This is the official repository for the research paper
xEEGNet: Towards Explainable AI in EEG Dementia Classification
Published in IOP, Journal of Neural Engineering xEEGNet.
In this work, we presents xEEGNet, a compact, explainable neural network for EEG data analysis. It is fully interpretable and mitigates overfitting by substantially reducing parameters. xEEGNet employs only 168 parameters, 200 times fewer than ShallowNet, yet maintains interpretability, resists overfitting, achieves similar median performance (-1.5%), and decreases performance variability across data splits. Its ability to filter specific EEG bands, learn band-specific topographies, and select the relevant ones for disease classification demonstrates its interpretability. While large deep learning models are usually emphasized for performance, this study shows that smaller architectures like xEEGNet can perform equally well in pathology classification using EEG data.
Scripts used to generate the results presented in the paper are available in this repository. In AllFnc folder, the Python file eegvislib.py was developed with the intent to be used as library for insepcting kernels and feature maps in DL applications, with a focus in EEG data. The Jupyter Notebook ModelVisualization.ipynb shows a proof of concept of the usefulness. In details, the library can help researchers to inspect both the weights and the feature maps, learned by the neural network.
If you find codes and results useful for your research, please concider citing our work. It would help us to continue our research.
Contributors:
- M.Sc. Andrea Zanola
- M.Sc. Louis Fabrice Tshimanga
- Eng. Federico Del Pup
- Prof. Marco Baiesi
- Prof. Manfredo Atzori
The code is released under the MIT License.