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.
Computer Methods and Programs in Biomedicine
|August 20, 2025
Summary
This study introduces a novel, interpretable deep learning framework for analyzing electroencephalographic (EEG) brain oscillations. The approach automatically identifies key frequency, spatial, and temporal patterns, enhancing brain function analysis.
Area of Science:
- Neuroscience
- Computational Neuroscience
- Signal Processing
Background:
- Analysis of neural oscillations using electroencephalography (EEG) is crucial for understanding brain function.
- Traditional EEG analysis methods face limitations in quality, reproducibility, and reliability due to manual processing choices and dimensionality reduction.
- A need exists for automated, end-to-end approaches to analyze EEG oscillations with minimal prior assumptions.
Purpose of the Study:
- To develop a novel, fully-interpretable deep learning framework for characterizing neural oscillations in EEG signals.
- To automatically identify optimal processing strategies for minimally pre-processed EEG data.
- To reveal salient signatures of brain oscillations across frequency, spatial, and temporal domains.
Main Methods:
- A novel framework utilizing a fully-interpretable convolutional neural network (CNN) was designed.
- The CNN learns a bank of bandpass filters for minimally pre-processed EEG signals.
- Frequency-specific spatial and temporal filtering identifies salient features, processed to reveal meaningful EEG signatures.
Main Results:
- The framework was successfully applied to real-world data from motor imagery tasks.
- The approach identified known modulations of brain oscillations during motor imagery, consistent with classic analyses.
- The alpha band (8-13 Hz), motor-related electrodes, and time samples near the cue were identified as most significant.
Conclusions:
- The proposed framework offers an automatic, optimal, and end-to-end method for characterizing brain oscillations.
- This approach can significantly enhance the understanding of brain functions in both healthy individuals and patients.
- It holds potential for tracking neuropathological alterations through EEG signal analysis.
