Learning joint space-time-frequency features for EEG decoding on small labeled data.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) are gaining attention for controlling external devices via brain activity.
• Decoding electroencephalography (EEG) signals into accurate commands is a key challenge.
• Existing methods often lack robustness due to separate feature extraction and classification, limiting user adaptability.

Purpose of the Study:

• To develop a robust EEG decoding method for BCIs.
• To jointly learn subject-specific features and classification rules.
• To enhance the performance and adaptability of BCIs across different users.

Main Methods:

• A deep convolutional network (ConvNet) was developed for end-to-end EEG signal decoding.
• The ConvNet integrates time-frequency transformation, spatial filtering, and classification.
• Morlet wavelet-like kernels were employed for efficient feature extraction (spectral amplitude) and parameter reduction.
• Subject-to-subject weight transfer was utilized to address limited labeled data for new users.

Main Results:

• The proposed ConvNet achieved superior classification performance on three public datasets.
• The joint space-time-frequency feature extraction scheme demonstrated effectiveness.
• Subject-to-subject weight transfer improved model training with limited data.

Conclusions:

• The developed deep ConvNet offers a robust and adaptable solution for EEG decoding in BCIs.
• The joint feature learning and weight transfer strategies overcome limitations of existing methods.
• This approach advances the performance and applicability of brain-computer interfaces.