Subject adaptation convolutional neural network for EEG-based motor imagery classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep transfer learning is crucial for non-stationary electroencephalogram (EEG) data in motor imagery (MI) classification.
• Existing deep learning methods struggle with limited accuracy due to ineffective temporal and spatial feature extraction.

Purpose of the Study:

• To propose a novel end-to-end deep subject adaptation convolutional neural network (SACNN) for improved EEG-based MI classification.
• To enhance the extraction of temporal and spatial features from raw EEG data.
• To reduce feature distribution shifts between subjects for better generalization.

Main Methods:

• Developed a SACNN model with jointly optimized feature extractor, classifier, and subject adapter.
• Employed a parallel multiscale convolution network for simultaneous temporal and spatial feature extraction.
• Utilized maximum mean discrepancy in the subject adapter to minimize distribution shift.

Main Results:

• Achieved average accuracies of 86.42% (BCI IV IIb), 81.71% (BCI III IVa), and 79.35% (BCI IV I).
• Demonstrated significant performance improvements through statistical analysis.
• Successfully extracted temporal-spatial features for accurate prediction on new subjects.

Conclusions:

• Highlights the critical role of temporal-spatial features in EEG-based MI classification.
• The proposed SACNN model effectively leverages temporal-spatial information for enhanced classification performance.