Fig. 6.

The top performing ten configurations for the prediction of each non-imaging variable by dataset after deconfounding. [A,B,C] (HCP Data) Each data point represents a different configuration strategy that may vary in terms of parcellation strategy, the functional connectivity estimation method, whether tangent space parameterization was employed, whether tangent space regularization was employed, and the predictor that was used. The first word indicates the parcellation strategy, and the second word refers to the functional connectivity estimation method. The third word refers to the geometry in which classifier/predictor is applied, ambient referring to non-tangent space and tangent referring to the projected covariance matrices in tangent space. If non-isotropic shrinkage was applied after projecting covariance matrices to tangent space, the fourth word will be “shrinkage”. The last word indicates the type of classifier/predictor that was used. The highlighted red blocks show the recommended pipelines (rationale explained in Section 5), and red dotted lines highlight the point when the error bar of pipeline after the dotted line is out of range from the error bar of the top (first) pipeline.