FourierMask: Explain EEG-Based End-to-End Deep Learning Models in the Frequency Domain

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning models for electroencephalography (EEG) analysis are increasingly used, but their decision-making processes remain opaque.
• Understanding how these models interpret time-series EEG signals is crucial for reliable applications.
• The frequency domain offers advantages for EEG analysis due to its correlation with cognitive states and ability to capture long-range dependencies, yet it's underexplored for model explanation.

Purpose of the Study:

• To introduce FourierMask, a novel mask perturbation framework for explaining EEG-based end-to-end deep learning models in the frequency domain.
• To enhance the transparency and interpretability of complex EEG models.
• To identify key frequency biomarkers driving model predictions.

Main Methods:

• Developed FourierMask, a framework utilizing Fourier-based domain transformation for direct spectral component manipulation.
• Implemented a learnable mask mechanism to jointly model spectral-spatial couplings for EEG explanation.
• Employed a perturbation generator with a target alignment loss and cluster-aware regularization to ensure natural perturbations and minimize distribution shift.

Main Results:

• FourierMask achieved significant accuracy drop gaps: 36.0% at the group-level and 17.8% at the instance-level, outperforming LIME and easyPEASI.
• Demonstrated superior performance compared to existing methods in quantifying the importance of frequency components.
• Validated effectiveness across multiple EEG deep learning architectures (EEGNet, TSCeption, DeepConvNet).

Conclusions:

• FourierMask provides a model-agnostic, plug-and-play solution for enhancing the transparency of EEG deep learning models.
• Effectively links model predictions to specific frequency biomarkers, offering insights into cognitive states.
• Presents potential applications in advancing neuromedicine and brain-computer interfaces through interpretable EEG analysis.