Subject-based dipole selection for decoding motor imagery tasks

Highlights

• A novel subject-based dipole selection method named PRDS is proposed for decoding motor imagery tasks in the source space.

• PRDS includes two steps: the preliminary selection of dipoles with the data-driven method and the refinement of dipoles based on continuous wavelet transform.

• The dipoles selected by PRDS are fully activated and can best reflect the differences among the multiclass MI-tasks, meanwhile have good adaptability for different subjects.

• The wavelet coefficient power of the selected dipoles are input to the OVO–CSP to extract the fusion features of time-frequency-spatial domains, thus yielding better decoding performance and calculation efficiency for the four classes of MI-tasks.

Abstract

In the BCI rehabilitation system, the decoding of motor imagery tasks (MI-tasks) with dipoles in the source domain has gradually become a new research focus. For complex multiclass MI-tasks, the number of activated dipoles is large, and the activation area, activation time and intensity are also different for different subjects. The means by which to identify fewer subject-based dipoles is very important. There exist two main methods of dipole selection: one method is based on the physiological functional partition theory, and the other method is based on human experience. However, the number of dipoles that are selected by the two methods is still large and contains information redundancy, and the selected dipoles are the same in both number and position for different subjects, which is not necessarily ideal for distinguishing different MI-tasks. In this paper, the data-driven method is used to preliminarily select fully activated dipoles with large amplitudes; the obtained dipoles are refined by using continuous wavelet transform (CWT) to best reflect the differences among the multiclass MI-tasks, thereby yielding a subject-based dipole selection method, which is named PRDS. PRDS is further used to decode multiclass MI-tasks in which some representative dipoles are found, and their wavelet coefficient power is calculated and input to one-vs.-one common spatial pattern (OVO-CSP) for feature extraction, and the features are classified by the support vector machine. We denote this decoding method as D-CWTCSP, which enhances the spatial resolution and also makes full use of the time-frequency-spatial domain information. Experiments are carried out using a public dataset with nine subjects and four classes of MI-tasks, and the proposed D-CWTCSP is compared with the relevant methods in sensor space and brain-source space in terms of the decoding accuracy, standard deviation, recall rate and kappa value. The experimental results show that D-CWTCSP reaches an average decoding accuracy of 82.66% among the nine subjects, which generates 8–20% improvement over other methods, thus reflecting its great superiority in decoding accuracy.

Introduction

The brain-computer interface (BCI) is a popular technology that achieves external communication and control [1,2]. The BCI does not rely on the peripheral nerve, muscle tissue or other conventional brain information output pathways but uses engineering techniques to establish connections between the brain and the computer or other external devices. As a promising rehabilitation technology, the BCI aims to improve the life quality of disabled people with dyskinesia [3,4]. There are many kinds of biological signals such as electroencephalogram (EEG), magnetoencephalogram (MEG), visual evoked potential (VEP) and functional magnetic resonance imaging (fMRI), and they are commonly used in BCI systems. Especially in active non-invasive BCIs, motor imagery EEG (MI-EEG), which has the characteristics of convenient acquisition, low price and high time resolution, can directly reflect the subject's motion intent, it is often decoded and the decoding result is employed to drive the robot arm or functional electrical stimulation instrument, realizing the active and intelligent motor function rehabilitation [5], [6], [7].

However, MI-EEG suffers from a low signal-to-noise ratio (SNR) and poor spatial resolution, which limits its application in the field of BCI. Increasing numbers of researchers use the spatial analysis method to decode MI-tasks in the sensor space, such as the common average reference (CAR), common spatial subspace decomposition (CSSD), and the common spatial pattern (CSP). The CSP is considered to be one of the most effective approaches to extracting the features of EEG, which is calculated by using spatial filters, to maximize the variance of one class and to minimize the variance of the opposite class [8]. Ramoser et al. demonstrated that the spatial filter of multichannel EEG calculated by the CSP can effectively extract different information from the two MI-tasks for the first time in 2000 [9]. In the 20 years that followed, several variants of the classical CSP algorithm were extended, including the filter bank common spatial pattern (FBCSP) algorithm that combines frequency band decomposition with the CSP [10], and the regularized common spatial pattern (RCSP) [11] with its derivative algorithm to solve the problem of overfitting with small training sets by adding regularization terms [13], [14], [15]. Generally, the CSP is directly applied to the preprocessed EEG signal in the sensor space and performs well in two-class MI-tasks whose active regions are non-overlapping (such as left and right limbs tasks) [16,17]. However, for complex multiclass MI-tasks, some of the active regions in the cerebral cortex are diffuse and similar. Due to the limited number of scalp electrodes, the spatial resolution of MI-EEG is lower, and the features after CSP-based filtering cannot obtain ideal decoding results.

EEG source imaging (ESI) refers to a commonly used technique for estimating the source distribution of the equivalent dipole on the cerebral cortex by solving an inverse problem with the EEG data recorded on the scalp [18], [19], [20]. A forward problem model based on prior physiological and anatomical information is firstly established, and then the mapping relationship (lead field matrix) between the scalp electrode recording and the electrical activity on the cerebral cortex is obtained in order to trace electrical activity on the scalp back to neuroanatomical structure. The electrical activity of the gyral and sulcal folds of the brain is reconstructed by modeling populations of cells as equivalent dipoles [21,22] or current density distributions [23,24] with solving an ill-posed inverse problem. With the rapid strides in neurobiology and artificial intelligence, ESI could bring the research of EEG from sensor space to source space and further promotes the development of biomedicine and brain science.

In recent years, some researchers have combined ESI with the classical CSP algorithm to decode MI-EEG in order to improve the classification accuracy [25], [26], [27], [28]. In 2010, the Beamformer algorithm was used to transfer left-right hand MI-EEG to the brain-source space. The regions of interest (ROI) were covered by electrodes C3 and C4. The CSP was applied to the feature extraction of the dipoles in the ROI, and Fisher's linear discriminant analysis (FLDA) was used for classification. The highest recognition rate reached 90%, which is much higher than that of the MI-EEG in sensor space [25]. In [26], Handiru et al. extended the classical CSP to the source space by using various regularization approaches. The weighted minimum norm estimate (wMNE) was selected as the ESI algorithm. Dipoles on the left and right dorsal premotor cortex were selected for feature extraction and classification according to the physiological functional partition theory. For a multidirectional hand movement dataset, the source space features can result in an increase of over 10% accuracy compared to the sensor space features. In [27], a decoding method that combined the wMNE algorithm with the one-vs.-rest common spatial pattern (OVR-CSP) was developed. Dipoles distributed in the parietal lobe region around electrode Cz were selected as the ROI, and LDA was used for classification. The average classification accuracy for four classes of MI-tasks reached 74.6% on the BCI Competition IV dataset 2a, which was nearly 8% higher than that of the competition winner. For the convenience of comparison, this method is denoted as wMNE-CSP1 in this paper. Based on the same dataset, Xygonakis et al. [28] applied the OVR-CSP to the cerebral cortex dipole time series and the LDA, based on a voting mechanism, was used as the classifier. Sixteen symmetrical regions on the left and right hemispheres were selected as ROIs based on the Brodmann partition theory and their classification model accuracy. This method achieved the highest classification recognition rate of 75%. Once again, for convenience, we denoted this method as wMNE-CSP2 in this paper.

The above studies provide a new concept for the development of BCI technology. However, there seems to be no in-depth study on the selection of dipoles in the cerebral cortex. There are two main methods for selecting dipoles: one is based on the physiological functional partition theory, such as the commonly used Brodmann theory [26,28], and the other is based on human experience [25,27]. The selected dipoles constitute the ROIs, defining a feasible range for the following feature extraction and pattern classification. These methods have worked well in decoding MI-tasks but still have some limitations. The number of selected dipoles is still large, and the decoding accuracy is hindered by feature information redundancy. In addition, for different subjects, even for the same MI-task, the dipole activation (number, location, time and intensity of activation) is different. If the dipoles are selected according to a unified partition standard, the performance of feature extraction and classification will be affected. Moreover, the time-frequency characteristics of the dipoles are not taken into account, so the features of the selected dipoles are not necessarily the most favorable to distinguish complex MI-tasks.

In this paper, a subject-based dipole selection method is proposed. This method innovatively performs two rounds of selection for dipoles: the fully activated dipoles are selected on the basis of amplitude by the data-driven method; then, these preliminary-selected dipoles are refined by a continuous wavelet transform (CWT) to obtain the subject-based optimal dipoles. The “preliminary selection + refinement” dipole selection method is denoted as PRDS. The proposed PRDS is applied to decode complex multiclass MI-tasks. To make better use of the time-frequency characteristics of the dipoles, the selected dipoles’ wavelet coefficient power is input into a one-vs.-one common spatial pattern (OVO-CSP) for feature extraction, and a support vector machine (SVM) is used to classify the features. The “CWT + OVO-CSP” based decoding method is called as D-CWTCSP. Many comparative experiments are conducted on a public dataset with 10  ×  10-fold cross validation (10  ×  10-fold CV) between D-CWTCSP and some related methods in the sensor and source spaces. In addition, we compare the influence of the classical CSP and its derivative algorithms on decoding effect. The results showed that D-CWTCSP reaches preferable average classification accuracy, which suggests that our method has obvious advantages.

The structure of this paper is organized as follows. In Section 2, the basic principles of standardized low-resolution electromagnetic tomography (sLORETA) and CWT are introduced. Section 3 describes the workflow of the proposed PRDS and D-CWTCSP in detail. In the next section, a large number of comparative experiments and results are presented. Sections 5 and 6 are the discussion and the conclusion of this paper, respectively.