Li-Chen Shi

EEG-based emotion recognition during watching movies

This study aims at finding the relationship between EEG signals and human emotions. EEG signals are used to classify two kinds of emotions, positive and negative. First, we extracted features from original EEG data and used a linear dynamic system approach to smooth these features. An average test accuracy of 87.53% was obtained by using all of the features together with a support vector machine. Next, we reduced the dimension of features through correlation coefficients. The top 100 and top 50 subject-independent features were achieved, with average test accuracies of 89.22% and 84.94%, respectively. Finally, a manifold model was applied to find the trajectory of emotion changes.

Off-line and on-line vigilance estimation based on linear dynamical system and manifold learning

For many human machine interaction systems, to ensure work safety, the techniques for continuously estimating the vigilance of operators are highly desirable. Up to now, various methods based on electroencephalogram (EEG) are proposed to solve this problem. However, most of them are static methods and are based on supervised learning strategy. The main deficiencies of the existing methods are that the label information is hard to get and the time dependency of vigilance changes are ignored. In this paper, we introduce the dynamic characteristics of vigilance changes into vigilance estimation and propose a novel model based on linear dynamical system and manifold learning techniques to implement off-line and online vigilance estimation. In this model, both spatial information of EEG and temporal information of vigilance changes are used. The label information what we need is merely to know which EEG indices are important for vigilance estimation. Experimental results show that the mean off-line and on-line correlation coefficients between estimated vigilance level and local error rate in second-scale without being averaged are 0.89 and 0.83, respectively.

Automatic artifact removal from EEG - a mixed approach based on double blind source separation and support vector machine

Electroencephalography (EEG) recordings are often obscured by physiological artifacts that can render huge amounts of data useless and thus constitute a key challenge in current brain-computer interface research. This paper presents a new algorithm that automatically and reliably removes artifacts from EEG based on blind source separation and support vector machine. Performance on a motor imagery task is compared for artifact-contaminated and preprocessed signals to verify the accuracy of the proposed approach. The results showed improved results over all datasets. Furthermore, the online applicability of the algorithm is investigated.

Differential entropy feature for EEG-based vigilance estimation

This paper proposes a novel feature called differential entropy for EEG-based vigilance estimation. By mathematical derivation, we find an interesting relationship between the proposed differential entropy and the existing logarithm energy spectrum. We present a physical interpretation of the logarithm energy spectrum which is widely used in EEG signal analysis. To evaluate the performance of the proposed differential entropy feature for vigilance estimation, we compare it with four existing features on an EEG data set of twenty-three subjects. All of the features are projected to the same dimension by principal component analysis algorithm. Experiment results show that differential entropy is the most accurate and stable EEG feature to reflect the vigilance changes.

Dynamic clustering for vigilance analysis based on EEG

Electroencephalogram (EEG) is the most commonly studied signal for vigilance estimation. Up to now, many researches mainly focus on using supervised learning methods for analyzing EEG data. However, it is hard to obtain enough labeled EEG data to cover the whole vigilance states, and sometimes the labeled EEG data may be not reliable in practice. In this paper, we propose a dynamic clustering method based on EEG to estimate vigilance states. This method uses temporal series information to supervise EEG data clustering. Experimental results show that our method can correctly discriminate between the wakefulness and the sleepiness for every 2 seconds through EEG, and can also distinguish two other middle states between wakefulness and sleepiness.

Semi-Supervised Clustering for Vigilance Analysis Based on EEG

Vigilance research is very useful and important to our daily lives. EEG has been proved very effective for measuring vigilance. Up to now, many researches mainly focus on using supervised learning methods to analyze the vigilance. However, the labelled information of vigilance is hard to get and sometimes not reliable. In this paper, we proposed a semi-supervised clustering method for vigilance analysis based on EEG. This method uses the insufficient labeled information to guide the vigilance related feature selection and uses prior knowledge of vigilance state transform to guide the clustering algorithm. The experiment results show that our method can almost correctly distinguish the awake state and the sleeping state by EEG, and can also represent the transform processes of reasonable middle states between the awake state and the sleeping state.

Vigilance estimation by using electrooculographic features

This study aims at using electrooculographic (EOG) features, mainly slow eye movements (SEM), to estimate the human vigilance changes during a monotonous task. In particular, SEMs are first automatically detected by a method based on discrete wavelet transform, then linear dynamic system is used to find the trajectory of vigilance changes according to the SEM proportion. The performance of this system is evaluated by the correlation coefficients between the final outputs and the local error rates of the subjects. The result suggests that SEMs perform better than rapid eye movements (REM) and blinks in estimating the vigilance. Using SEM alone, the correlation can achieve 0.75 for off-line, while combined with a feature from blinks it reaches 0.79.

A novel method for EOG features extraction from the forehead

We have shown that the slow eye movements extracted from electrooculogram (EOG) signals can be used to estimate human vigilance in our previous work. However, the traditional method for recording EOG signals is to place the electrodes near the eyes of subjects. This placement is inconvenient for users in real-world applications. This paper aims to find a more practical placement for acquiring EOG signals for vigilance estimation. Instead of placing the electrodes near the eyes, we place them on the forehead. We extract EOG features from the forehead EOG signals using both independent component analysis and support vector machines. The performance of our proposed method is evaluated using the correlation coefficients between the forehead EOG signals and the traditional EOG signals. The results show that a correlation of 0.84 can be obtained when the users make 14 different face movements and for merely eye movements it reaches 0.93.

A robust principal component analysis algorithm for EEG-based vigilance estimation

Feature dimensionality reduction methods with robustness have a great significance for making better use of EEG data, since EEG features are usually high-dimensional and contain a lot of noise. In this paper, a robust principal component analysis (PCA) algorithm is introduced to reduce the dimension of EEG features for vigilance estimation. The performance is compared with that of standard PCA, L1-norm PCA, sparse PCA, and robust PCA in feature dimension reduction on an EEG data set of twenty-three subjects. To evaluate the performance of these algorithms, smoothed differential entropy features are used as the vigilance related EEG features. Experimental results demonstrate that the robustness and performance of robust PCA are better than other algorithms for both off-line and on-line vigilance estimation. The average RMSE (root mean square errors) of vigilance estimation was 0.158 when robust PCA was applied to reduce the dimensionality of features, while the average RMSE was 0.172 when standard PCA was used in the same task.

A sparse common spatial pattern algorithm for brain-computer interface

Common spatial pattern (CSP) algorithm and principal component analysis (PCA) are two commonly used key techniques for EEG component selection and EEG feature extraction for EEG-based brain-computer interfaces (BCIs). However, both the ordinary CSP and PCA algorithms face a loading problem, i.e., their weights in linear combinations are non-zero. This problem makes a BCI system easy to be over-fitted during training process, because not all of the information from EEG data are relevant to the given tasks. To deal with the loading problem, this paper proposes a spare CSP algorithm and introduces a sparse PCA algorithm to BCIs. The performance of BCIs using the proposed sparse CSP and sparse PCA techniques is evaluated on a motor imagery classification task and a vigilance estimation task. Experimental results demonstrate that the BCI system with sparse PCA and sparse CSP techniques are superior to that using the ordinary PCA and CSP algorithms.