A Domain Generalization Method for EEG Based on Domain-Invariant Feature and Data Augmentation

Area of Science:

• Neuroscience and Biomedical Engineering
• Machine Learning and Artificial Intelligence

Background:

• Brain-computer interface (BCI) technology demonstrates significant progress and practical applications.
• Domain bias in cross-domain BCI applications, especially with limited target data, poses a major challenge.
• Electroencephalography (EEG) signal limitations, including noise sensitivity and nonstationarity, complicate BCI generalization.

Purpose of the Study:

• To address the domain bias issue in BCI technology for improved cross-domain generalization.
• To develop a robust method for processing nonstationary EEG signals in data-scarce scenarios.
• To enhance the stability and performance of BCI models across different datasets.

Main Methods:

• Proposed a hybrid approach integrating domain-invariant feature learning and data enhancement strategies.
• Introduced a 'fixed' structure enhancement method to decouple domain-invariant features.
• Optimized cross-domain feature extraction and reduced noise effects in EEG data.

Main Results:

• The proposed hybrid model significantly outperforms existing state-of-the-art methods.
• Demonstrated superior performance across multiple publicly available EEG datasets.
• Effectively reduced the impact of noise and improved feature extraction for cross-domain generalization.

Conclusions:

• The hybrid approach offers a novel and effective solution to the domain bias problem in BCI.
• The method enhances the generalization capability of BCI systems, particularly for EEG signals.
• This work contributes to more reliable and practical BCI applications in diverse real-world scenarios.