Comprehensive Summary
The study presented by Borra and Magosso introduces a deep learning-enriched framework, using a novel interpretable convolutional neural network (FCNet), to analyze spectral directed functional connectivity and characterize the information flow in brain functional networks. The research was performed by designing and training FCNet to decode right-hand versus left-hand motor imagery based on functional connectivity derived from EEG signals at both the scalp and cortex levels, using data from two distinct datasets. The findings show that the network explanations align with known neurophysiological markers, specifically identifying the upper alpha-band (10–12 Hz) and beta-band (20–22 Hz) as the most relevant frequency components for discrimination. Furthermore, the FCNet-based inflow and outflow measures were highly sensitive, demonstrating high statistical significance and effect size in detecting changes between motor states, similar to traditional graph theory measures, and consistently revealed a hemispheric lateralization in connectivity modulations.
Outcomes and Implications
This research is clinically important because characterizing patterns of inter-regional communication is important for describing normal mechanisms of perception, attention, and learning, as well as their pathological changes in conditions such as neurodegenerative diseases. The framework was developed using a within-subject training strategy to highlight participant-specific connectivity signatures, a choice intended to open important application scenarios for individual patients to better support the analysis of neuropathology and guide the personalization of treatments. Although presented as a preliminary proof-of-concept for neural data analysis, the methodology provides a useful tool for neuroscientists to prospectively support data-driven investigations of brain oscillatory interactions during cognitive tasks. Future research includes performing simulations and scaling up experiments to different cognitive tasks and recording modalities to enhance robustness, and suggests wider clinical deployment would follow.