Measuring the agreement between brain connectivity networks.

Investigating the level of similarity between two brain networks, resulting from measures of effective connectivity in the brain, can be of interest from many respects. In this study, we propose and test the idea to borrow measures of association used in machine learning to provide a measure of similarity between the structure of (un-weighted) brain connectivity networks. The measures here explored are the accuracy, Cohen's Kappa (K) and Area Under Curve (AUC). We implemented two simulation studies, reproducing two contexts of application that can be particularly interesting for practical applications, namely: i) in methodological studies, performed on surrogate data, aiming at comparing the estimated network with the corresponding ground-truth network; ii) in applications to real data, when it is necessary to compare the structure of a network obtained in a specific subject with a reference (e.g. a baseline condition or normative data). In the simulations, the level of similarity between two networks was manipulated through different factors. We then investigated the effect of such manipulations on the measures of association. Results showed how the three parameters modulated their values according to the level of similarity between the two networks. In particular, the AUC provided the better performances in terms of its capability to synthetize the similarity between two networks, showing high dynamic and sensitivity.