Explainable cross-task adaptive transfer learning for motor imagery EEG classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep transfer learning (TL) is crucial for motor imagery (MI) electroencephalography (EEG)-based brain-computer interfaces (BCIs) due to limited subject-specific data.
• Existing TL methods have advanced cross-subject/session and cross-device scenarios, but cross-task deep TL for MI EEG remains underexplored.

Purpose of the Study:

• To develop a novel, explainable cross-task adaptive TL method for MI EEG decoding.
• To investigate the feasibility of leveraging motor execution (ME) EEG data for enhanced MI EEG decoding.
• To address the challenge of limited training samples in MI EEG decoding.

Main Methods:

• Proposed a cross-task adaptive TL approach involving similarity analysis and data alignment between ME and MI EEG data.
• Pre-trained a deep learning model using extensive ME EEG data and fine-tuned it with partial MI EEG data.
• Employed expected gradient-based post-hoc explainability analysis to visualize key temporal-spatial features.

Main Results:

• Achieved high classification accuracies of 80.00% on the OpenBMI dataset and 72.73% on the GIST dataset.
• Outperformed several state-of-the-art algorithms in MI EEG decoding.
• Explainability analysis confirmed the correlation between ME and MI EEG data and the effectiveness of cross-task adaptation.

Conclusions:

• Decoding of MI EEG can be significantly improved by utilizing pre-existing ME EEG data.
• The proposed method effectively relaxes the constraint of limited training samples for MI EEG decoding.
• This approach holds practical significance for developing more robust and accessible BCIs.