A Graph Signal Processing Framework for the Classification of Temporal Brain Data

Abstract

Graph Signal Processing (GSP) addresses the analysis of data living on an irregular domain which can be modeled with a graph. This capability is of great interest for the study of brain connectomes. In this case, data lying on the nodes of the graph are considered as signals (e.g., fMRI time-series) that have a strong dependency on the graph topology (e.g., brain structural connectivity). In this paper, we adopt GSP tools to build features related to the frequency content of the signals. To make these features highly discriminative, we apply an extension of the Fukunaga-Koontz transform. We then use these new features to train a decision tree for the prediction of autism spectrum disorder. Interestingly, our framework outperforms state-of-the-art methods on the publicly available ABIDE dataset.