Related Experiment Video
Updated: Oct 12, 2025
10:14
Author Spotlight: Enhancing Neurorehabilitation Through EEG, Motor Imagery, and Virtual Reality
Published on: May 10, 2024
1.3K
Selective Cross-Subject Transfer Learning Based on Riemannian Tangent Space for Motor Imagery Brain-Computer
Yilu Xu1, Xin Huang2, Quan Lan3
1School of Software, Jiangxi Agricultural University, Nanchang, China.
Frontiers in Neuroscience
|November 22, 2021
Summary
This study introduces a new transfer learning method for brain-computer interfaces (BCIs) to improve motor imagery (MI) detection in stroke rehabilitation. The approach reduces calibration time by effectively using data from other subjects.
Area of Science:
- Neuroscience
- Biomedical Engineering
- Machine Learning
Background:
- Motor imagery (MI) brain-computer interfaces (BCIs) are crucial for stroke patient rehabilitation, offering real-time control via electroencephalogram (EEG).
- Identifying MI EEG signals is challenging due to significant inter-subject and inter-session variability, necessitating extensive subject-specific calibration.
- Current methods require substantial labeled data for subject-specific models, leading to lengthy and tedious calibration processes.
Purpose of the Study:
- To address the challenge of long calibration times in EEG-based MI BCI for stroke rehabilitation.
- To develop a novel transfer learning approach that effectively utilizes labeled and unlabeled data from both target and source subjects.
- To improve the accuracy and efficiency of MI signal identification in BCI applications.
Main Methods:
- A supervised selective cross-subject transfer learning (sSCSTL) approach is proposed, leveraging Riemannian tangent space for EEG signal analysis.
- Covariance matrices representing multi-channel EEG signals undergo Riemannian alignment to minimize subject-specific variations.
- Aligned covariance matrices are transformed into Riemannian tangent space features for Euclidean space classification, with semi-supervised and unsupervised versions explored.
Main Results:
- The proposed sSCSTL algorithms demonstrated superior performance compared to state-of-the-art methods on two public MI datasets, particularly with limited target subject labeled data.
- The algorithms effectively transferred knowledge from suitable source subjects to the target subject's feature space, significantly reducing calibration requirements.
- Performance was especially notable for 'good' target subjects, indicating the method's adaptability.
Conclusions:
- The developed transfer learning framework significantly enhances the efficiency and effectiveness of EEG-based MI BCIs for stroke rehabilitation.
- The approach mitigates the need for extensive subject-specific calibration, making BCI technology more accessible and practical for patients.
- Future work can further explore the utility of unlabeled data and optimize source subject selection for even greater BCI performance gains.
