Multi-class motor imagery EEG classification method with high accuracy and low individual differences based on hybrid...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Current motor imagery (MI) classification methods for electroencephalography (EEG) signals are complex, time-consuming, and fail to account for individual differences.
• Improving algorithm performance is crucial for increased classes and subject diversity in EEG-based applications.

Purpose of the Study:

• To develop an end-to-end deep learning method for classifying four-class MI tasks.
• To enhance recognition rates and balance classification accuracy across diverse subjects.
• To address limitations of traditional EEG signal processing in brain-computer interface (BCI) systems.

Main Methods:

• Proposed a novel one-dimensional input data representation to increase samples and mitigate channel correlation effects.
• Designed a cascade network combining convolutional neural networks (CNN) and gated recurrent units (GRU) for automated time-frequency feature learning.
• Implemented a subject-dependent neural network training approach to capture individual EEG patterns.

Main Results:

• Achieved high classification accuracies of 99.40% on the BCI Competition 2a dataset and 92.56% on a collected dataset.
• Demonstrated superior performance compared to advanced methods and baseline models.
• Reported low standard deviations of 0.34% and 1.35%, indicating robust and consistent results.

Conclusions:

• The proposed deep learning method significantly improves multi-class MI classification accuracy.
• The approach effectively overcomes the impact of individual differences in EEG signals.
• This work advances the development of practical and adaptable brain-computer interface systems.