Automatic Classification of SSVEP Responses with Limited Training Samples and Shrinkage-Based Regularized Covariance Matrix

Document Type : Original Article

Authors

1 School of Engineering Science, College of Engineering, University of Tehran, Tehran, Iran

2 Biomedical Engineering Faculty, Sahand University of Technology, Tabriz, Iran

Abstract

A major hurdle in brain computer interface (BCI) development is the low information transfer rate (ITR). Using a short stimulation time is a solution that offers the advantages of increasing the ITR value and reducing the mental fatigue of users. When using short stimulation time, algorithms based on linearly constrained minimum variance beamforming (LCMV) provide better performance over other classifiers. However, their performance in aforementioned condition is still low due to the ill-conditioned estimation of the data covariance matrix. To address this problem, this study proposes the use of four Shrinkage-based regularized covariance matrices, including convex combination (CC), generalized linear combination (GLC), modified CC (MCC), and modified GLC (MGLC). The proposed covariance matrices are applied in the spatial-temporal beamformer LCMVst to construct a better weight vector, thereby improving the classification performance. The results showed that when using the shortest stimulation time (0.25s), the proposed beamformers LCMVst-CC, LCMVst-GLC, LCMVst-MCC, and LCMVst-MGLC provided a significant improvement of about 27% in average classification accuracy over conventional LCMVst. Also, the LCMVst-MCC and LCMVst-MGLC methods compared to LCMVst-CC and LCMVst-GLC methods provided approximately 9% improvement in classification accuracy. The results of this study show that the proposed beamformers have high potential in the development of SSVEP-based BCI systems.

Keywords

Main Subjects

TABRIZ JOURNAL OF ELECTRICAL ENGINEERING

Articles in Press, Corrected Proof
Available Online from 16 October 2024

Files

Share

How to cite

Statistics