Federated Transfer Learning for EEG Signal Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning (DL) methods show promise for electroencephalographic (EEG) signal classification in brain-computer interfaces (BCI).
• Limited availability of large, diverse EEG datasets hinders the widespread success of DL in BCI applications.
• Privacy concerns restrict the aggregation of multi-subject EEG data, preventing the creation of comprehensive training datasets.

Purpose of the Study:

• To propose a novel, privacy-preserving deep learning architecture for EEG classification.
• To address the challenge of limited data by leveraging federated learning and domain adaptation techniques.
• To improve the accuracy of motor imagery classification in BCI systems.

Main Methods:

• Development of a federated transfer learning (FTL) architecture based on the federated learning framework.
• Utilizing the single-trial covariance matrix for feature extraction from multi-subject EEG data.
• Employing domain adaptation techniques to extract common discriminative information across subjects.

Main Results:

• The FTL approach achieved a 2% higher classification accuracy in subject-adaptive analysis compared to baseline methods, while ensuring data privacy.
• In scenarios lacking multi-subject data, the proposed FTL architecture demonstrated a 6% improvement in accuracy over existing state-of-the-art DL architectures.
• Evaluation was performed on the PhysioNet dataset for a 2-class motor imagery classification task.

Conclusions:

• Federated transfer learning offers a viable solution for privacy-preserving EEG classification in BCI.
• The proposed FTL architecture effectively extracts discriminative information from decentralized EEG data.
• FTL significantly enhances classification accuracy, overcoming limitations posed by small and private datasets.