Transfer learning for target user in motor imagery EEG recognition

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Accurate electroencephalography (EEG) identification is vital for brain-computer interface (BCI) systems, especially for two-class motor imagery tasks.
• Existing BCI methods require improvement in efficiency and accuracy for multi-user applications.

Purpose of the Study:

• To develop a lightweight and effective method for target user EEG identification in motor imagery BCI.
• To enhance the accuracy and practicality of BCI systems for diverse users.

Main Methods:

• Utilized wavelet packet transform (WPT) for EEG decomposition and feature extraction (variance, energy mean).
• Applied transfer learning (TL) with a TL classifier for robust classification.
• Preprocessed EEG data using common average reference.

Main Results:

• Achieved an average accuracy of 91.8% on the BCI Competition III dataset IVa.
• Outperformed the top two methods from the BCI Competition III.
• Demonstrated the effectiveness of combining WPT and TL for motor imagery recognition.

Conclusions:

• The proposed lightweight method combining WPT and TL is simple, effective, and practical for multi-user motor imagery recognition.
• This approach promotes robust BCI operations and has significant potential for real-world applications.
• The method offers a promising solution for advancing BCI technology.