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

TL;DR

The paper addresses interpretability in EEG-based motor-imagery BCIs by integrating post-hoc explanations with domain knowledge. It applies Grad-CAM to an EEG Conformer model to produce spatial and temporal relevance maps and compares full-channel, GradCAM-channel, and motor-imagery-channel configurations. Results show a full 64-channel model achieves $72.60\%$ accuracy, while GradCAM-relevant channels lower accuracy by $5.97\%$ (p=$0.002$) and motor-imagery channels lower by $1.75\%$ (not significant), with neurophysiological explanations revealing differing feature reliance. The work argues that neurophysiological validation is essential for trustworthy BCIs and calls for cross-dataset benchmarks of explanations across models and datasets.

Abstract

Decoding EEG during motor imagery is pivotal for the Brain-Computer Interface (BCI) system, influencing its overall performance significantly. As end-to-end data-driven learning methods advance, the challenge lies in balancing model complexity with the need for human interpretability and trust. Despite strides in EEG-based BCIs, challenges like artefacts and low signal-to-noise ratio emphasise the ongoing importance of model transparency. This work proposes using post-hoc explanations to interpret model outcomes and validate them against domain knowledge. Leveraging the GradCAM post-hoc explanation technique on the motor imagery dataset, this work demonstrates that relying solely on accuracy metrics may be inadequate to ensure BCI performance and acceptability. A model trained using all EEG channels of the dataset achieves 72.60% accuracy, while a model trained with motor-imagery/movement-relevant channel data has a statistically insignificant decrease of 1.75%. However, the relevant features for both are very different based on neurophysiological facts. This work demonstrates that integrating domain-specific knowledge with XAI techniques emerges as a promising paradigm for validating the neurophysiological basis of model outcomes in BCIs. Our results reveal the significance of neurophysiological validation in evaluating BCI performance, highlighting the potential risks of exclusively relying on performance metrics when selecting models for dependable and transparent BCIs.

Figures (4)

• Figure 1: Overview of design space for XAI applied to Motor Imagery and Execution task.
• Figure 2: Montage image highlighting the significant channels identified by GradCAM when model is trained with all 64 EEG channels (Left) and 21 EEG channels (Right) relevant for Motor Imagery and movement.
• Figure 3: Raw activations (Left column) during the task compared with Class Activation Topography(CAT) for both left(Centre column) and right (Right column) predictions for Subject ID 42 using top 17 relevant channels via GradCAM(Top row) and 21 MI-relevant channels(Bottom row).
• Figure 4: (Top) Time-frequency plots with topo maps in the Mu-Beta band depicting relevant frequencies during right fist movements, marked by GradCAM when trained with the top 17 channels. For Subject ID 42, model accuracy dropped to $42.50\%$ when trained with the top 17 channels. (Bottom) In contrast, the bottom plots explain the model outcomes, where the model achieved $80.00\%$ accuracy for right fist movements when trained with 21 MI-relevant channels. This comparison highlights the difference in relevant timeframes affecting model predictions.