Large-scale comparative analysis reveals top graph signal processing features for subject identification

In magnetic resonance imaging, graph signal processing (GSP) is an analytical framework that enables to express regional functional activity time courses in terms of the underlying structural connectivity backbone. To this end, several parameters must be set during the processing of structural and functional data, and a variety of output features have been proposed. While emerging applications of the GSP framework have shown clear merits to reveal the neural underpinnings of brain disorders, behavioural facets or individuality, at present, the optimal parameter choices and feature types for an outcome of interest remain unknown. Here, we fill this gap by conducting a large-scale comparative analysis across parameter choices and candidate feature types. First, we show that all the studied factors of variation within the GSP pipeline significantly modulate feature vector patterns and feature coefficient values, evidencing the importance of an exhaustive characterization. Second, focusing on the ability to fingerprint individual subjects, we demonstrate that power spectral density and the structural decoupling index are the most all-around feature types, which harmoniously balance robustness to external sources of variation (head movement and acquisition settings), parsimony of the telling feature set, and generalization to altered parcellation specificities. Our results emphasize the importance, for future GSP studies, of carefully considering the undertaken structural connectivity and functional parameter choices as a function of the outcome measure of interest. More globally, they also highlight the relevance of large-scale comparative strategies in optimizing an analytical pipeline towards a specific goal. Our reported methodology can seamlessly be extended to other analytical approaches and outcome measures of interest, which we hope will be of use for future researchers in and outside the GSP subfield.