Yuankui Yang

A Brain-Computer Interface Based on a Few-Channel EEG-fNIRS Bimodal System

With the development of the wearable brain–computer interface (BCI), a few-channel BCI system is necessary for its application to daily life. In this paper, we proposed a bimodal BCI system that uses only a few channels of electroencephalograph (EEG) and functional near-infrared spectroscopy (fNIRS) signals to obtain relatively high accuracy. We developed new approaches for signal acquisition and signal processing to improve the performance of this few-channel BCI system. At the signal acquisition stage, source analysis was applied for both EEG and fNIRS signals to select the optimal channels for bimodal signal collection. At the feature extraction stage, phase-space reconstruction was applied to the selected three-channel EEG signals to expand them into multichannel signals, thus allowing the use of the traditional effective common spatial pattern to extract EEG features. For the fNIRS signal, the Hurst exponents for the selected ten channels were calculated and composed of the fNIRS data feature. At the classification stage, EEG and fNIRS features were joined and classified with the support vector machine. The averaged classification accuracy of 12 participants was 81.2% for the bimodal EEG-fNIRS signals, which was significantly higher than that for either single modality.

Comparative study on wavelet coherence and HHT coherence

Classic coherence analysis has been commonly used as a effective method for the analysis of stationary signals. To study the instantaneous coherence between non-stationary signals, we extended the concept of coherence to time-varying coherence using some time-frequency analysis methods. Wavelet-based coherence is one of the most widely used time-varying coherence methods, but few researchers have applied Hilbert-Huang transform (HHT) to coherence analysis, which also has excellent characteristics of time-frequency analysis. Therefore, this paper proposed the concept of HHT coherence, derived its method based on wavelet coherence and verified its feasibility. Then, we compared wavelet coherence and HHT coherence from three different aspects: the time-frequency resolution, effects of noise and adaptivity. The results of different simulating signals demonstrated that HHT coherence had higher time resolution, frequency resolution and more adaptivity than wavelet coherence under ideal conditions. However, due to its imperfect algorithm, the time-frequency resolution of HHT coherence was reduced by the effect of mode mixing, boundary distortion and noise. By contrast, wavelet coherence is more stable.

A partial least squares-based stimulus frequency recognition model for steady-state visual evoked potentials detection

With shorter calibration times and higher information transfer rates, steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs) have been studied most activity in recent years. Target identification is the ongoing core task in BCI researches, and plays a significant role in practical applications. In order to improve the performance of SSVEP-based BCI system, we proposed a partial least squares (PLS)-based stimulus frequency recognition model for SSVEP detection. Moreover, we compared the proposed method with canonical correlation analysis (CCA) and least absolute shrinkage and selection operator (LASSO) method, respectively. The experiment results showed that PLS can not only extract the SSVEP features effectively, but also can increase the classification accuracies of SSVEP-based BCI systems.

Rapid Detection of Targets from Complex Backgrounds Based on Eye-tracking Data

In this study1, we used a remote eye-tracker in a head-free setting to measure target detection in visual scenes. Participants underwent two kinds of tasks that were designed to simulate different situations and to study the validity and accuracy of the remote eye-tracking system. We found that the average target detection rate in the simulation task reached 88.95%, whereas in the real scene task the average accuracy was 83.20%. Our results show that the remote eye-tracker possesses enough precision to be used for the measurement of target detection in complex visual scenes.

Action understanding based on a combination of one-versus-rest and one-versus-one multi-classification methods

When people observe the actions of others, they naturally try to understand the underlying intentions. This behavior is called action understanding, and it has an important influence on mental development, language comprehension, and socialization. In this study, we used functional near-infrared spectroscopy (fNIRS) to obtain brain signals related to action understanding and then classified different intentions. Aiming to overcome the drawbacks of traditional multiclass classification methods of one-versus-rest (OVR) and one-versus-one (OVO), in this paper, we propose a new effective method to solve multiclass classification that is a combination of OVR and OVO. Compared with OVO, this new method effectively improved the accuracy of four-class classification from 25% to 48%.

One class support vector machine based filter for improving the classification accuracy of SSVEP BCI

Canonical correlation analysis (CCA) has been proved to be effective in the detection of steady state visual evoked potential (SSVEP) signals. However, the CCA method only chooses the frequency in the reference mode that corresponds to the maximum correlation value as the target. This may make the CCA output less robust. In this study, we propose a one-class support vector machine based filter to filter the sequences of correlation values in the process of the detection of SSVEP signals. The results demonstrate that the classification accuracy improved over different time windows for all subjects and the improvement achieved approximately 10% for some subjects. Moreover, the ratio of instructions that were filtered incorrectly was relative low (less than 5%) if the SSVEP signals were generated effectively.

Evaluating the Feasibility of a Novel Approach for SSVEP Detection Accuracy Improvement Using Phase Shifts

The canonical correlation analysis (CCA), double-partial least-squares (DPLS) methods and least absolute shrinkage and selection operator (LASSO) have been proven effectively in detecting the steady-state visual evoked potential (SSVEP) in SSVEP-based brain-computer interface systems. However, the accuracy of SSVEP classification can be affected by phase shifts of the electroencephalography data, so we explored the possibility of improving SSEVP detection using these methods at different phase shifts. After calculating the accuracy at different phases, we found that the phase shifts could affect the accuracy of SSVEP classification, the classification accuracy could improved about 1.1% mostly using the CCA method, meanwhile the comparison of the three methods was made at the same time and some differences between the CCA, DPLS and LASSO methods at the different phase shifts also be found. The results indicated that on the one hand, the accuracy of SSVEP detection was improved with the change of the phase, but on the other hand, although the three methods could obtain high classification accuracy, the DPLS and LASSO method showed larger fluctuations than the CCA method as the phase of the electroencephalography data of each participant or their average changed.

A sinusoidal-signal-assisted method of improving multivariate empirical mode decomposition

As the multivariate extension of empirical mode decomposition, multivariate empirical mode decomposition still suffers the problem of mode mixing. A noise-assisted method has been proposed to reduce mode mixing in multivariate empirical mode decomposition by using the dyadic filter bank property of multivariate empirical mode decomposition when applied to white Gaussian noise. However, the noise-assisted method generates redundant components that do not exist in original signals because the added noise occupies a broad range in the frequency spectrum. We propose a method of using sinusoidal signals, occupying the same frequency spectrum as the original signal, instead of white Gaussian noise to solve this problem. Results show that the new method not only solves the problem of redundant components successfully but also obtains purer modes.

Phase space reconstruction for improvement of classification in few-channel BCI systems

With advances in brain-computer interface (BCI) research for the practical use of BCI systems, few-channel BCI systems have become necessary. The common spatial pattern (CSP) algorithm is a classic and powerful tool for extraction of features for motor imagery in BCI systems. However, previous studies show that this algorithm is not suitable for few-channel systems. In this study, phase space reconstruction (PSR) was used to decompose few-channel electroencephalography (EEG) signals into multichannel information. Using the reconstructed data, CSP and a support vector machine (SVM) were combined to obtain high classification accuracies from a small number of channels. The mean accuracy for the EEG signals from three channels was 0.74 for PSR + CSP + SVM, while this accuracy was only 0.43 for CSP + SVM, which suggests that PSR + CSP + SVM is practicable for few-channel BCI systems.

Classification of intention understanding using EEG-NIRS bimodal system

When people observe the actions of others, they always try to understand the intentions underlying the actions. The neural mechanism of this understanding is referred to as the mirror neuron system (MNS). Different actions may correspond to different intentions, and the activation of the MNS in the human brain may also be slightly different. The present study distinguishes these differences according to functional brain imaging signals analyzed with machine learning. Brain signals were detected when the participants observed two types of actions: (1) grasping a cup for drinking, and (2) no meaningful contact. A synchronous measurement method for EEG and NIRS was adopted to increase the information contained in the brain signals. In order to obtain better classification accuracies, the method of functional brain networks was used. This method can be used to examine the relationship between brain regions. First, phase synchronization and Pearson correlation were used to calculate correlations for EEG channels and NIRS channels, respectively. Next, correlation matrices were converted into binary matrices, and the local properties of the networks were obtained. Finally, the feature vectors for the classifier were selected by analysis of their significance. In addition, EEG data and NIRS data were combined at the feature level and better classification results were obtained.