Different classification techniques considering brain computer interface applications

In this work the application of different machine learning techniques for classification of mental tasks from electroencephalograph (EEG) signals is investigated. The main application for this research is the improvement of brain computer interface (BCI) systems. For this purpose, Bayesian graphical network, neural network, Bayesian quadratic, Fisher linear and hidden Markov model classifiers are applied to two known EEG datasets in the BCI field. The Bayesian network classifier is used for the first time in this work for classification of EEG signals. The Bayesian network appeared to have a significant accuracy and more consistent classification compared to the other four methods. In addition to classical correct classification accuracy criteria, the mutual information is also used to compare the classification results with other BCI groups.     

EEG-based motor imagery analysis using weighted wavelet transform features

In this study, an electroencephalogram (EEG) analysis system for single-trial classification of motor imagery (MI) data is proposed. Feature extraction in brain-computer interface (BCI) work is an important task that significantly affects the success of brain signal classification. The continuous wavelet transform (CWT) is applied together with Student's two-sample t-statistics for 2D time-scale feature extraction, where features are extracted from EEG signals recorded from subjects performing left and right MI. First, we utilize the CWT to construct a 2D time-scale feature, which yields a highly redundant representation of EEG signals in the time-frequency domain, from which we can obtain precise localization of event-related brain desynchronization and synchronization (ERD and ERS) components. We then weight the 2D time-scale feature with Student's two-sample t-statistics, representing a time-scale plot of discriminant information between left and right MI. These important characteristics, including precise localization and significant discriminative ability, substantially enhance the classification of mental tasks. Finally, a correlation coefficient is used to classify the MI data. Due to its simplicity, it will enable the performance of our proposed method to be clearly demonstrated. Compared to a conventional 2D time-frequency feature and three well-known time-frequency approaches, the experimental results show that the proposed method provides reliable 2D time-scale features for BCI classification.     

A Twin Multi-Class Classification Support Vector Machine

Twin support vector machine (TSVM) is a novel machine learning algorithm, which aims at finding two nonparallel planes for each class. In order to do so, one needs to resolve a pair of smaller-sized quadratic programming problems rather than a single large one. Classical TSVM is proposed for the binary classification problem. However, multi-class classification problem is often met in our real world. For this problem, a new multi-class classification algorithm, called Twin-KSVC, is proposed in this paper. It takes the advantages of both TSVM and K-SVCR (support vector classification-regression machine for k-class classification) and evaluates all the training points into a “1-versus-1-versus-rest” structure, so it generates ternary outputs { ?1, 0, +1}. As all the samples are utilized in constructing the classification hyper-plane, our proposed algorithm yields higher classification accuracy in comparison with other two algorithms. Experimental results on eleven benchmark datasets demonstrate the feasibility and validity of our proposed algorithm.     

基于功率谱峰值及时变线性分类算法的运动意识分类

背景：脑电信号的特征提取是脑机接口系统中一个重要的环节,如何快速有效地提取反映大脑意识任务状态的脑电特征是进行分类、正确解读意识任务的关键。目前,提取脑电信号特征通常采用功率谱密度估计、自回归模型和小波变换等方法,这些特征都是以脑电信号的线性化为前提,上述方法不能很好地反映出大脑的非线性动力学性质。 目的：分析脑电信号功率谱峰值在识别左右手想象运动中的作用。 方法：采用脑机接口2003竞赛中Graz科技大学提供的脑电数据,用小波包分解获取8~24 Hz脑电信号,计算C3,C4电极脑电信号的功率谱峰值作为脑电特征向量,运用时变线性分类算法对运动意识任务运行分类。 结果与结论：对140次实验的测试样本进行数据分析,最大分类正确率可达89.29%,最大互信息和信噪比分别为0.626 9 bit和1.384 8。C3,C4电极8~24 Hz脑电信号功率谱峰值能很好地反映左右手运动想象脑电特征的变化,与事件相关去同步/事件相关同步现象变化一致,可在线识别左右手想象运动。     

基于Mu/Beta节律的想象运动脑电信号特征提取

背景：不同的运动会产生不同的脑电信号,脑机接口技术就是利用脑电信号的特异性,通过现代信号处理技术和外部的连接实现人脑与外部设备的通信。以实现脑机接口在线研究的目标,首先要解决的是脑电信号处理的速度问题。 目的：研究快速、准确地提取脑电信号特征及分类的方法。 方法：充分利用想象运动过程中,脑电信号中Mu/Beta节律的事件相关同步化和去同步化特性,以2003年BCI竞赛数据为处理对象,采用带通滤波和小波包分析的方法提取Mu、Beta节律,提取C3、C4两通道上的能量平均值形成二维特征向量,利用matlab自带的classify函数进行分类。 结果与结论：通过对训练数据进行测试得到较为合适的参数,利用该参数对同等条件下的训练数据和测试数据分别进行判别,分类正确率分别达到87.857%和88.571%。 关键词：特征提取与分类;脑电信号;事件相关同步化/去同步化;想象运动;小波包分析     

基于二阶盲辨识的运动想象脑电信号分类

人类的思维活动会引起脑电信号的变化，而对特定的脑电信号的分析可以设计脑机接口通信系统。文章提出了一种应用于脑机接口系统的运动想象脑电信号分类方法，即采用一种盲源分离算法——二阶盲辨识对运动想象脑电信号进行处理，进而根据不同类型间的Fisher距离提取特征，最后运用多层BP神经网络进行分类，3个样本的平均分类准确率达到90.9%，而直接对原始脑电信号进行分类，其平均分类准确率仅为72.6%，这表明二阶盲辨识算法的运用明显改善了运动想象脑电信号的分类准确率。不同受试者及不同训练样本的分类准确率表明，以BP神经网络作为分类器，其结果与训练样本关系密切，尤其是在样本数较少的情况下。从实时性方面来说，二阶盲辨识算法完全满足要求。     

A high performance sensorimotor beta rhythm-based brain-computer interface associated with human natural motor behavior

To explore the reliability of a high performance brain-computer interface (BCI) using non-invasive EEG signals associated with human natural motor behavior does not require extensive training. We propose a new BCI method, where users perform either sustaining or stopping a motor task with time locking to a predefined time window. Nine healthy volunteers, one stroke survivor with right-sided hemiparesis and one patient with amyotrophic lateral sclerosis (ALS) participated in this study. Subjects did not receive BCI training before participating in this study. We investigated tasks of both physical movement and motor imagery. The surface Laplacian derivation was used for enhancing EEG spatial resolution. A model-free threshold setting method was used for the classification of motor intentions. The performance of the proposed BCI was validated by an online sequential binary-cursor-control game for two-dimensional cursor movement. Event-related desynchronization and synchronization were observed when subjects sustained or stopped either motor execution or motor imagery. Feature analysis showed that EEG beta band activity over sensorimotor area provided the largest discrimination. With simple model-free classification of beta band EEG activity from a single electrode (with surface Laplacian derivation), the online classifications of the EEG activity with motor execution/motor imagery were: >90%/ approximately 80% for six healthy volunteers, >80%/ approximately 80% for the stroke patient and approximately 90%/ approximately 80% for the ALS patient. The EEG activities of the other three healthy volunteers were not classifiable. The sensorimotor beta rhythm of EEG associated with human natural motor behavior can be used for a reliable and high performance BCI for both healthy subjects and patients with neurological disorders. Significance: The proposed new non-invasive BCI method highlights a practical BCI for clinical applications, where the user does not require extensive training.     

Neural network classification of autoregressive features from electroencephalogram signals for brain-computer interface design

In this paper, we have designed a two-state brain-computer interface (BCI) using neural network (NN) classification of autoregressive (AR) features from electroencephalogram (EEG) signals extracted during mental tasks. The main purpose of the study is to use Keirn and Aunon's data to investigate the performance of different mental task combinations and different AR features for BCI design for individual subjects. In the experimental study, EEG signals from five mental tasks were recorded from four subjects. Different combinations of two mental tasks were studied for each subject. Six different feature extraction methods were used to extract the features from the EEG signals: AR coefficients computed with Burg's algorithm, AR coefficients computed with a least-squares (LS) algorithm and adaptive autoregressive (AAR) coefficients computed with a least-mean-square (LMS) algorithm. All the methods used order six applied to 125 data points and these three methods were repeated with the same data but with segmentation into five segments in increments of 25 data points. The multilayer perceptron NN trained by the back-propagation algorithm (MLP-BP) and linear discriminant analysis (LDA) were used to classify the computed features into different categories that represent the mental tasks. We compared the classification performances among the six different feature extraction methods. The results showed that sixth-order AR coefficients with the LS algorithm without segmentation gave the best performance (93.10%) using MLP-BP and (97.00%) using LDA. The results also showed that the segmentation and AAR methods are not suitable for this set of EEG signals. We conclude that, for different subjects, the best mental task combinations are different and proper selection of mental tasks and feature extraction methods are essential for the BCI design.     

Plausibility assessment of a 2-state self-paced mental task-based BCI using the no-control performance analysis

The feasibility of having a self-paced brain–computer interface (BCI) based on mental tasks is investigated. The EEG signals of four subjects performing five mental tasks each are used in the design of a 2-state self-paced BCI. The output of the BCI should only be activated when the subject performs a specific mental task and should remain inactive otherwise. For each subject and each task, the feature coefficient and the classifier that yield the best performance are selected, using the autoregressive coefficients as the features. The classifier with a zero false positive rate and the highest true positive rate is selected as the best classifier. The classifiers tested include: linear discriminant analysis, quadratic discriminant analysis, Mahalanobis discriminant analysis, support vector machine, and radial basis function neural network. The results show that: (1) some classifiers obtained the desired zero false positive rate; (2) the linear discriminant analysis classifier does not yield acceptable performance; (3) the quadratic discriminant analysis classifier outperforms the Mahalanobis discriminant analysis classifier and performs almost as well as the radial basis function neural network; and (4) the support vector machine classifier has the highest true positive rates but unfortunately has nonzero false positive rates in most cases.     

Improving BCI performance by task-related trial pruning

Noise in electroencephalography data (EEG) is an ubiquitous problem that limits the performance of brain computer interfaces (BCI). While typical EEG artifacts are usually removed by trial rejection or by filtering, noise induced in the data by the subject’s failure to produce the required mental state is very harmful. Such “noise” effects are rather common, especially for naive subjects in their training phase and, thus, standard artifact removal methods would inevitably fail. In this paper, we present a novel method which aims to detect such defected trials taking into account the intended task by use of Relevant Dimensionality Estimation (RDE), a new machine learning method for denoising in feature space. In this manner, our method effectively “cleans” the training data and thus allows better BCI classification. Preliminary results conducted on a data set of 43 naive subjects show a significant improvement for 74% of the subjects.     

Characterization of four-class motor imagery EEG data for the BCI-competition 2005

To determine and compare the performance of different classifiers applied to four-class EEG data is the goal of this communication. The EEG data were recorded with 60 electrodes from five subjects performing four different motor-imagery tasks. The EEG signal was modeled by an adaptive autoregressive (AAR) process whose parameters were extracted by Kalman filtering. By these AAR parameters four classifiers were obtained, namely minimum distance analysis (MDA)--for single-channel analysis, and linear discriminant analysis (LDA), k-nearest-neighbor (kNN) classifiers as well as support vector machine (SVM) classifiers for multi-channel analysis. The performance of all four classifiers was quantified and evaluated by Cohen's kappa coefficient, an advantageous measure we introduced here to BCI research for the first time. The single-channel results gave rise to topographic maps that revealed the channels with the highest level of separability between classes for each subject. Our results of the multi-channel analysis indicate SVM as the most successful classifier, whereas kNN performed worst.     

Adaptive estimation of EEG-rhythms for optimal band identification in BCI

The amplitude of EEG μ-rhythm is large when the subject does not perform or imagine movement and attenuates when the subject either performs or imagines movement. The knowledge of EEG individual frequency components in the time-domain provides useful insight into the classification process. Identification of subject-specific reactive band is crucial for accurate event classification in brain-computer interfaces (BCI). This work develops a simple time-frequency decomposition method for EEG μ rhythm by adaptive modeling. With the time-domain decomposition of the signal, subject-specific reactive band identification method is proposed. Study is conducted on 30 subjects for optimal band selection for four movement classes. Our results show that over 93% the subjects have an optimal band and selection of this band improves the relative power spectral density by 200% with respect to normalized power.     

Imagery of motor actions: differential effects of kinesthetic and visual-motor mode of imagery in single-trial EEG

Single-trial motor imagery classification is an integral part of a number of brain-computer interface (BCI) systems. The possible significance of the kind of imagery, involving rather kinesthetic or visual representations of actions, was addressed using the following experimental conditions: kinesthetic motor imagery (MIK), visual-motor imagery (MIV), motor execution (ME) and observation of movement (OOM). Based on multi-channel EEG recordings in 14 right-handed participants, we applied a learning classifier, the distinction sensitive learning vector quantization (DSLVQ) to identify relevant features (i.e., frequency bands, electrode sites) for recognition of the respective mental states. For ME and OOM, the overall classification accuracies were about 80%. The rates obtained for MIK (67%) were better than the results of MIV (56%). Moreover, the focus of activity during kinesthetic imagery was found close to the sensorimotor hand area, whereas visual-motor imagery did not reveal a clear spatial pattern. Consequently, to improve motor-imagery-based BCI control, user training should emphasize kinesthetic experiences instead of visual representations of actions.     

Motor imagery and action observation: Modulation of sensorimotor brain rhythms during mental control of a brain–computer interface

ObjectiveThis study investigates the impact of a continuously presented visual feedback in the form of a grasping hand on the modulation of sensorimotor EEG rhythms during online control of a brain–computer interface (BCI).MethodsTwo groups of participants were trained to use left or right hand motor imagery to control a specific output signal on a computer monitor: the experimental group controlled a moving hand performing an object-related grasp (‘realistic feedback’), whereas the control group controlled a moving bar (‘abstract feedback’). Continuous feedback was realized by using the outcome of a real-time classifier which was based on EEG signals recorded from left and right central sites.ResultsThe classification results show no difference between the two feedback groups. For both groups, ERD/ERS analysis revealed a significant larger ERD during feedback presentation compared to an initial motor imagery screening session without feedback. Increased ERD during online BCI control was particularly found for the lower alpha (8–10 Hz) and for the beta bands (16–20, 20–24 Hz).ConclusionsThe present study demonstrates that visual BCI feedback clearly modulates sensorimotor EEG rhythms. When the feedback provides equivalent information on both the continuous and final outcomes of mental actions, the presentation form (abstract versus realistic) does not influence the performance in a BCI, at least in initial training sessions.SignificanceThe present results are of practical interest for classifier development and BCI use in the field of motor restoration.     

Brain-Computer Interface with Corrupted EEG Data: a Tensor Completion Approach

One of the current issues in brain-computer interface (BCI) is how to deal with noisy electroencephalography (EEG) measurements organized as multidimensional datasets (tensors). On the other hand, recently, significant advances have been made in multidimensional signal completion algorithms that exploit tensor decomposition models to capture the intricate relationship among entries in a multidimensional signal. We propose to use tensor completion applied to EEG data for improving the classification performance in a motor imagery BCI system with corrupted measurements. Noisy measurements (electrode misconnections, subject movements, etc.) are considered as unknowns (missing samples) that are inferred from a tensor decomposition model (tensor completion). We evaluate the performance of four recently proposed tensor completion algorithms, CP-WOPT (Acar et al. Chemom Intell Lab Syst. 106:41-56, 2011), 3DPB-TC (Caiafa et al. 2013), BCPF (Zhao et al. IEEE Trans Pattern Anal Mach Intell. 37(9):1751-1763, 2015), and HaLRT (Liu et al. IEEE Trans Pattern Anal Mach Intell. 35(1):208-220, 2013), plus a simple interpolation strategy, first with random missing entries and then with missing samples constrained to have a specific structure (random missing channels), which is a more realistic assumption in BCI applications. We measured the ability of these algorithms to reconstruct the tensor from observed data. Then, we tested the classification accuracy of imagined movement in a BCI experiment with missing samples. We show that for random missing entries, all tensor completion algorithms can recover missing samples increasing the classification performance compared to a simple interpolation approach. For the random missing channels case, we show that tensor completion algorithms help to reconstruct missing channels, significantly improving the accuracy in the classification of motor imagery (MI), however, not at the same level as clean data. Summarizing, compared to the interpolation case, all tensor completion algorithms succeed to increase the classification performance by 7–9% (LDA–SVD) for random missing entries and 15–8% (LDA–SVD) for random missing channels. Tensor completion algorithms are useful in real BCI applications. The proposed strategy could allow using motor imagery BCI systems even when EEG data is highly affected by missing channels and/or samples, avoiding the need of new acquisitions in the calibration stage.     

Wavelet-based fractal features with active segment selection: application to single-trial EEG data

Feature extraction in brain-computer interface (BCI) work is one of the most important issues that significantly affect the success of brain signal classification. A new electroencephalogram (EEG) analysis system utilizing active segment selection and multiresolution fractal features is designed and tested for single-trial EEG classification. Applying event-related brain potential (ERP) data acquired from the sensorimotor cortices, the proposed system consists of three main procedures including active segment selection, feature extraction, and classification. The active segment selection is based on the continuous wavelet transform (CWT) and Student's two-sample t-statistics, and is used to obtain the optimal active time segment in the time-frequency domain. We then utilize a modified fractal dimension to extract multiresolution fractal feature vectors from the discrete wavelet transform (DWT) data for movement classification. By using a simple linear classifier, we find significant improvements in the rate of correct classification over the conventional approaches in all of our single-trial experiments for real finger movement. These results can be extended to see the good adaptability of the proposed method to imaginary movement data acquired from the public databases.     

Walking from thought

Online analysis and classification of single electroencephalogram (EEG) trials during motor imagery were used for navigation in the virtual environment (VE). The EEG was recorded bipolarly with electrode placement over the hand and foot representation areas. The aim of the study was to demonstrate for the first time that it is possible to move through a virtual street without muscular activity when the participant only imagines feet movements. This is achieved by exploiting a brain-computer interface (BCI) which transforms thought-modulated EEG signals into an output signal that controls events within the VE. The experiments were carried out in an immersive projection environment, commonly referred to as a "Cave" (Cruz-Neira, C., Sandin, D.J., DeFanti, T.A., Surround-screen projection-based virtual reality: the design and implementation of the CAVE. Proceedings of the 20th annual conference on Computer graphics and interactive techniques, ACM Press, 1993, pp. 135-142) where participants were able to move through a virtual street by foot imagery only. Prior to the final experiments in the Cave, the participants underwent an extensive BCI training.     

Brain-computer interface using a simplified functional near-infrared spectroscopy system

A brain-computer interface (BCI) is a device that allows a user to communicate with external devices through thought processes alone. A novel signal acquisition tool for BCIs is near-infrared spectroscopy (NIRS), an optical technique to measure localized cortical brain activity. The benefits of using this non-invasive modality are safety, portability and accessibility. A number of commercial multi-channel NIRS system are available; however we have developed a straightforward custom-built system to investigate the functionality of a fNIRS-BCI system. This work describes the construction of the device, the principles of operation and the implementation of a fNIRS-BCI application, 'Mindswitch' that harnesses motor imagery for control. Analysis is performed online and feedback of performance is presented to the user. Mindswitch presents a basic 'on/off' switching option to the user, where selection of either state takes 1 min. Initial results show that fNIRS can support simple BCI functionality and shows much potential. Although performance may be currently inferior to many EEG systems, there is much scope for development particularly with more sophisticated signal processing and classification techniques. We hope that by presenting fNIRS as an accessible and affordable option, a new avenue of exploration will open within the BCI research community and stimulate further research in fNIRS-BCIs.     

A Voting Optimized Strategy Based on ELM for Improving Classification of Motor Imagery BCI Data

This paper presents an approach to classifying electroencephalogram (EEG) signals for brain–computer interfaces (BCI). To eliminate redundancy in high-dimensional EEG signals and reduce the coupling among different classes of EEG signals, we use principle component analysis and linear discriminant analysis to extract features that represent the raw signals. Next, we introduce the voting-based extreme learning machine to classify the features. Experiments performed on real-world data from the 2003 BCI competition indicate that our classification method outperforms state-of-the-art methods in speed and accuracy.     

Robust modeling based on optimized EEG bands for functional brain state inference

The need to infer brain states in a data driven approach is crucial for BCI applications as well as for neuroscience research. In this work we present a novel classification framework based on Regularized Linear Regression classifier constructed from time-frequency decomposition of an EEG (electro-encephalography) signal. The regression is then used to derive a model of frequency distributions that identifies brain states. The process of classifier construction, preprocessing and selection of optimal regularization parameter by means of cross-validation is presented and discussed. The framework and the feature selection technique are demonstrated on EEG data recorded from 10 healthy subjects while requested to open and close their eyes every 30 s. This paradigm is well known in inducing Alpha power modulations that differ from low power (during eyes opened) to high (during eyes closed). The classifier was trained to infer eyes opened or eyes closed states and achieved higher than 90% classification accuracy. Furthermore, our findings reveal interesting patterns of relations between experimental conditions, EEG frequencies, regularization parameters and classifier choice. This viable tool enables identification of the most contributing frequency bands to any given brain state and their optimal combination in inferring this state. These features allow for much greater detail than the standard Fourier Transform power analysis, making it an essential method for both BCI proposes and neuroimaging research.