EEG classification across sessions and across subjects through transfer learning in motor imagery-based brain-machine...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Transfer learning is crucial for adapting models to distribution mismatches across sessions or subjects in electroencephalography (EEG) data.
• Classifying motor imagery EEG signals is vital for brain-computer interfaces (BCIs).

Purpose of the Study:

• To propose a novel transfer learning algorithm for classifying motor imagery EEG data.
• To improve classification accuracy by effectively handling cross-session and cross-subject variations.

Main Methods:

• Extracting shared features (mean and variance of model parameters) from the power spectrum of motor imagery EEG data.
• Updating shared features using relevant datasets identified by Euclidean distance.
• Jointly utilizing shared and subject/session-specific features to generate a new classification model.

Main Results:

• The proposed transfer learning algorithm demonstrated significantly higher classification accuracy compared to traditional machine learning algorithms (PSD and CSP).
• Paired t-tests showed significant differences in classification results (p < 3e-9).
• Achieved 85.7% ± 5.4% accuracy on BCI competition IV dataset 2a, outperforming existing methods.

Conclusions:

• The developed transfer learning algorithm is effective for classifying motor imagery EEG signals across sessions and subjects.
• The algorithm shows superior performance compared to traditional machine learning techniques.
• The approach holds promise for advancing brain-computer interface (BCI) applications.