Motor imagery EEG classification based on flexible analytic wavelet transform

https://doi.org/10.1016/j.bspc.2020.102069Get rights and content

Highlights

  • A novel classification system for MI-EEG signals is proposed based on flexible analytic wavelet transform (FAWT).

  • Feature dimensions is reduced by MDS, which is seldom applied in MI-EEG recognition.

  • The propsoed method is an automatic and simple recognition method for LH and RH MI-EEG signals, and it achieves a better trade-off between classification performance and time consuming.

Abstract

Motor imagery electroencephalogram (MI-EEG) based brain-computer interface (BCI) is a burgeoning auxiliary means to realize rehabilitation therapy. One of the major concerns in MI-EEG based BCI is to have an accurate classification, and effective and fast feature extraction is the key to build a successful MI-EEG classification model. In this paper, a novel classification system for MI-EEG signals is proposed based on flexible analytic wavelet transform (FAWT). The filtered MI-EEG signals are firstly subjected to the FAWT to obtain sub-bands, and time-frequency features are calculated from the sub-bands. MDS is then adopted to reduce the dimension of the extracted features, and principal component analysis (PCA), kernel principal component analysis (KPCA), locally linear embedding (LLE) and Laplacian Eigenmaps (LE) are utilized as comparison. Finally, linear discriminant analysis (LDA) is utilized to complete the classification of left-hand (LH) and right-hand (RH) MI-EEG signals. The proposed method is experimentally validated on BCI Competition II Data Set III (BCI Dataset III) and BCI Competition III Data Set IIIb (BCI Dataset IIIb). As a result, the combined method of FAWT, MDS attains the maximal mutual information (MaI) of 0.95 and the maximum accuracy (ACC) of 94.29% using BCI Dataset III, and the mean of the maximal MaI steepness of 0.3740 using BCI Dataset IIIb. The proposed method yields better performance in comparison to the existing methods. Overall, the effectiveness of the proposed approach suggests that it can be a worthwhile and promising method for a MI-EEG based BCI system.

Introduction

Brain-computer interface (BCI) is a state-of-the-art technology that builds a direct information transmission pathway between the human brain and the outside world [[1], [2], [3]]. As an all-important brain-computer interaction strategy, motor imagery-based BCI is characterized by controlling the external devices through the imagination of a specific movement in the brain rather than the actual execution of the movement [4]. This interactive strategy has been successfully applied to control robot devices, which lays the foundation for the disabled people to communicate with the real world [5].

Electroencephalogram (EEG) is a bioelectric signal generated by the electrical activity of brain neurons, which covers a wealth of physiological and pathological information and can be closely related to the state of consciousness [6]. EEG is the best and the most widely applied technique in a BCI system to evaluate the motor imagery tasks, owing to its merits of non-invasiveness, high temporal resolution and low-cost devices [7,8]. Thus, the motor imagery EEG (MI-EEG) recorded from patients are recognized using a classification model (including preprocessing, feature extraction, feature selection and classification) to distinguish different types of motor imagery tasks. However, conclusions have drawn that MI-EEG are the time series signals with non-stationary, non-linear and low signal-to-noise ratio (SNR) [9]. In order to achieve a successful classification of MI-EEG signals, the accurate and quick extraction of MI-EEG features is indeed and is also the key point of a classification model.

In addition to the above characteristics of MI-EEG signals, MI-EEG signals are also time-varying sensitive in nature. Hence, the time-frequency analysis methods are the most widely employed in MI-EEG recognition. Wavelet transform (WT), which is an advanced time-frequency representation method, arouses researchers’ enormous interests and attention owing to its abilities of multiscale decomposition and obtaining the instantaneous information of EEG signal. Various WT approaches including discrete wavelet transform (DWT) [[10], [11], [12], [13],47], wavelet packet transformation (WPT) [14,48], dual-tree complex wavelet transform (DTCWT) [15], tunable-Q wavelet transform (TQWT) [16,45] have been applied for MI-EEG classification. In [10], time-frequency features computed from the detail components of DWT were cooperated with the non-linear features obtained by parametric t-distributed stochastic neighbor embedding (P. t-SNE) to identify left-hand (LH) and right-hand (RH) MI-EEG signals. In [11], DWT decomposed the MI-EEG signals into sub-bands which cover the rhythms related to event-related desynchronization (ERD) and event-related synchronization (ERS), and features were extracted from the sub-bands for MI-EEG signals. In [12], DWT was employed to extract statistical measures to discriminate the MI-EEG signals and principal component analysis (PCA) was used to reduce the dimension of the feature vector, a good performance with high accuracy was achieved using the reduced feature vector. In [13], the energy distribution was estimated by computing the detailed and approximated coefficients of DWT to select the most appropriate band, and then the power spectral density of the optimal frequency bands was considered as the features for four-class MI-EEG tasks. In [47], DWT was coordinated with Hilbert transform to extract significant features from LH and RH MI-EEG signals. In [14], the applicability and effectiveness of various wavelet basis functions were investigated in WPT domain for feature extraction of MI-EEG signals. In [48], WPT combined with spectrum analysis (SC) was conducted on MI-EEG signals to extract features and followed by MI-EEG recognition. In [15], the time, frequency, and phase features were explored in the DTCWT domain for classifying the LH and RH MI-EEG tasks. In [16], TQWT domain based activity, clearance factor, range and mean parameters were utilized as the input features to least-squares support vector machines (LS-SVM) for right-hand and right-foot MI-EEG classification. In [45], the LH and RH MI-EEG signals were differentiated using TQWT and linear discriminant analysis (LDA).

Recently, a novel time-frequency analysis method termed flexible analytic wavelet transform (FAWT) has been introduced to analyze the oscillatory signals [17]. In FAWT, the quality-factor (QF), dilation factor and shift parameters can be adjusted flexibly that other wavelet transforms like DWT, DTCWT and TQWT cannot compare favorably with it [17]. Hence, the merits of FAWT make it a feasible approach for analysis of bio-medical signals like EEG, and the applications of FAWT on EEG signals have been successfully investigated including sleep stages classification [18], epileptic seizures [19], emotion recognition [20] and alcoholism detection [21]. In [18], a FAWT-based framework was proposed to extract time domain measures of each band-limited sub-band for the EEG signals of various sleep stages, and the best classification accuracies for two to six sleep stages classification are 97.56%, 96.48%, 96.39%, 96.03%, respectively. In [19], the complex-valued distribution entropy was explored in FAWT domain to discriminate epileptic EEG signals and a good performance with high accuracy was obtained. In [20], the emotion EEG signals were decomposed into sub-bands using FAWT, information potential was subsequently calculated from the sub-band signals, which are tested on random forest and support vector machine (SVM) for various emotion recognition. In [21], the FAWT domain extracted wavelet coefficients were employed for computing statistical features and Shannon entropy, which classify the alcoholic and nonalcoholic EEG signals and 99.17% accuracy was yielded. To the best of our knowledge, FAWT is rarely explored in MI-EEG classification tasks. Considering the FAWT’s advantaged of good shift-invariance, tunable oscillatory bases and flexible time-frequency partition manner and the satisfactory performance in analyzing EEG signals, a FAWT-based feature extraction of MI-EEG signals is proposed in this paper. MI-EEG signals are firstly decomposed into multiple sub-bands using FAWT. Five statistical parameters, including the mean energy (TE¯), the absolute mean value (Taav), the standard deviation (Tstd), the volatility index (Tcov) and the center frequency (Pcf), are then extracted from the reconstructed sub-bands. Since the high dimensions of the extracted features, multidimensional scaling (MDS) is subsequently conducted on the features which can not only reduce the feature dimension but also increase the classification performance. Then the reduced features were test on LDA that possesses the merits of low computational complexity, stability and no parameter adjustment.

The objective of this study is to enjoy a two-class MI-EEG classification model to extract more effective features and thereby improving the classification performance. The main contributions of the proposed method are as follows:

  • (1)

    FAWT is adopted to capture the time-frequency information of the raw MI-EEG signals. There are hardly research reports on the use of FAWT for MI-EEG recognition. Compared with other wavelet transform methods, FAWT possesses desirable properties of flexible time-frequency covering, better shift-invariance and tunable oscillatory nature of the bases.

  • (2)

    MDS is utilized to reduce the feature dimensions to enhance the discriminative capability of features and improve the classification performance, and comparative experiments are conducted with other dimension reduction methods.

  • (3)

    A combination method of FAWT, MDS and LDA is introduced for classification of LH and RH MI-EEG signals. We show that the proposed method is superior to the existing methods.

The subsequent sections of this paper are organized as follows: In Section 2, the adopted EEG datasets are briefly introduced and the details of the proposed method are then described. Subsequently, experimental results and discussion are presented in Section 3 and 4, respectively. Finally, the conclusion of the whole work is summed up in Section 5.

Section snippets

Dataset description

To verify the feasibility of the proposed method, two public BCI EEG datasets released by BCI Competition II and III are utilized [22,23]: One is the Dataset III of BCI Competition II (BCI Dataset III) and the other is the Dataset IIIb of BCI Competition III (BCI Dataset IIIb). Both of the datasets were recorded over channels C3, Cz and C4 from the test subjects. The electrode positions are shown in Fig. 1.

BCI Dataset III [22] contains two-class MI-EEG signals which were recorded during

Results

In this paper, the LH and RH MI-EEG signals are first decomposed into sub-band signals using a five-level FAWT. Then, based on extensive experiments, features extracted and selected at every sampling point with 3 s sliding rectangular window size lead to the optimal classification results. For BCI Dataset III, the sliding window starts at t = 3 s and ends at t = 7 s, while it begins at t = 4 s and ends at t = 9 s or 8 s for BCI Dataset IIIb. The plots of sub-band signals of C3, C4 electrode

Discussion

Many researchers concentrate on developing strategies for the recognition of MI-EEG signals. Hence, in order to appraise the usefulness of proposed framework for two-class MI-EEG signals classification, a comparison between the proposed technology and the existing methods based on the same datasets are discussed in this section. Previous studies based on the BCI Data III and BCI Data IIIb are listed in Table 6, Table 7, respectively.

For BCI Data III [22], Morlet-wavelets, AR and AAR were

Conclusion

In this paper, a novel framework of FAWT-based MDS is put forward for the recognition of LH and RH MI-EEG signals. The proposed method well combines the advantages of FAWT and MDS, capturing the hidden information, and achieving feature dimension reduction of extracted features thereby extracting the most informative features in MI-EEG. Through the visualization, MDS achieves better feature dimension reduction than other DimR techniques including PCA, KPCA, LLE and LE. After evaluating our

CRediT authorship contribution statement

Yang You: Methodology, Software, Investigation, Writing - original draft, Writing - review & editing. Wanzhong Chen: Conceptualization, Supervision, Data curation. Tao Zhang: Validation, Software.

Acknowledgments

This work is supported by the Science and technology development project of Jilin Province (Grant No.20190302034GX), the Fundamental Research Funds for the Central Universities (Grant No. 451170301193) and the Natural Science Foundation for Science and Technology Development Plan of Jilin Province, China (Grant No. 20150101191JC).

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (50)

  • M. Kumar et al.

    Characterization of coronary artery disease using flexible analytic wavelet transform applied on ECG signals

    Biomed. Signal Process. Control

    (2017)
  • M. Kumar et al.

    Automated diagnosis of atrial fibrillation ECG signals using entropy features extracted from flexible analytic wavelet transform

    Biocybern. Biomed. Eng.

    (2018)
  • Z.J. Tang et al.

    Robust image hashing with multidimensional scaling

    Signal. Process.

    (2017)
  • H. Zhou et al.

    Age-invariant face recognition based on identity inference from appearance age

    Pattern. Recog.

    (2018)
  • S. Taran et al.

    Features based on analytic IMF for classifying motor imagery EEG signals in BCI applications

    Measurement.

    (2018)
  • Y. Wang et al.

    Research of EEG recognition algorithm based on motor imagery

    Proc. 2nd Int. Conf. Robot. Autom. Sci. (ICRAS)

    (2018)
  • M. Chen et al.

    Phase space reconstruction for improving the classification of single trial EEG

    Biomed. Signal Process. Control

    (2014)
  • Y. Fang et al.

    Extracting features from phase space of EEG signals in brain-computer interfaces

    Neurocomputing.

    (2015)
  • P.J. Benson

    Decoding brain-computer interfaces

    Science.

    (2018)
  • H. Yuan et al.

    Brain-computer interfaces using sensorimotor rhythms: current state and future perspectives

    IEEE Trans. Biomed. Eng.

    (2014)
  • S. Ge et al.

    A double-partial least-squares model for the detection of steady-state visual evoked potentials

    IEEE J. Biomed. Health Inform.

    (2016)
  • M. Hamedi et al.

    Electroencephalographic motor imagery brain connectivity analysis for BCI: a review

    Neural Comput.

    (2016)
  • Z. Qiu et al.

    Optimized motor imagery paradigm based on imagining Chinese characters writing movement

    IEEE Trans. Neural Syst. Rehabil.

    (2017)
  • M. Li et al.

    Applying a locally linear embedding algorithm for feature extraction and visualization of MI-EEG

    J. Sensors

    (2016)
  • S. Ge et al.

    Sinusoidal signal assisted multivariate empirical mode decomposition for brain-computer interfaces

    IEEE J. Biomed. Health Inform.

    (2017)
  • Cited by (0)

    View full text