Quantifying Spatial Domain Explanations in BCI using Earth Mover's Distance

TL;DR

The paper tackles the challenge of interpretable MI-based BCIs by linking spatial explanation maps to neuroscience knowledge through an Earth Mover's Distance ($EMD$) metric. It benchmarks three model families—MDM (Riemannian geometry on SPD covariances), EEGNet, and EEG Conformer—using GradCAM-derived relevance maps and evaluates alignment with domain knowledge on a large 64-channel EEG dataset ($N$-channel setup) for motor imagery tasks. The key contributions are (1) an $EMD$-based framework to quantify spatial explanations against neuroscience priors, and (2) a comprehensive performance and explanation benchmark across 109 participants, revealing that MI-relevant channels beat data-driven relevance and that Riemannian methods provide robust performance signals. This work advances explainable BCIs by formalizing spatial explanations and illustrating how domain knowledge can guide more reliable, interpretable models in real-world neurotechnology contexts.

Abstract

Brain-computer interface (BCI) systems facilitate unique communication between humans and computers, benefiting severely disabled individuals. Despite decades of research, BCIs are not fully integrated into clinical and commercial settings. It's crucial to assess and explain BCI performance, offering clear explanations for potential users to avoid frustration when it doesn't work as expected. This work investigates the efficacy of different deep learning and Riemannian geometry-based classification models in the context of motor imagery (MI) based BCI using electroencephalography (EEG). We then propose an optimal transport theory-based approach using earth mover's distance (EMD) to quantify the comparison of the feature relevance map with the domain knowledge of neuroscience. For this, we utilized explainable AI (XAI) techniques for generating feature relevance in the spatial domain to identify important channels for model outcomes. Three state-of-the-art models are implemented - 1) Riemannian geometry-based classifier, 2) EEGNet, and 3) EEG Conformer, and the observed trend in the model's accuracy across different architectures on the dataset correlates with the proposed feature relevance metrics. The models with diverse architectures perform significantly better when trained on channels relevant to motor imagery than data-driven channel selection. This work focuses attention on the necessity for interpretability and incorporating metrics beyond accuracy, underscores the value of combining domain knowledge and quantifying model interpretations with data-driven approaches in creating reliable and robust Brain-Computer Interfaces (BCIs).

Figures (3)

• Figure 1: Montage images highlighting the significant channels among the 64 channels, identified by a) using feature relevance scores based on Riemannian distance on covariance matrix, b) using GradCAM with EEGConformer model, and c) using GradCAM with EEGNet model.
• Figure 2: 2D spatial representation of the significant channels among the 64 channels used for the comparison based on EMD. a) Baseline representation using 21 channels proximal to motor cortical regions, b) Relevance from MDM classifier using Riemannian distance is closest to baseline, b) followed by EEGConformer, and c) EEGNet model, when trained on top 14 participants.
• Figure 3: 2D spatial representation of the significant channels with weighted relevance among the 64 channels used for the comparison based on EMD. a) Baseline representation with equal importance to 21 channels proximal to motor cortical regions, b) Relevance from MDM classifier using Riemannian distance is closest to baseline, b) followed by EEGConformer, and c) EEGNet model, when trained on top 14 participants.