Feature alignment and enhancement network with guided tuning for non-stationary EEG classification.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) signal variability due to external factors and individual differences poses challenges for motor imagery (MI) classification in brain-computer interfaces (BCIs).
• Existing domain alignment techniques often fail to fully leverage source and target domain information, leading to negative transfer issues and suboptimal performance.

Purpose of the Study:

• To propose a Feature Alignment and Enhancement Framework (FAEF) designed to overcome limitations in cross-domain MI-EEG classification.
• To improve the adaptability and accuracy of BCI models in diverse BCI scenarios, including cross-session and cross-subject classifications.

Main Methods:

• Aligning covariance matrices of source and target domains to harmonize spatial distributions.
• Utilizing a conditional domain adversarial network to reduce distribution discrepancies and enhance cross-domain representation discriminability.
• Implementing an EEG feature-based guided tuning method with centroid features for dynamic optimization of input representations.

Main Results:

• Achieved cross-session classification accuracy of 76.89% and cross-subject accuracy of 57.91% on the BCI Competition IV-2a dataset.
• Obtained classification accuracies of 84.61% and 82.78% on the BCI Competition IV-2b and High Gamma Datasets, respectively.
• Demonstrated robust performance with additional accuracies of 84.09% and 70.81% on specific datasets.

Conclusions:

• The proposed framework effectively mitigates cross-domain variations in EEG signals.
• This approach offers a reliable solution for enhancing cross-session and cross-subject MI-EEG classification in BCIs.
• The FAEF framework shows significant potential for advancing the practical application of BCI technology.