Fig. 6.

The effects of selective network occlusion on model accuracy. (Top) the process by which occlusion AUROCs are estimated; either all inner edges of a given network, or all edges connecting to a network, are selected. The network edges are then scrambled (see Fig. 1), and the selected edges are placed among one half of the scrambled edges, and in the other half left out. These two sets are then trained on $2\times 300$ independent neural networks, and the resulting AUROCs are compared. (Bottom) The results. Considering only inner edges, the only statistically significant effect, after Bonferroni-Holmes correction, was the salience networks on resting-state data. Considering all connecting edges, all three networks had a significant effect on the classification of sex in resting-state data, while both the default mode network and, more strongly, the central executive network, appeared to have an effect in classification of task data. The nonparametric Mann-Whitney $U$-test was used to test for statistical significance. Final model means and ensemble results are shown in Table 1.