Related Experiment Video
Updated: Sep 7, 2025
08:51
Statistical Modelling of Cortical Connectivity Using Non-invasive Electroencephalograms
Published on: November 1, 2019
5.7K
A Systematic Approach for Explaining Time and Frequency Features Extracted by Convolutional Neural Networks From Raw
Charles A Ellis1,2, Robyn L Miller2,3, Vince D Calhoun1,2,3
1Wallace H. Coulter Department of Biomedical Engineering, Georgia Institute of Technology and Emory University, Atlanta, GA, United States.
Frontiers in Neuroinformatics
|June 17, 2022
Summary
This study introduces a novel method for visualizing convolutional neural networks (CNNs) in electroencephalography (EEG) analysis. The approach enhances understanding of how CNNs interpret spectral and waveform features for sleep stage classification.
Area of Science:
- Neuroscience
- Machine Learning
- Signal Processing
Background:
- Convolutional Neural Networks (CNNs) are increasingly used for analyzing raw resting-state electroencephalography (EEG) data.
- Explainability of CNNs in EEG analysis remains a challenge compared to traditional methods.
- Existing methods focus on spectral features, neglecting the importance of multispectral waveforms.
Purpose of the Study:
- To develop a novel, interpretable CNN architecture for raw EEG analysis.
- To systematically evaluate the importance of spectral and waveform features learned by CNNs.
- To enhance the understanding of CNN decision-making processes in neurophysiological data.
Main Methods:
- Developed a model visualization-based approach adapting CNN architecture for increased interpretability.
- Implemented novel explainability methods to evaluate feature importance.
- Assessed the contribution of spectrally distinct filter clusters, waveforms, and spectra to model predictions.
Main Results:
- The proposed approach systematically evaluates both waveform and spectral feature importance in CNNs trained on resting-state EEG.
- Explainability results for automated sleep stage classification align well with clinical guidelines.
- Identified specific contributions of identified waveforms and spectra to cluster importance.
Conclusions:
- The novel approach provides a powerful tool for interpreting CNNs in EEG analysis.
- It offers a systematic way to understand the role of both spectral and waveform features.
- This method advances the explainability of deep learning models in neuroscience applications.
