EEGDM: Learning EEG Representation with Latent Diffusion Model

TL;DR

The paper tackles the limited ability of traditional self-supervised EEG methods to capture global temporal dynamics by introducing EEGDM, a latent diffusion framework. EEGDM conditions a diffusion denoiser on PCA-projected EEG latent space guided by an EEG encoder and channel augmentations, enabling both high-fidelity signal generation and robust representation learning. It demonstrates competitive downstream performance across diverse EEG tasks and provides extensive ablations and visualizations to validate the approach. The work highlights the practicality and generalizability of diffusion-based pretraining for EEG, offering a new pathway for cross-dataset EEG understanding.

Abstract

Recent advances in self-supervised learning for EEG representation have largely relied on masked reconstruction, where models are trained to recover randomly masked signal segments. While effective at modeling local dependencies, such objectives are inherently limited in capturing the global dynamics and long-range dependencies essential for characterizing neural activity. To address this limitation, we propose EEGDM, a novel self-supervised framework that leverages latent diffusion models to generate EEG signals as an objective. Unlike masked reconstruction, diffusion-based generation progressively denoises signals from noise to realism, compelling the model to capture holistic temporal patterns and cross-channel relationships. Specifically, EEGDM incorporates an EEG encoder that distills raw signals and their channel augmentations into a compact representation, acting as conditional information to guide the diffusion model for generating EEG signals. This design endows EEGDM with a compact latent space, which not only offers ample control over the generative process but also can be leveraged for downstream tasks. Experimental results show that EEGDM (1) reconstructs high-quality EEG signals, (2) learns robust representations, and (3) achieves competitive performance across diverse downstream tasks, thus exploring a new direction for self-supervised EEG representation learning.

Figures (5)

• Figure 1: The overall architecture of EEGDM. The original EEG data is projected onto a latent space using PCA. Noise is added in this latent space, and the denoised latent representation is predicted by learning the conditional latent diffusion model. The prediction is then projected back to the original space by inverse PCA to obtain the generated target EEG signal. The encoder transforms the source EEG signal (including original data and augmentations) into a concise EEG representation. It is combined with the noise timestep representation and then used as conditional information to modulate the DiT block, thus guiding the denoising of the latent representation.
• Figure 2: The pre-trained encoder is connected to a linear prediction head to perform downstream tasks.
• Figure 3: Performance comparison with different numbers of PCA components on TUEV and TUAB datasets.
• Figure 4: The t-SNE visualizations. Clear inter-class separation and intra-class compactness indicate that the learned representations are discriminative.
• Figure 5: Visualization of reconstructed EEG signals and their PSD.