Transfer learning with data alignment and optimal transport for EEG based motor imagery classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) signal non-stationarity and inter-subject variability pose challenges for Brain-Computer Interfaces (BCI).
• Current BCI systems require lengthy calibration procedures to adapt to individual users.
• Transfer Learning (TL) is explored as a method to leverage existing data for faster BCI adaptation.

Purpose of the Study:

• To propose a novel Multi-source domain Transfer Learning Fusion (MTLF) framework.
• To address the significant calibration problem in BCI research.
• To reduce the calibration workload and improve classification accuracy.

Main Methods:

• The MTLF framework transforms source domain data using resting-state segments to minimize domain discrepancies.
• Common Spatial Pattern (CSP) is utilized for effective feature extraction.
• An improved TL classifier is employed for target sample classification without requiring target domain labels.

Main Results:

• The MTLF framework achieved the best performance compared to other algorithms on BCI Competition IV Datasets 2a and 2b.
• Mean classification accuracies of 73.69% on Dataset 2a and 70.83% on Dataset 2b were obtained.
• The framework effectively reduced domain discrepancies between source and target data.

Conclusions:

• The proposed MTLF framework significantly reduces the calibration burden in BCI systems.
• MTLF demonstrates superior classification performance on motor imagery tasks.
• This approach offers a promising solution for more efficient and accurate BCI applications.