Classification of multi-class motor imagery with a novel hierarchical SVM algorithm for brain-computer interfaces.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Pattern classification is vital for brain-computer interface (BCI) development.
• Electroencephalography (EEG) signals are commonly used for BCI applications, but accurate classification remains challenging.

Purpose of the Study:

• To propose and validate a novel hierarchical support vector machine (HSVM) algorithm for a four-class motor imagery classification task using EEG signals.
• To enhance the accuracy and effectiveness of EEG-based BCI systems.

Main Methods:

• Wavelet packet transform was used for raw EEG signal decomposition.
• Effective frequency sub-bands were grouped and reconstructed.
• Feature vectors were extracted using one versus the rest common spatial patterns (OVR-CSP) and one versus one common spatial patterns (OVO-CSP).
• A two-layer HSVM classifier was designed, utilizing OVO classifiers in the first layer and OVR in the second.

Main Results:

• The proposed HSVM algorithm achieved an average classification accuracy of 67.5% ± 17.7% in the first layer and 60.3% ± 14.7% in the second layer.
• The overall average classification accuracy was 64.4% ± 16.7% using the BCI Competition IV-II-a dataset.
• Fivefold cross-validation confirmed the method's effectiveness.

Conclusions:

• The proposed hierarchical support vector machine (HSVM) algorithm is effective for four-class motor imagery classification in EEG-based BCI.
• This approach offers a promising solution for improving the performance of BCI applications.