EEGChaT: A Transformer-Based Modular Channel Selector for SEEG Analysis

TL;DR

EEGChaT addresses the challenge of high-dimensional SEEG data by introducing a Transformer-based channel selector that uses Channel Aggregation Tokens (CATs) and Attention Rollout to produce interpretable, task-relevant channel weights. By downsampling inputs to 250 Hz and integrating CAT-attention with a downstream classifier, EEGChaT yields up to approximately 17% absolute gains in decoding accuracy on the DuIN dataset while providing channel-weighted interpretability aligned with manually selected channels. The work demonstrates consistent improvements across baseline classifiers (e.g., EEGConformer and Du-IN Encoder) and provides ablations showing the importance of Attention Rollout, masking, and CAT configuration, along with cross-subject considerations and limitations. This approach offers a generalizable, interpretable solution for high-dimensional SEEG analysis with potential broader impact on invasive EEG research and BCI applications.

Abstract

Analyzing stereoelectroencephalography (SEEG) signals is critical for brain-computer interface (BCI) applications and neuroscience research, yet poses significant challenges due to the large number of input channels and their heterogeneous relevance. Traditional channel selection methods struggle to scale or provide meaningful interpretability for SEEG data. In this work, we propose EEGChaT, a novel Transformer-based channel selection module designed to automatically identify the most task-relevant channels in SEEG recordings. EEGChaT introduces Channel Aggregation Tokens (CATs) to aggregate information across channels, and leverages an improved Attention Rollout technique to compute interpretable, quantitative channel importance scores. We evaluate EEGChaT on the DuIN dataset, demonstrating that integrating EEGChaT with existing classification models consistently improves decoding accuracy, achieving up to 17\% absolute gains. Furthermore, the channel weights produced by EEGChaT show substantial overlap with manually selected channels, supporting the interpretability of the approach. Our results suggest that EEGChaT is an effective and generalizable solution for channel selection in high-dimensional SEEG analysis, offering both enhanced performance and insights into neural signal relevance.

Figures (5)

• Figure 1: The framework of EEGChaT. (a) Overview of our methods. The input data to EEGChaT is downsampled to 250Hz to reduce the complexity of channel feature extraction for the model. (b) EEGChaT and downstream classifier integration diagram. (c) Default attention mask strategy in EEGChaT's multi-head self-attention. The gray areas in the figure indicate regions where attention score computation is not performed.
• Figure 2: Deriving channel weights from attention matrices. (a)Tracing back attention scores from raw channels to the CAT using Attention Rollout across multiple layers. (b) Deriving the weight tensor of raw channels via the Attention Rollout.
• Figure 3: The averaged Attention Rollout of Subject 1 on the test set.
• Figure 4: Different attention matrix masking strategies.
• Figure 5: The overall training pipeline combining cross-subject pretraining and downstream task fine-tuning.