A systematic evaluation of Euclidean alignment with deep learning for EEG decoding

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) signals are crucial for Brain-Computer Interface (BCI) tasks.
• Deep learning (DL) models show promise for BCI but require substantial data.
• Transfer learning using multi-subject data can improve DL model training efficiency.

Purpose of the Study:

• To systematically evaluate the impact of Euclidean alignment (EA) on DL model training for BCI signal decoding.
• To assess EA's effectiveness in improving both shared and individual DL models for BCI tasks.
• To investigate EA's role in enhancing transfer learning performance in BCI applications.

Main Methods:

• Utilized Euclidean alignment (EA) as a pre-processing step for EEG data.
• Trained shared DL models using multi-subject data with EA pre-processing.
• Evaluated the transferability of trained models to new subjects.
• Compared EA's performance against individual DL models in an ensemble classifier.

Main Results:

• EA pre-processing improved decoding accuracy in the target subject by 4.33%.
• EA significantly decreased DL model convergence time by over 70%.
• For ensemble classifiers, EA improved accuracy by 3.71%, though shared models with EA still outperformed ensembles.

Conclusions:

• Euclidean alignment is effective in enhancing transfer learning for DL models in BCI.
• EA can be a valuable pre-processing technique to improve BCI performance and efficiency.
• The findings suggest EA could become a standard pre-processing method for BCI research and applications.