Self-Supervised EEG Representation Learning with Contrastive Predictive Coding for Post-Stroke Patients

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Stroke patients face challenges in EEG acquisition due to fatigue and physical limitations.
• Effective feature representation is crucial for motor imagery (MI) based brain-computer interfaces (BCIs).
• Deep learning shows promise in improving BCI performance.

Purpose of the Study:

• To propose a novel framework for generating effective feature representations for MI-BCI.
• To enhance decoding performance for MI task recognition in stroke patients.
• To validate the proposed method's efficiency and accuracy.

Main Methods:

• A contrast predictive coding (CPC) framework based on modified s-transform (MST) was developed.
• MST was used for temporal-frequency feature extraction.
• EEG2Image converted multi-channel EEG into 2D topography for CPC processing.
• K-means clustering validated feature effectiveness.

Main Results:

• The MST-CPC model achieved an average classification accuracy of 89% across 40 subjects.
• The generated features demonstrated high efficiency and good clustering effects.
• The proposed method outperformed other self-supervised methods on a public dataset.

Conclusions:

• The MST-CPC framework effectively generates robust feature representations for MI-BCI.
• This approach significantly improves MI-BCI system performance, particularly for stroke patients.
• The integration of self-supervised learning and EEG image processing represents a breakthrough for BCI applications in neurorehabilitation.