Fig. 4.

Support network of EC links to identify subjects and generalization capability. A) Extracted links that contribute to the classification for Datasets A1 and B, obtained using recursive feature elimination (RFE). The ROIs are grouped in anatomical pools, as detailed in Supplementary Table S1. B) Average RFE ranking over the ROI anatomical pools in A. Darker color indicates more important links for the classification. C) Overlap between the two signatures for Datasets A1 and B as a function of selected links. The curve represents the amount of common links in the data. Shaded areas represent different quantiles of the surrogate distribution of common links under the null-hypothesis of random rankings. The color of the curve indicates the probability of the corresponding amount of common links under the null-hypothesis (here p-value ¡ 0.001 when considering more than 1% of the total links, namely 40 links). D) Extrapolation of the size of the support network in A for Dataset B when increasing the number of subjects up to 1000 (x-axis), for 10 repetitions. The shaded area corresponds to 10 repetitions. The curves compare three approximations, the best one displayed in orange (as indicated by the fitting error SSE) corresponds to a sublinear power law (exponent equal 0.6). E) Same as D for the accuracy based on EC. The plotted data points correspond to EC in Fig. 3C and the variability to 100 repetitions. The best approximation also corresponds to a power law (in orange), which is very close to a lognormal relationship.