Infinite Feature Selection on SHORE based biomarkers reveals connectivity modulation after stroke

Connectomics is gaining increasing interest in the scientificand clinical communities. It consists in deriving models ofstructural or functional brain connections based on some localmeasures. Here we focus on structural connectivity asdetected by diffusion MRI. Connectivity matrices are derivedfrom microstructural indices obtained by the 3D-SHORE.Typically, graphs are derived from connectivity matrices andused for inferring node properties that allow identifying thosenodes that play a prominent role in the network. This informationcan then be used to detect network modulationsinduced by diseases. In this paper we take a complementaryapproach and focus on link as opposed to node properties. Wehypothesize that network modulation can be better describedby measuring the connectivity alteration directly in the formof modulation of the properties of white matter fiber bundlesconstituting the network communication backbone. The goalof this paper is to detect the paths that are most altered by thepathology by exploiting a feature selection paradigm. Temporalchanges on connection weights are treated as featuresand those playing a leading role in a patient versus healthycontrols classification task are detected by the Infinite FeatureSelection (Inf-FS) method. Results show that connectionpaths with high discriminative power can be identified that areshared by the considered microstructural descriptor allowinga classification accuracy ranging between 83% and 98% forthe different indices.