Simple adaptive sparse representation based classification schemes for EEG based brain–computer interface applications
Introduction
Brain–computer interface (BCI) systems provide a new communication and control channel between human brain and an external device without any muscle movements [1]. Due to the convenient usability and high temporal resolution compared to other brain imaging equipment such as functional magnetic resonance imaging (fMRI) and magnetoencephalogram (MEG), research of noninvasive electroencephalogram (EEG) based brain–computer interface (BCI) systems is continuously progressed [1], [2], [3].
In the beginning of BCI research, BCI systems have been developed mostly to provide alternative communication means to people who have severe motor disabilities [2], [4], [5]. Recently, much research effort focused on development of portable BCI systems for normal person by using headset shaped scalp electrodes [6], [7] and also dry electrodes which do not need conductive gel for preparation of EEG recording [8], [9]. In addition, with the progress of portable BCI systems and EEG sensor technologies, many BCI applications are developed for general public [9], [10]. However, for the BCI systems going beyond laboratory researches, the most important issue is stable classification performance.
Normally, EEG based BCI experiment can be categorized as a training (calibration) stage and a real time testing (feedback) stage. In the training stage, translation algorithm such as classification is designed using collected training signals. Then, an application device such as neural prosthesis is controlled by using the classification algorithm in real time testing stage. However, EEG signals have inherent non-stationary characteristics and there exist significant day-to-day and even session-to-session variability [12], [27], [29]. Thus, features of experimental EEG signals are changed from the offline training sessions to online testing sessions [11]. Due to this, classification performance is unavoidably deteriorated in BCI experiment with time. In addition, the training session (15–35 min) is conventionally carried out every time before using the BCI systems even for experienced subjects [12]. These are major obstacles of real-time online BCI applications.
To overcome the performance decrease caused by the non-stationarity of EEG signals, many adaptive signal processing methods are proposed. In [27], [28], [29], adaptive feature extraction methods are proposed for the motor imagery based BCI systems. For the adaptive classification scheme, in [13], mean and covariance matrix of a statistical classifier are iteratively updated using each class data. The study [11] proposes a bias adaptation scheme of linear discriminant analysis (LDA) classification using class labels of several test trials. They have shown that simple bias adaptation is effective for online test data. In [14], they propose an expectation-maximization (EM) algorithm based unsupervised adaptive classification method. Using EM algorithm, common spatial pattern (CSP) features are re-extracted and parameters of Bayes classifier are updated in each iteration step. Similarly, [15] suggest unsupervised bias adaptation of LDA without using class label information. Previous studies for adaptive classification method need classifier re-adjustment (training) such as parameters and bias adaptation for new test trials. However, for this re-training, additional computation is needed in each update (adjustment) step.
Recently, with much progress of L1 minimization technique in compressive sensing field [21], [22], sparse representation has received a lot of attention in signal processing and pattern recognition fields. Especially, sparse representation based classification (SRC) has shown an increased interest [16], [23], [24]. SRC framework is first introduced by Huang et al [16]. A test data from one class is predominantly represented by the same class training data from dictionary. The dictionary is composed by all class training data and usually underdetermined. Sparse representation of the test data using the dictionary can effectively be solved by the L1 minimization tool, and the classification is performed by comparing the representation error for each class.
SRC have been also studied for EEG signal classification [17], [18], [25]. In [18] and [25], SRC scheme is applied to vigilance detection and epileptic seizure detection problem respectively. In addition, SRC scheme is first introduced for motor imagery based BCI application in [17]. They have shown that the SRC exhibits better classification performance than the conventional LDA method using two experimental datasets. Another study [31] also revealed that the SRC shows better classification accuracy and noise robustness than the well-known SVM method. However, no research has been studied for adaptive SRC scheme for online BCI applications.
Compared to other fixed decision rule based classification method such as linear discriminant analysis (LDA) and support vector machine (SVM), in the SRC, the sparse representation is adaptively performed for each test data by utilizing all training data in the dictionary. Along with this inherent adaptive characteristic of the SRC, in this study, we propose simple adaptive SRC schemes for real-time BCI applications. We suggest a dictionary update rule and an incoherence based dictionary modification (IDM) method. For the dictionary update rule, supervised and unsupervised adaptive schemes and also accumulated and fixed update rules are considered. Proposed dictionary update methods are very simple and additional computation for adaptation is not needed. In the part of IDM method, our aim is to create a maximally incoherent dictionary via an incoherence measure of training data. This method is applied to the training data before performing the sparse representation. Using online motor imagery based BCI experimental datasets, we evaluate classification performance of the proposed adaptive method by comparing with the conventional SRC and other adaptive classification methods.
This paper is organized as follows. In Section 2, our experiment and dataset are explained. In Section 3, technical methods such as feature extraction, sparse representation based classification (SRC) method and proposed adaptive SRC schemes are introduced. We explain experimental evaluation strategy and results in Section 4. In Section 5, we discuss some experimental results. Finally, we conclude the paper in Section 6.
Section snippets
Experiment
For evaluation of adaptive classification scheme, we performed online motor imagery based BCI experiment. The experiment was approved by the Institutional Review Board of Gwangju Institute of Science and Technology. Ten subjects who signed a written informed consent letter participated in our online experiment. The experiment was performed on multiple days (two or three days). In each day, just one session experiment was executed. The number of sessions for each subject was determined by
Preprocessing and feature extraction
For preprocessing of experimental EEG dataset, we apply same procedures to all datasets and classification methods. First, we perform band pass filtering to eliminate the frequencies which are not related to motor imagery signals. In this study, we use fourth order Butterworth filter with 5 and 30 of cut off frequencies.
EEG signals are very noisy and have poor spatial resolution. Thus, an electrode placed on the scalp measures the EEG signals generated not only from the motor cortex area but
Evaluation strategy
Using the online experimental dataset, we aim to evaluate proposed adaptive SRC schemes, i.e., four dictionary update methods (supervised accumulated update (SAU), supervised fixed update (SFU), unsupervised accumulated update (UAU) and unsupervised fixed update (UFU) rule) and an incoherence based dictionary modification (IDM) method. From the multi session datasets of 10 subjects, 12 session datasets are selected for evaluation of proposed methods. In this selection, for a reliable assessment
Experimental results
To evaluate classification performance of the proposed adaptive SRC schemes, we compare classification accuracy (%) of proposed methods with that of conventional SRC method using the online experimental dataset of 12 motor imagery sessions. Table 1 shows the classification accuracy of the SRC and the proposed dictionary update based SRC methods with and without IDM method. For fair comparison, we set the same value of n (the number of elimination trials of IDM) of 10 for all subjects and all
Results for public dataset
For the evaluation of the proposed methods, we use a public dataset obtained from Dataset IVc of BCI Competition III [32]. In this dataset, the test data were separately recorded for more than 3 h after the acquisition of the training data. Therefore, the distribution of some EEG features could be effected by non-stationarities. This dataset was recorded from a healthy subject. He sat in a comfortable chair with his arms resting on the armrests. The training dataset consists of the data of the
Conclusion
Because of the inherent non-stationarity of EEG signals, performance degradation is an inevitable phenomenon in EEG based BCI systems. In particular, an already designed classifier by the training data does not guarantee satisfactory classification accuracy for new test data in the online feedback stage. In this paper, we propose dictionary update methods with incoherence based dictionary modification (IDM) as adaptive SRC schemes to compensate for the non-stationary effects. We consider
Conflict of interest statement
The authors declare that there is no conflict of interests regarding the publication of this article.
Acknowledgments
This work was supported by the National Research Foundation (NRF) of Korea Grant funded by the Korean government (NRF-2015R1A2A1A05001826).
References (32)
- et al.
Brain–computer interfaces for communication and control
Clin. Neurophysiol.
(2002) - et al.
An EEG-based brain–computer interface for cursor control
Electroencephalogr. Clin. Neurophysiol.
(1991) - et al.
Brain–computer interface – a new communication device for handicapped persons
J. Microcomput. Appl.
(1993) - et al.
A P300-based brain–computer interface: Initial tests by ALS patient
Clin. Neurophysiol.
(2006) - et al.
Toward Brain–Computer Interfacing
(2007) - ...
- ...
- et al.
Dry and noncontact EEG sensors for mobile brain–computer interfaces
IEEE Trans. Neural Syst. Rehabil. Eng.
(2012) - et al.
Gaming control using a wearable and wireless EEG-based brain–computer interface device with novel dry foam-based sensors,
J. Neuroeng. Rehabil.
(2012) - et al.
Brain computer interface-based smart living environmental auto-adjustment control system in UPnP home networking
IEEE Syst. J.
(2014)
Towards adaptive classification for BCI
J. Neural Eng.
Optimizing spatial filters for robust EEG single-trial analysis
IEEE Signal Process. Mag.
An extended EM algorithm for joint feature extraction and classification in brain–computer interfaces
Neural Comput.
Toward unsupervised adaptation of LDA for brain–computer interfaces
IEEE Trans. Biomed. Eng.
Sparse representation for signal classification
Adv. Neural Inf. Process. Syst.
Cited by (33)
A comprehensive review of the movement imaginary brain-computer interface methods: Challenges and future directions
2022, Artificial Intelligence-Based Brain-Computer InterfaceSparse representation-based classification with two-dimensional dictionary optimization for motor imagery EEG pattern recognition
2021, Journal of Neuroscience MethodsCitation Excerpt :Recently, sparse representation-based classification (SRC) began to be applied to neural information processing and achieved more superiority results than traditional classifiers. Its basic theory is to use the interpretability of sparse representation to make a discriminant classification, that is, an over-complete dictionary is constructed from the original dataset containing all the training samples, and then the over-complete dictionary is used to sparsely represent the test samples (Shin et al., 2015a,b, 2012). Therefore, the construction of dictionary directly affects the classification performance of SRC.
Classification of multiclass motor imagery EEG signal using sparsity approach
2019, NeurocomputingCitation Excerpt :However, the computational time still need to be reduced for real-time BCI applications. In fact, the performances of the existing SRC methods for MI-based BCIs [10,11,37,40,41,43] are solely governed by either band-pass or CSP filtering. In this work, a new sparsity-based framework for multiclass MI EEG signals has been introduced to classify different MI tasks accurately in a lesser time.
Nonparametric kernel sparse representation-based classifier
2017, Pattern Recognition LettersCitation Excerpt :SRC has found various applications in which non-stationary signals were investigated. In EEG-based brain computer interfaces, SRC-based methods provided better performance with respect to the other approaches from both classification and robustness points of view [20,21]. Moreover, in different disease assessment problems like tongue geometric feature analysis [24], SRC was utilized for distinguishing between healthy and disease patterns.
Recognition of emotions using multimodal physiological signals and an ensemble deep learning model
2017, Computer Methods and Programs in BiomedicineCitation Excerpt :In addition, the multimodal PNS physiological signals, e.g., galvanic skin response (GSR) [31], electrooculargram (EOG) [32], electromyogram (EMG) [33], and electrocardiogram (ECG) [34], were extensively explored. Considering high spatial and temporal resolutions of sophisticated CNS and PNS signal acquisition devices, machine learning approaches facilitate analyzing the massive volume of neurophysiological data [35–39]. In particular, pattern classifiers could fuse physiological features of different modality.
Cross-session classification of mental workload levels using EEG and an adaptive deep learning model
2017, Biomedical Signal Processing and ControlCitation Excerpt :The comparison between the adaptive SDAE and the 8 classical MW classifiers for the cross-session case implies the former is quite promising to tackle the data distribution variation between the training and testing EEG features. The reason behind is that the decision function of a static classifier is invariant [41–43] with the time while the weights in the shallow layer of the adaptive SDAE could be iteratively tuned to learn the novel statistical information from the estimated testing EEG samples. To give an insight of hierarchical representations for the high-dimensional EEG power features between the adaptive and the static deep model, three-dimensional activations for each hidden layer of the final adaptive SDAE and the standard SDAE that are specifically built for the participant A are visualized in Fig. 16.