Towards Optimising EEG Decoding using Post-hoc Explanations and Domain Knowledge.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Decoding Electoencephalography (EEG) signals for motor imagery is essential for Brain-Computer Interface (BCI) performance.
• Increasing complexity of data-driven models challenges interpretability and trust in BCI systems.
• Artefacts and low signal-to-noise ratio in EEG necessitate transparent and reliable BCI models.

Purpose of the Study:

• To investigate the adequacy of accuracy metrics for BCI performance evaluation.
• To propose and validate the use of post-hoc explanations for interpreting BCI model outcomes.
• To assess the integration of domain-specific knowledge with Explainable AI (XAI) for neurophysiological validation.

Main Methods:

• Application of the GradCAM post-hoc explanation technique to an EEG motor movement/imagery dataset.
• Comparison of model performance and feature relevance between models trained on all EEG channels versus relevant channels.
• Validation of model-derived features against established neurophysiological facts.

Main Results:

• A model using all EEG channels achieved 72.60% accuracy; a model using relevant channels showed a statistically insignificant 1.75% decrease.
• Significant differences in relevant features were identified between the two models, despite comparable accuracy.
• Neurophysiological validation revealed discrepancies not captured by accuracy metrics alone.

Conclusions:

• Accuracy metrics are insufficient for ensuring the performance and acceptability of BCI systems.
• Integrating domain-specific knowledge with XAI techniques provides a robust method for validating the neurophysiological basis of BCI models.
• Neurophysiological validation is critical for developing dependable and transparent BCIs, mitigating risks associated with over-reliance on performance metrics.