Keywords

1 Introduction

Stroke is one of the most common cerebrovascular diseases. It is the leading cause of serious and long-term disability worldwide [1,2,3]. A major consequence of stroke is impairment of motor function, often including hemiplegia of the upper limbs. After rehabilitation therapy, some stroke survivors can partially regain their motor control, but most survivors are left with permanent motor disability, which may affect their ability to work, drive, type and conduct other daily activities [4].

At present, treatments to improve upper limb function may be divided into two categories: peripheral nerve intervention and of central nerve intervention. The methods of peripheral nerve intervention include specialized exercises, sensory integration training, and functional electrical stimulation (FES). The methods of central nerve intervention, which are much more experimental, include brain computer interfaces (BCIs), repetitive transcranial magnetic stimulation (rTMS), and transcranial direct current stimulation (tDCS).

Brain computer interface (BCI) systems have gained considerable attention in the research literature recently. BCI systems provide a new communication method for people with severe neuromuscular disabilities [5,6,7]. Various neurological phenomena can be used in BCI systems. Steady state visual evoked potentials (SSVEP) [8, 9], slow cortical potentials, P300 evoked potentials [10,11,12,13,14,15] and event-related desynchronization (ERD) [16]/event-related synchronization (ERS) [17] are the most prominent approaches in BCI systems.

Even if persons with stroke can no longer move an affected limb, they can still imagine moving it. Research has shown that motor imagery based BCIs can help induce brain plasticity in stroke survivors [1,2,3, 18]. It is well established that imagination of limb movement could result in event-related desynchronization (ERD) and event-related synchronization (ERS). Specifically, movement imagery (MI) affects the mu (8–12 Hz) and beta waves (13–30 Hz) in the EEG. Movement imagery produces significant ERD over the contralateral central area during imagination of right and left hand movement [19]. However, as stroke patients suffer neurological damage, the brain regions associated with motor function might be compromised [4]. This neurological damage could impair the ability to imagine movement. This study will explore any differences in BCI accuracy when classifying imagined movements of the paretic versus non-paretic hands.

2 Methods

2.1 Patients

Ten stroke patients (5 male, aged 21–69 years, mean 48.4 ± 15.4) participated in this study. Five of them, designated patients 1–5, have left hemiplegia, while patients 6–10 have right hemiplegia. All patients signed a written consent form prior to this experiment. The local ethics committee approved the consent form and experimental procedure before any patients participated. All patients were right handed according to self-reports. All subjects’ native language was Mandarin Chinese.

2.2 Data Acquisition

In this study, EEG signals were sampled at 256 Hz through the g.USBamp (Guger Technologies, Graz, Austria). The band pass filter was set to 0.1 Hz–30 Hz. A 16-channel cap (FC3, FCZ, FC4, C5, C3, C1, CZ, C2, C4, C6, CP3, CP1, CPZ, CP2, CP4 and PZ) following the 10–20 international system was used for signal recording. Data were referenced to electrode REF located over the right mastoid with a forehead ground (GND), shown in Fig. 1.

Fig. 1.
figure 1

The electrode distribution used in this study.

Full size image

2.3 Experimental Paradigms

After being prepared for EEG recording, the patients were seated in a comfortable chair in a shielded room. During data acquisition, patients were asked to relax and avoid unnecessary movement. The experimenter informed the patients that they would hear cues over a speaker that would instruct them to imagine moving either the left or the right hand. Figure 2 shows that each trial lasts eight seconds and starts with a warning “beep”. Two seconds later, the cue (the command to imagine a left or right hand movement) is played. Patients were asked to imagine whatever type of hand movements that they felt was easiest for them. Six second later, a “relax” command is played, informing patients that the trial is over. Each patient participated in sixty trials within one recording session.

Fig. 2.
figure 2

Timing of a trial of the motor imagery paradigm. Each trial consisted of task and rest periods. Patients started to execute motor imagery tasks while the cue (“left” or “right”) was played.

Full size image

2.4 Feature Extraction

The EEG data were band-pass filtered using a third order Butterworth band pass filter from 8 to 30 Hz, since this frequency band included the range of frequencies that are most relevant in classifying motor imagery with EEG data.

The CSP algorithm is an effective feature extraction algorithm. It has been widely used in processing EEG data from motor imagery [20,21,22,23,24,25]. CSP is based on the simultaneous diagonalization of two covariance matrices. It finds a spatial filter to maximize variance for one class and minimize variance for another class at the same time to improve classification.

For the analysis, the original EEG signal data of a single trial is represented as a matrix \( \text{E}_{N \times T} \), where \( N \) is the number of channels, and \( T \) is the number of sampling points for each channel. The CSP algorithm process is as follows:

Calculate spatial covariance of the EEG data:

$$ C = (EE^{T} )/(tr(EE^{T} )) $$
(1)

\( C_{l} \) and \( C_{r} \) represent two spatial covariance matrixes (two classes of motor imagery). The composite spatial covariance matrix is \( C_{c} \).

\( C_{c} \) can be decomposed as:

$$ C_{c} = U_{c}\uplambda_{c} U_{c}^{T} $$
(2)

\( U_{c} \) is the matrix of eigenvectors and \( \uplambda_{c} \) is the diagonal matrix of eigenvalues. In the process, the eigenvalues are arranged in descending order.

After whitening transformation:

$$ P = \sqrt {\uplambda_{c}^{ - 1} } U_{c}^{T} $$
(3)

Then \( \overline{{C_{l} }} \) and \( \overline{{C_{r} }} \) can be transformed into:

$$ S_{l} = P\overline{{C_{l} }} P^{T} ,S_{r} = P\overline{{C_{r} }} P^{T} $$
(4)

\( S_{l} \) and \( S_{r} \) share the same eigenvectors. If \( S_{l} = B\uplambda_{r} B^{T} \), then

$$ S_{r} = B\uplambda_{r} B^{T} ,\,\uplambda_{l} +\uplambda_{r} = I $$
(5)

\( I \) is the identity matrix. The projection matrix is achieved by the following equation:

$$ W = (B^{T} P)^{T} $$
(6)

This is the expected spatial filter. Then the EEG signals can be projected on the first \( m \) and last \( m \) columns of \( B \). So a single trial EEG data can be transformed into:

$$ Z = WE $$
(7)

2.5 Classification Scheme

The support vector machine (SVM) is a machine learning method proposed in the 1990s [26]. It is mainly proposed for pattern recognition problems. Suppose the EEG data set \( A \in R^{d} \) is from two classes, and could be divided by a hyperplane linearly. The resulting hyperplane could be expressed as: \( WA + b = 0 \). \( W \in R^{d} \) is weight vector and \( b \) is the intercept (scalar). The problem is transformed into finding the optimal hyperplane as follows:

$$ \begin{aligned} & \hbox{min} \,J\left( {W,\upvarepsilon} \right) = \frac{1}{2}||W||^{2} + c\mathop \sum \limits_{i = 1}^{n}\upvarepsilon_{i} ,c \ge 0 \\ & s.t.\text{ }y_{i} (W^{T} A^{(i)} + b) \ge 1 -\upvarepsilon_{i} ,\upvarepsilon_{i} \ge 0,i = (1, \cdots n) \\ \end{aligned} $$
(8)

Where \( A^{(i)} \) is a feature vector of a training sample, and \( y_{i} \) is the category with labels, {−1, 1}, to which \( A^{(i)} \) belongs. \( W \) is the hyperplane coefficients vector. c is used to control the trade-off between the model complexity and empirical risk [27]. In this paper, the 10*10-fold cross validation accuracy approach were used to evaluate the performance of each patient.

3 Results

3.1 Feature Extraction

Figure 3 shows the ERD maps from patients 4 and 9 over channels C3 and C4. When patient 4 imagined right hand movement, ERD phenomena at channel C3 were stronger than at C4. When patient 9 performed the left-hand MI, ERD phenomenon at channel C4 were stronger than at C3.

Fig. 3.
figure 3

The figure described two patients’ ERD map during the right hand and left hand MI. “P” represents “patient”.

Full size image

In the CSP algorithm, W is the projection matrix, and W−1 is the inverse matrix of W. The columns of W−1 are the time invariant vectors of EEG source distribution vectors called common spatial patterns [28]. Figure 4 shows the ten patients’ first and last common spatial patterns. The first pattern was related to the ERD phenomena over left motor area of the cortex (obtained by maximizing the variance of the right hand MI EEG data). Accordingly, the ERD phenomenon over the right sensorimotor area of the cortex was related to the last pattern, corresponding to the left hand MI.

Fig. 4.
figure 4

The topographic maps of the first and last spatial patterns of ten patients. The patterns were extracted from averages of 60 trials. “P” represents “patient”.

Full size image

The topographic maps in Fig. 4 show that left vs. right left hand MI produced clearly different spatial patterns for patients 1, 6, 7 and 10. The spatial patterns from patients 2, 4 and 8 did not show clear differences between left and right hand MI.

3.2 Classification Results

Calculation of the accuracy was done via 10-fold cross validation. The accuracies in Fig. 5 were calculated for all trials within a time window of 1.5 s after the attention beep until the end of the trial, in steps of half a second. Figure 5 shows the right hand MI, left hand MI, and overall classification results. Patients 1 to 5 have left hemiplegia, and patients 6 to 10 have right hemiplegia.

Fig. 5.
figure 5

The accuracy plots across all 60 trials for ten patients. The green, red and blue lines represent right hand motor imagery, left hand motor imagery and overall classification accuracies respectively. “P” represents “patient”. (Color figure online)

Full size image

Results showed that patients 2, 4, 6 and 10 exhibited higher accuracies with the right hand MI than the left hand MI during most of the movement imagery period. For patients 3 and 9, the accuracies of left hand MI are higher than right hand MI. For the remaining patients, differences between the left and right hand MI are not obvious. Overall, these results do not support a relationship between which hand is paretic and classification of left vs. right MI. The higher accuracy during right hand MI (compared to left hand MI) may result from handedness or greater use of the right hand, but this suggestion requires further testing.

Figure 5 also shows that the accuracies varied across patients. The accuracies from patients 5 and 7 were above 80% during most of the MI period, while accuracies of patients 1, 8, and 9 were close to the chance level 50%. These performance results are generally consistent with work with healthy patients who have not trained to use a motor imagery BCI. There was no clear effect of paretic hand on accuracy.

4 Discussion

The primary goal of this study was to investigate whether there are differences between the motor imagery performance of the paretic and non-paretic hands for stroke patients. Ten stroke patients participated in our experiment. The results did not reveal a relationship between which hand was paretic (left vs. right) and either BCI performance or the separability of left vs. right hand MI.

Figure 3 displays the ERD maps from two patients. For patient 4, ERD phenomena at channel C3 were stronger than at C4 during right hand motor imagery. However, ERD phenomena at channel C4 were not stronger than at C3 during left hand motor imagery, but lower. This result is consistent with the classification accuracies in Fig. 5. Results in Fig. 5 showed that right hand MI yielded better performance than left hand MI for patient 4. For patient 9, ERD phenomena at channel C4 were significantly stronger than at C3 during left hand MI, while there were not obvious ERD changes over either channel during right hand MI. Again, the results in Fig. 5 are consistent, as patient p performed better with left hand MI than right hand MI.

Figure 5 shows that the MI accuracies do not show a strong relationship between the paretic vs. non-paretic hand. However, for many patients, regardless of paretic side, right hand MI yielded better classification accuracy than left hand MI. [29] reported that motor evoked potential amplitudes induced by MI of right (dominant hand) finger movement were significantly larger than those induced over the left (non-dominant hand). According to some relevant research, patients would perform better when they are familiar with (or skilled in) the action of complex MI. [30] reported that sports experts showed significant activation in the parahippocampus during imagery of professional skills relative to the novices. [31] found that the mean execution rate was significantly faster in a skilled MI condition relative to a novel condition. Hence, patients would perform better when they were familiar with the action of motor imagery. Since all patients were right handed, their right hands would be more skilled in some actions. This might lead to the better performance during right hand MI.

The overall accuracies from most patients were not high enough for reliable communication. In a motor imagery based BCI, users should learn how to modulate their brain activity proficiently. Future research could explore new ways to train stroke patients based partly on ongoing measures of EEG activity [1,2,3, 18], including paradigms in which users can choose their own mental activities relating to hand movement. Instead of learning to perform specific hand MI tasks such as imagining grasping or lifting, users could choose tasks such as writing, typing, or painting.

5 Conclusion

This study explored the differences between the MI of the paretic versus non-paretic hands for stroke patients. Results from ten patients did not show a strong relationship between BCI classification accuracy and which hand was paretic. The MI accuracies of the paretic hand might not be lower than the non-paretic hand. On the other hand, MI was more accurate for right than left hand MI, which may result from handedness and the subjects’ freedom to choose their own MI task. Future work could explore training with self-selected imagery to improve performance, and explore the impact on rehabilitation of motor impairment resulting from stroke.