Figure 3

Connectivity, directionality and model errors. (A) shows the different networks captured by functional connectivity fMRI analysis based on correlation coefficients (Pearson “r”), and the pattern of activation/inhibition expressed by coefficients of the EGN-B adjacency matrix A. The first captures strong positive within-network correlation, as well as negative ones between nodes of different networks. The EGN-B model unveils a more complex pattern, where positive and negative “directed” modulations are present even within a single network. (B) shows the predictive capabilities of the model by comparing real fMRI data of one node (red line) with simulated data (blue line). The latter has been obtained by using an EGN-B connectivity matrix estimated by means of the first 90 real data samples (yellow area). The green area highlights optimal prediction for approximately 45 samples, while prediction accuracy tends to drop thereafter. (C) shows the average prediction errors for different size of the estimation dataset across the entire sample. (D) depicts the differences between EGN-B and a linear model (reported in the Supplementary Information) in terms of mean and standard deviation of prediction errors in 40 healthy participants.