Improved Brain-Computer Interface Signal Recognition Algorithm Based on Few-Channel Motor Imagery.

Area of Science:

• Neuroscience and Biomedical Engineering
• Signal Processing and Machine Learning

Background:

• Common Spatial Pattern (CSP) is effective for multichannel EEG but performs poorly with limited channels.
• Extracting robust motor imagery (MI) features from low-channel EEG remains a challenge for brain-computer interface (BCI) development.

Purpose of the Study:

• To develop a novel combined feature extraction method for MI-EEG signals using fewer channels.
• To improve the performance of BCI systems with limited EEG channel data.

Main Methods:

• Decomposition of band-pass filtered EEG into time-frequency components using Wavelet Packet Transform, Fast Ensemble Empirical Mode Decomposition, and Local Mean Decomposition.
• Selection of relevant components based on MI frequency characteristics or correlation coefficients.
• Phase Space Reconstruction (PSR), calculation of the maximum Lyapunov index, feature reconstruction, CSP projection mapping, Support Vector Machine (SVM) probability output, and Dempster-Shafer evidence theory fusion for classification.

Main Results:

• Achieved 95.71% accuracy on BCI Competition II dataset III (left- and right-hand MI), outperforming existing methods by 2.88%.
• Attained an average accuracy of 86.60% on BCI Competition IV dataset IIb, exceeding existing methods by 2.3%.
• Demonstrated the effectiveness of the proposed method for MI-BCI systems with fewer channels.

Conclusions:

• The proposed combined feature extraction method significantly enhances MI classification accuracy from low-channel EEG.
• This approach offers a viable solution for developing practical and efficient MI-BCI systems with reduced hardware requirements.