EAD: An EEG Adapter for Automated Classification

TL;DR

The paper introduces EEG Adapter (EAD), a modular adapter layered on the LaBraM EEG foundation model, to produce robust, task-tailored embeddings for diverse EEG classification tasks across different acquisition devices. Through manual montage alignment and automatic channel distillation, EAD achieves state-of-the-art results on EEG-ImageNet (99.33% accuracy) and BrainLat (92.31% at the subject level), including zero-shot demonstrations that highlight cross-task and cross-subject generalizability. The approach reduces preprocessing requirements and supports end-to-end training, enabling a unified pipeline for stimulus decoding and resting-state classification. This work advances cross-device EEG embeddings with practical implications for visual brain decoding and clinical EEG analysis.

Abstract

While electroencephalography (EEG) has been a popular modality for neural decoding, it often involves task specific acquisition of the EEG data. This poses challenges for the development of a unified pipeline to learn embeddings for various EEG signal classification, which is often involved in various decoding tasks. Traditionally, EEG classification involves the step of signal preprocessing and the use of deep learning techniques, which are highly dependent on the number of EEG channels in each sample. However, the same pipeline cannot be applied even if the EEG data is collected for the same experiment but with different acquisition devices. This necessitates the development of a framework for learning EEG embeddings, which could be highly beneficial for tasks involving multiple EEG samples for the same task but with varying numbers of EEG channels. In this work, we propose EEG Adapter (EAD), a flexible framework compatible with any signal acquisition device. More specifically, we leverage a recent EEG foundational model with significant adaptations to learn robust representations from the EEG data for the classification task. We evaluate EAD on two publicly available datasets achieving state-of-the-art accuracies 99.33% and 92.31% on EEG-ImageNet and BrainLat respectively. This illustrates the effectiveness of the proposed framework across diverse EEG datasets containing two different perception tasks: stimulus and resting-state EEG signals. We also perform zero-shot EEG classification on EEG-ImageNet task to demonstrate the generalization capability of the proposed approach.

Figures (4)

• Figure 1: The figure demonstrates our approach to utilize the EAD for classification tasks. We utilize two setups resting state and image decoding multichannel EEG dataset (MCED) to showcase the effectiveness of our approach.
• Figure 2: This image portrays a magnified view of the EAD architecture. $X$ refers to the input EEG signal with $E$ channels and $T$ time steps. $T^{'}$ is dependent upon the convolution kernel, $I_{Ch}$ is the number of input channels and $I_{Tsteps}$ is the number of time steps. $L_{CE}$ is categorical cross entropy loss.
• Figure 3: On the left, the tSNE plot is for the manual channel selection without mixing data. On the right, the tSNE plot is for the automatic distillation using EAD. Each color in both plots denotes a different class. This provides evidence that using EAD, effective learning is happening for the data.
• Figure 4: tSNE plot for BrainLat dataset.