xEEGNet: towards explainable AI in EEG dementia classification | JoVE Visualize

Area of Science:

Computational Neuroscience
Artificial Intelligence in Medicine
Biomedical Signal Processing

Background:

Deep learning models for electroencephalography (EEG) analysis often function as 'black boxes', limiting clinical interpretability.
Existing models can be parameter-heavy, leading to overfitting and reduced generalizability in neurological condition classification.
There is a need for compact, interpretable, and robust neural networks for analyzing spectral alterations in EEG data.

Purpose of the Study:

To introduce xEEGNet, a novel, compact, and fully interpretable neural network for EEG data analysis.
To demonstrate the model's efficacy in classifying dementia conditions (Alzheimer's and frontotemporal dementia) versus controls.
To showcase the broad applicability of xEEGNet to other neurological conditions characterized by spectral changes.

Main Methods:

Developed xEEGNet by modifying the ShallowNet architecture to enhance transparency and reduce parameters.
Employed a nested-leave-n-subjects out cross-validation strategy for unbiased performance estimation.
Analyzed learned kernels, weights, and embedded EEG representations to assess clinical significance and explain performance variability.

Main Results:

xEEGNet utilizes only 168 parameters, a 200-fold reduction compared to ShallowNet, while maintaining interpretability.
The model achieves comparable median performance to ShallowNet, with a mere -1.5% difference, and exhibits reduced performance variability.
Higher classification accuracy correlated with greater separability of embedded EEG representations between control and Alzheimer's groups.

Conclusions:

xEEGNet offers a compact and interpretable alternative to larger deep learning models for EEG-based pathology classification.
The model's ability to filter specific EEG bands and learn band-specific topographies underscores its clinical interpretability.
This study validates the effectiveness of smaller, transparent neural network architectures in identifying neurological conditions using EEG data.