noncarriers of an allele associated with the high risk of Alzheimer disease. However, the results obtained are extremely sensitive to the parameters of the empirical density reconstruction needed to compute the kernels. Ultimately, similar methods bypassing this issue might be of research interest.