An empirical comparison of deep learning explainability approaches for EEG using simulated ground truth

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning (DL) enhances electroencephalography (EEG) decoding but lacks interpretability.
• Understanding DL model explanations is vital for reliable EEG-based applications.

Approach:

• Developed a simulation framework to evaluate 12 back-propagation-based visualization methods for EEG.
• Assessed method robustness and sensitivity against ground truth features.
• Tested reliability after randomizing model weights and labels.

Key Points:

• Saliency methods, common in EEG, lack class and model specificity.
• DeepLift demonstrated consistent accuracy and robustness in detecting temporal, spatial, and spectral features.
• Several methods showed reliability issues when tested against randomized data.

Conclusions:

• Provides a review of model explanation methods for DL neural decoders in EEG.
• Offers recommendations for selecting and interpreting explanation methods.
• Highlights the importance of robust methods like DeepLift for EEG analysis.