Subject Representation Learning from EEG using Graph Convolutional Variational Autoencoders

TL;DR

This work introduces GC-VASE, a graph convolutional variational autoencoder for EEG subject representation that splits the latent space into a subject-specific component $z^{S}$ and a residual component $z^{T}$, and leverages contrastive learning to disentangle subject variability from task content. The model combines GCNN-based encoders/decoders with transformers and an XGB classifier, and uses lightweight attention adapters to fine-tune efficiently for unseen subjects. Empirical results on ERP-Core and SleepEDFx-20 show state-of-the-art subject identification (89.81% zero-shot on ERP-Core, 90.31% after adapters; 70.85% on SleepEDFx-20), with ablations highlighting the critical roles of GCNN layers and contrastive loss in disentangling subject-specific features. The approach promises practical impact for personalized EEG-based diagnostics and biometrics, offering low-cost domain adaptation and a path toward knowledge distillation from large EEG foundation models.

Abstract

We propose GC-VASE, a graph convolutional-based variational autoencoder that leverages contrastive learning for subject representation learning from EEG data. Our method successfully learns robust subject-specific latent representations using the split-latent space architecture tailored for subject identification. To enhance the model's adaptability to unseen subjects without extensive retraining, we introduce an attention-based adapter network for fine-tuning, which reduces the computational cost of adapting the model to new subjects. Our method significantly outperforms other deep learning approaches, achieving state-of-the-art results with a subject balanced accuracy of 89.81% on the ERP-Core dataset and 70.85% on the SleepEDFx-20 dataset. After subject adaptive fine-tuning using adapters and attention layers, GC-VASE further improves the subject balanced accuracy to 90.31% on ERP-Core. Additionally, we perform a detailed ablation study to highlight the impact of the key components of our method.

Figures (2)

• Figure 1: The proposed GC-VASE model incorporates a split latent space. The encoder splits the latent space into subject-specific space and residual latent space that are subsequently used for subject and task classification through an XGB classifier.
• Figure 2: t-SNE plots of split-latents encoded on the test set (unseen subjetcs), colored by their true labels. The plots, in order, depict: subject space colored by subject, subject space colored by task, task space colored by task, and task space colored by subject.