Multivariate empirical mode decomposition and multiscale entropy analysis of EEG signals from SSVEP-based BCI system

Abstract

The steady-state visual evoked potential (SSVEP)-based Brain-Computer Interface (BCI) has been employed in the brain-controlled wheelchair system for patients with severe dyskinesia disease. However, a long-time operation brings about users fatigue, leading to a decrease of performance of the BCI system in practical applications. The characterization of the fatigued mechanism and the improvement of the SSVEP classification accuracy remains a challenging problem of significant importance. In this letter, we first conduct SSVEP experiments to acquire the EEG signals during both normal and fatigued states. Then we develop a novel framework, which integrates the advantages of multivariate empirical mode decomposition (MEMD) and Support Vector Machine (SVM), for improving the SSVEP classification accuracy, especially during the fatigued state. The results suggest that the novel framework enables us to obtain a higher SSVEP classification accuracy compared with the method without MEMD. Furthermore, in order to reveal the fatigued behavior, we use the multivariate multiscale sample entropy (MMSE) to analyze the multi-channel EEG signals corresponding to normal and fatigued states. We interestingly find that the MMSE values in the fatigued state are lower than that in the normal state, which reflects the increase of regularity in SSVEP signals during the fatigued state. That is, a greater synchronization of neural assemblies is required to realize cognitive impairment when fatigue happens. The knowledge for the understanding of the brain fatigued behavior underlying SSVEP-based BCI experiments is gained by our analysis.

Enhancing SSVEP-BCI Performance Under Fatigue State Using Dynamic Stopping Strategy

“…In the condition of processing data of length 0.5-1s, statistical analysis shows that the accuracy of each method is significantly higher than that of FS 0.5s strategy (hypothesis testing-based DS strategy: t low (23)=4.850, p low <0.0001, t high (23)=6.854, p high <0.0001; Bayesian-based DS strategy: t low (23)=5.371, p low <0.0001, t high (23)=6.782, p high <0.0001) and ITR is significantly higher than that of FS 1s strategy (hypothesis testing-based DS strategy: t low (23)=12.770, p low <0.0001, t high (23)=11.270, p high <0.0001; Bayesian-based DS strategy: t low (23)=12.110, p low <0.0001, t high (23)=9.908, p high <0.0001). We can also find that the ITR of Bayesian-based DS strategy is significantly higher than that of FS 0.5s strategy (t low (23)=4.389, p low =0.0002, t high (23)=5.105, p high <0.0001), which we believe is due to the high accuracy and low selection time guaranteed by Bayesianbased DS strategy. In the condition of processing data of length 1-2s, statistical analysis shows that the accuracy of each method is significantly higher than that of FS 1s strategy (hypothesis testing-based DS strategy: t low ( 23 23)=2.245, p high <0.0001), which is consistent with the conclusion we mentioned above.…”

“…Concerning algorithms, Gao et al developed an adaptive optimal-Kernel time-frequency representation complex network method to analyze the fatigue on the SSVEP classification accuracy from the perspective of brain networks [22]. Later, they proposed an alternative approach, combining the multivariate empirical mode decomposition with the Support Vector Machine, to improve the detection of SSVEPs under fatigue state [23]. Ajami et al employed the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm for frequency recognition and successfully compensated for the fatigue effect [24].…”