A deep learning-enriched framework for analyzing brain functional connectivity

Cognitive and motor functions require a coordinated communication among brain regions, with the directionality of interactions playing a key role, as the brain relies on functional asymmetries of reciprocal connections. Predictive models based on deep learning approaches could represent valuable tools for processing functional connectivity. However, these approaches are mainly adopted for decoding different brain states, but not for characterizing the information flow of functional networks. Here, we design a deep learning-enriched framework for analyzing spectral directed functional connectivity. The knowledge learned by a novel interpretable convolutional neural network ('Functional-Connectivity-Net', FCNet) - trained to discriminate brain states from functional connectivity - is used to define novel inflow and outflow measures, characterized for being non-linear, and for combining the information across brain regions and frequencies in an optimally discriminative way. Moreover, network decision is explained via DeepLIFT, revealing the most relevant frequency contents and connectivity inflow/outflow. We apply our approach to EEG functional connectivity estimated at both the scalp and cortex level, during motor imagery tasks. The network explanations match the known markers of spectral connectivity changes underlying motor imagery, and the network-based measures capture connectivity changes with high strength and significance like graph theory measures (in degree, out degree, authority, hubness). Our framework is helpful to elucidate the predictability of brain functional networks, and the most informative frequencies and connectivity inflow/outflows for the analyzed brain states.