Fig. 3.

Comparison between EC and FC in subject identification. A) Classification pipeline used to assess the generalization of performance. The full set of connectivity measures (here EC) over all fMRI sessions was split into two groups: a train set and a test set. We use z-scores calculated over the elements of each session matrix; see Eq. (16) in Methods for details. We trained the classifier —with or without previously applying PCA— and evaluated the classification accuracy on the test set. Details about the algorithms are given in Section 4.3.5. B) Performance of multinomial logistic regression (MLR, left panel) and 1-nearest-neighbor (1NN, right panel) classifiers when increasing the number of sessions per subject used as training set with Dataset A1. The mean (solid curve) and standard deviation (colored area) were calculated for 100 folds of randomly-split train-test sets (the number of sessions per subject for the training set is indicated by the value of the x-axis; the test set comprised all remaining sessions). C) Same as B when varying the number of subjects using Dataset B, using a single training session per subject (leaving 9 sessions per subject as test).