AutoEncoder Filter Bank Common Spatial Patterns to Decode Motor Imagery From EEG.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Decoding imagined movements from electroencephalography (EEG) is crucial for brain-computer interfaces (BCIs).
• Existing methods like Filter Bank Common Spatial Patterns (FBCSP) have limitations in subject-specific performance and accuracy.
• The need for robust and accurate EEG decoding methods is paramount for practical BCI applications.

Purpose of the Study:

• Introduce a novel EEG decoding method, AutoEncoder-Filter Bank Common Spatial Patterns (AE-FBCSP).
• Enhance the performance of FBCSP through unsupervised feature projection into a latent space.
• Improve subject-specific classification accuracy for motor imagery tasks.

Main Methods:

• Extracted features from high-density EEG using FBCSP.
• Employed an unsupervised custom AutoEncoder (AE) to learn a compressed latent representation of FBCSP features.
• Trained a supervised feed-forward neural network classifier on latent features for decoding imagined movements.
• Utilized a public dataset from 109 subjects with various motor imagery tasks (right hand, left hand, both hands, both feet, and resting states).

Main Results:

• AE-FBCSP significantly outperformed standard FBCSP in statistical significance (p > 0.05).
• Achieved a subject-specific average accuracy of 89.09% in 3-way classification (right hand vs. left hand vs. resting).
• Demonstrated superior subject-specific classification performance compared to other literature methods across 2-way, 4-way, and 5-way tasks.

Conclusions:

• AE-FBCSP offers a significant advancement in decoding imagined movements from EEG.
• The method effectively enhances classification accuracy and increases the number of subjects achieving high performance.
• AE-FBCSP shows strong potential for practical and effective BCI system development.