Foundation models for EEG decoding: current progress and prospective research.

Area of Science:

• Neuroscience and Artificial Intelligence
• Brain-computer interfaces
• Machine learning for neuroimaging

Background:

• Traditional electroencephalography (EEG) decoding methods face limitations in task specificity and dataset dependency.
• Deep learning is increasingly applied to EEG decoding, but supervised approaches restrict model generalization.
• EEG foundation models (EEG FMs), inspired by large language models, present a unified paradigm for EEG decoding.

Purpose of the Study:

• To review representative studies on EEG foundation models (EEG FMs).
• To extract trends in EEG FM development and application.
• To provide recommendations for future EEG FM research.

Main Methods:

• Comprehensive analysis of recent advances in EEG FMs.
• Focus on downstream tasks, benchmark datasets, model architectures, and pre-training techniques.
• Systematic comparison of core FM components, performance, and generalizability.

Main Results:

• EEG FMs are pre-trained on large-scale datasets (up to 14,987 subjects, 27,062 hours).
• Mask-based reconstruction and transformer-based architectures are common pre-training strategies.
• EEG FMs show potential in seizure detection but have limited performance in complex tasks like motor imagery decoding.

Conclusions:

• EEG FMs represent a significant advancement in EEG decoding, offering improved generalization.
• Current limitations exist, particularly in complex cognitive tasks.
• Further research is needed to enhance EEG FM capabilities and address current challenges.