EEG decoding for datasets with heterogenous electrode configurations using transfer learning graph neural networks.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-machine interfaces (BMI) rely on machine learning for feature extraction, requiring large datasets.
• Combining diverse datasets is challenging due to variations in recording equipment and electrode layouts, leading to data distribution shifts.

Purpose of the Study:

• To develop a machine learning framework that overcomes domain adaptation challenges in BMI.
• To enable learning from diverse datasets with varying electrode configurations and experimental protocols.

Main Methods:

• A novel framework combining graph neural networks (GNNs) and transfer learning was developed.
• The approach was applied to non-invasive motor imagery (MI) electroencephalography (EEG) decoding.
• Three MI EEG databases with different electrode numbers (22-64) and layouts were utilized.

Main Results:

• The GNN-based transfer learning framework achieved high accuracy and low standard deviations on testing datasets.
• The model effectively aggregated knowledge from datasets with differing electrode layouts.
• Improved generalization in subject-independent MI EEG classification was observed.

Conclusions:

• The proposed framework effectively addresses domain adaptation issues in BMI by integrating diverse EEG datasets.
• This approach enhances the generalization capabilities of motor imagery classification, overcoming limitations of non-unified experimental setups.
• The findings advance the development and application of brain-computer interface technologies.