DMAE-EEG: A Pretraining Framework for EEG Spatiotemporal Representation Learning.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Electroencephalography (EEG) is vital in neuroscience and clinical settings but faces challenges with data uniformity, noise, and labeling.
• Developing methods to extract generalizable spatiotemporal representations from unlabeled EEG is crucial for advancing its applications.

Purpose of the Study:

• To introduce DMAE-EEG, a denoising masked autoencoder framework for mining generalizable spatiotemporal representations from large unlabeled EEG datasets.
• To address limitations in EEG data quality and labeling for improved downstream task performance.

Main Methods:

• Proposed a brain region topological heterogeneity (BRTH) method for nonuniform EEG data partitioning.
• Designed a denoised pseudo-label generator (DPLG) using a denoising reconstruction pretext task to learn robust representations.
• Utilized an asymmetric autoencoder with self-attention as the backbone to capture long-range spatiotemporal dependencies across 14 public EEG datasets.

Main Results:

• DMAE-EEG significantly improved EEG signal quality, reducing normalized mean squared error (nMSE) by 27.78%-50.00% compared to statistical methods.
• Achieved relative improvements of 2.71%-6.14% in motion intention recognition balanced accuracy, outperforming state-of-the-art methods.
• Demonstrated enhanced knowledge transferability across sessions, subjects, and tasks.

Conclusions:

• DMAE-EEG effectively captures generalizable spatiotemporal representations from massive unlabeled EEG data.
• The framework advances EEG-aided diagnosis, brain-computer interfaces, and clinical practice by improving data utility and model generalizability.