Multi-source deep domain adaptation ensemble framework for cross-dataset motor imagery EEG transfer learning.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is crucial for brain activity measurement, with motor imagery (MI) EEG showing clinical potential.
• Convolutional neural networks (CNNs) are widely used for MI EEG classification, but performance is limited by subject-specific data scarcity.
• Lack of subject-specific data hinders decoding accuracy and generalization in MI EEG classification.

Purpose of the Study:

• To propose a novel transfer learning (TL) framework to enhance MI EEG classification performance for target subjects using auxiliary datasets.
• To develop a multi-source deep domain adaptation ensemble framework (MSDDAEF) for robust cross-dataset MI EEG decoding.
• To investigate the feasibility and effectiveness of cross-dataset TL for improving MI EEG classification.

Main Methods:

• Developed a multi-source deep domain adaptation ensemble framework (MSDDAEF) for cross-dataset MI EEG decoding.
• The MSDDAEF integrates model pre-training, deep domain adaptation, and multi-source ensemble techniques.
• Evaluated the framework's robustness by examining different designs within each component.

Main Results:

• Achieved highest average classification accuracy of 74.28% with openBMI as the target dataset and GIST as the source dataset.
• Reached an average classification accuracy of 69.85% when GIST was the target dataset and openBMI was the source dataset.
• Demonstrated superior classification performance compared to several established studies and state-of-the-art algorithms.

Conclusions:

• Cross-dataset TL is a viable approach for left/right-hand MI EEG decoding.
• The MSDDAEF presents a promising solution for mitigating cross-dataset variability in MI EEG analysis.
• The proposed framework enhances the accuracy and generalization of MI EEG classification models.