Unveiling multi-domain signatures of EEG oscillations using a fully-interpretable convolutional neural network
1Department of Electrical, Electronic and Information Engineering "Guglielmo Marconi" (DEI), University of Bologna, Cesena, Forlì-Cesena, Italy.
Background And Objective:
Neural oscillations are widely investigated to characterize brain functions. However, their analysis in the frequency, spatial, and temporal domains from electroencephalographic (EEG) signals (e.g., via event-related spectral perturbations) is affected by choices that limit the quality, reproducibility, and reliability of results. For example, different pre-processing and processing steps strongly affect the results, and the steps are often implemented with manual/semi-automatic algorithms. Moreover, due to the high dimensionality of the involved measures, generally a few frequency intervals/brain regions/time intervals of interest are selected exploiting a priori knowledge, and then analyzed. Therefore, it is desired an end-to-end approach that automatically learns the optimal strategy to process minimally pre-processed EEG to highlight the most relevant signatures of brain oscillations in the frequency, spatial, and temporal domains with minimal a priori assumptions.
Methods:
In this study, we design a novel framework for characterizing EEG oscillations in the frequency, spatial, and temporal domains, based on the features learned by a fully-interpretable convolutional neural network. The network learns a bank of bandpass filters to be applied to minimally pre-processed EEG. Then, frequency-specific spatial and temporal filtering allow the learning of the most salient spatial and time samples, separately for each frequency component. Finally, the framework processes the learned interpretable features to reveal meaningful EEG signatures.
Results:
We test our approach by applying it to real data recorded during motor imagery tasks. Our neural network-empowered approach reveals the modulations of brain oscillations known to occur during motor imagery, and match results obtained with classic analyses. Specifically, the alpha band (8-13 Hz) was the most important, together with the electrodes covering motor areas and the time samples closer to the cue indicating the action to imagine.
Conclusions:
The proposed framework enables the characterization of brain oscillations in an automatic, optimal and end-to-end way, and could be conveniently exploited for boosting our comprehension of brain functions in healthy participants and in patients, tracking their neuropathological alterations.
