A deep domain adaptation framework with correlation alignment for EEG-based motor imagery classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) face challenges with electroencephalography (EEG) data due to time-consuming acquisition and annotation.
• Conventional classification methods struggle with motor imagery tasks when EEG data varies across subjects or time periods (cross-domain).
• This distribution divergence significantly reduces classification accuracy in BCIs.

Purpose of the Study:

• To propose a novel deep domain adaptation framework with correlation alignment (DDAF-CORAL) for robust motor imagery classification.
• To address the distribution divergence issue in EEG data across different domains (subjects, sessions).
• To enhance the accuracy and reliability of BCIs for motor imagery tasks.

Main Methods:

• A two-stage deep framework (DDAF-CORAL) was developed to extract deep features from raw EEG data.
• Correlation alignment (CORAL) was employed to minimize distribution divergence by aligning feature distribution covariances.
• Simultaneous optimization of classification and adaptation losses was performed to achieve discriminative classification and low feature divergence.

Main Results:

• The DDAF-CORAL method effectively reduced distribution divergence between source and target EEG data across three experimental datasets.
• Demonstrated superior performance in two-class motor imagery classification tasks.
• Achieved high accuracy rates: 92.9% within-session, 0.761 kappa cross-session, and 83.3% cross-subject.

Conclusions:

• The proposed DDAF-CORAL framework significantly improves motor imagery classification accuracy in cross-domain BCI applications.
• This method offers a viable solution for overcoming data variability challenges in EEG-based BCIs.
• The findings highlight the potential of deep domain adaptation for enhancing BCI performance and reducing reliance on extensive labeled data.