Interpretable Decoding of Frequency-Resolved Functional Connectivity

Whole-brain functional connectivity, derived from magnetoencephalography (MEG) data, delivers a concise depiction of distant neuronal communication, rendering it ideal for predicting potential biomarkers associated with brain disorders. This research introduces a deep learning architecture (FC-CNN) designed to forecast brain states based on frequency-resolved functional connectivity metrics obtained from resting-state MEG recordings.

The performance of FC-CNN is rigorously evaluated against traditional regression techniques utilizing both amplitude and phase-related functional connectivity measures within the well-documented age-prediction task involving the Cam-CAN cohort, comprising 576 participants. The findings reveal that FC-CNN surpasses conventional methods, and that, when compared to phase synchronization, amplitude envelope correlation yields superior prediction performance.

Furthermore, the study provides compelling quantitative evidence that the weights assigned to a trained deep learning model can facilitate neurological interpretation of the neural activity patterns that contribute to accurate predictions. The results of this investigation demonstrate that the presented approach effectively decodes brain states from MEG functional connectivity, suggesting its potential to uncover predictive biomarkers for various brain disorders.