Does feedback based on FES-evoked nociceptive withdrawal reflex condition event-related desynchronization? An exploratory study with brain-computer interfaces

Twelve healthy naïve volunteers without neurological or cognitive conditions (mean age 32 ± 9 years, 6 female) were invited to participate, of whom 8 were enrolled in the study after assessing the following inclusion criteria: ability to modulate their sensorimotor rhythms achieving ERD and visible foot dorsiflexion following FES-evoked NWR. The study was conducted following the Declaration of Helsinki. The protocol was approved by the Central Ethics Committee for Research and Medical Practice of the Entre Ríos, Argentina, and registered in the National Registry of Health Research (RENIS IS001667). All subjects signed a written informed consent prior to participation in the experiment.

The purpose of this stage was twofold: (1) to determine the desynchronization frequency (${f}_{ERD}$) to configure the BCI and (2) to identify the NWR stimulation intensity to configure the electrical stimulator. The ${f}_{ERD}$ was defined as the EEG frequency that showed the largest desynchronization, assessed with the coefficient of determination (r²) at Cz [27]. NWR stimulation intensity was defined as the minimum comfortable electrical stimulation intensity that achieved visible foot dorsiflexion. The limb with the highest r² in Cz for ${f}_{ERD}$ was selected for further use in the Feedback stage.

2.2.1. Materials

2.2.2. EEG signal processing

2.2.3. Protocols

This stage consisted of 2 sessions, in which the two ERD-based BCIs were tested: BCI_NWR and BCI_VR. During these sessions, volunteers performed MI tasks, which consisted of ankle dorsiflexion intention without muscle contraction, to imitate the paresis of the stroke users.

2.3.1. Materials

Figure 1 shows the block diagram of the ERD-based BCI used in this study. It consisted of 3 modules: Signal Sensor and Conditioner, Control Signal Generator and the Actuator Device. The first two modules comprised the following blocks: Acquisition, Filtering, Segmentation, Spatial Filtering, Feature Extraction, Feature Selection, and Classification.

Zoom In Zoom Out Reset image size

The electrodes positions placed in the Parameters Selection stage were preserved in this stage. Acquisition and Filtering were the same blocks that were used in the previous stage. Then, data from each channel were segmented and sent to the next module in sets of 8 samples (total length: 31.25 ms) [27]. In the Control Signal Generator module, a Laplacian spatial filter was implemented to improve the signal to noise ratio and to enhance the signal from Cz (Cz' signal) [31]. Then, in the Feature Extraction block, the power spectral density from Cz' was computed in 500-ms epochs via a 16th-order autoregressive model estimated by the Burg Method [32], in the range of 8–30 Hz and using 3-Hz bins. In the Feature Selection block, the spectral amplitude from Cz' corresponding to the ${f}_{ERD}$ (${A}_{{f}_{ERD}}$) was selected, resulting in 64 ${A}_{{f}_{ERD}}$ per task [16]. The first 500 ms after the cue were not used because they can present cue-evoked potentials [33]. The Classification block identified the ERD when the mean ${A}_{{f}_{ERD}}$ (${\bar{A}}_{{f}_{ERD}}$) between 0.5 and 2 s after the cue falls below an adaptive threshold. This threshold was set as the lowest of the last two ${\bar{A}}_{{f}_{ERD}}.$ The initial classification threshold was established as the first ${\bar{A}}_{{f}_{ERD}}.$ In this way, the strategy to update the detection threshold compelled the volunteers to try decreasing the amplitude of their sensorimotor rhythms to achieve the activation of the BCI, in a trial-by-trial basis. The output of the classification block provided a single activation command for each detected MI task, to trigger the stimulation pulse train [30]. Figure 2 shows the BCI operation protocol.

Zoom In Zoom Out Reset image size

The control signal was sent through virtual ports to the two actuators, so the following two BCIs were available:

• BCI_VR: the actuator was a VR software developed on the V-REP PRO EDU 3.0 platform (Coppelia Robotics GmbH, Zürich, Switzerland) that imitates the foot dorsiflexion in a mirror-like fashion, giving extrinsic feedback [6].
• BCI_NWR: the actuator consisted of the electric stimulation system used in the Parameters Selection stage that generates the electrical stimuli to evoke the NWR, producing reflex responses resulting in foot dorsiflexion and therefore, intrinsic feedback [30].

The Filtering, Segmentation, Spatial Filtering, Feature Extraction, and Feature Selection blocks were implemented using the BCI2000 platform [27]. The Classification block and the communication with the actuator devices were implemented in Matlab 9.01, R2016a (The Mathworks, Natick, MA, USA).

2.3.2. Recording protocol

2.3.3. ERD quantification

Data from 528 trials of EEG signals were processed. Following the approach put forward by [32, 34], the percentage of ERD at ${f}_{ERD}$ was estimated as:

$$\begin{eqnarray*}&&ERD \% =\left(\left({\bar{A}}_{ITI}-{\bar{A}}_{MI}\right)/{\bar{A}}_{ITI}\right)\,* \,100\end{eqnarray*}$$

where ${\bar{A}}_{ITI}\,\,$is the ${\bar{A}}_{{f}_{ERD}}\,$of the ITI and ${\bar{A}}_{MI}$ is the ${\bar{A}}_{{f}_{ERD}}\,$during the MI task.

2.3.4. NWR quantification

The NWR response was assessed by surface electromyography (EMG) of the tibialis anterior muscle (TA). Two Ag/ClAg electrodes of 1 cm in diameter, were placed 2 cm apart on the belly of the TA and the ground electrode was placed on the malleolus [35]. A BioAmp amplifier [36] and the BrainBay software platform [37] were used to record EMG. The EMG signals were filtered between 0.5 and 500 Hz and sampled at 2 kHz.

The EMG signal was analysed using Matlab 9.01, R2016a (The Mathworks, Natick, MA, USA). The analysis window, defined as the period of time in which a NWR is expected, was established between 60 and 180 ms after the start of the electric stimulation train [19]. The mean amplitude of the rectified TA EMG signal was calculated by means of the Root Mean Square value ($RM{S}_{NWR}$). The baseline $RMS$ ($RM{S}_{baseline\,}$) was defined as the $RMS$ amplitude of the TA EMG starting 120 ms before the stimulation onset and ending at the stimulation onset. The $RM{S}_{baseline\,}$was subtracted from the $RM{S}_{NWR\,}$ in order to quantify the reflex response and, finally, the difference was normalized with respect to the baseline ($RM{S}_{norm\,}$):

$$\begin{eqnarray*}&&RM{S}_{norm\,}=\left(\,RM{S}_{NWR}-\,RM{S}_{baseline\,}\right)/RM{S}_{baseline\,}\end{eqnarray*}$$

To detect the presence of the NWR, a z score was calculated as:

$$\begin{eqnarray*}&&z=\left({A}_{{\rm{\max }}}-{\mu }_{baseline}\right)/{\sigma }_{baseline}\end{eqnarray*}$$

where ${A}_{\max \,}$ is the maximum amplitude of the EMG in the analysis window, and ${\mu }_{baseline}$ and ${\sigma }_{baseline}$ are the means and standard deviations of the baseline, respectively. A z score greater than 12 indicates the presence of a NWR [38].

2.3.5. BCI performance

To estimate the BCI performance, accuracy was defined as the number of correct classifications of user´s states divided by the total amount of trials (in percentage):

$$\begin{eqnarray*}&&Accuaracy\, \% =\left(\left(TP+TN\right)/N\right)\,* \,100\end{eqnarray*}$$

where $TP$ is the number of True Positive; $TN$ is the number of True Negative trials and $N$ is the number of the total trials. Then, because the BCI was disabled during the resting trials (ITI), $TN$ was equal to the number of resting trials.

2.3.6. Pain assessment

Statistics were performed using SPSS^® v.23. Summary measures for the data are presented as mean ± standard deviation. Linear mixed models were used to analyze the relationship between the reflex response (i.e. the NWR size estimated by the RMS_norm ) and both the ${\bar{A}}_{MI}$ and the $ERD \% \,\,$subsequent the FES-evoked NWR triggered by a MI; and to determine if the type of feedback (extrinsic feedback by RV or intrinsic feedback by NWR) affects the ${\bar{A}}_{MI}$ or the $ERD \% \,\,$subsequent to the activation of the ERD-based BCI. A compound symmetry structure was used to model the covariance matrix and p values smaller than 0.05 were considered statistically significant.

The stimulation intensity required to evoke a withdrawal reflex is often sufficient to activate the nociceptive system [19]. Regarding the use of a nociceptive stimuli that may evoke painful sensations, the volunteers in this study were aware that there was a chance that they would receive a nociceptive stimulus to evoke the NWR after a MI. The stimulation intensity selected in our study was high enough to achieve visible ankle dorsiflexion but was not scored as painful by the volunteers, i.e. none of them rated above 5, which was the pain threshold. Also, the mean stimulation intensity was similar to the one used in NWR experiments with healthy subjects (14.6 ± 5.8 mA) [19]. Stroke patients usually present sensory problems with a concurrent decrease in sensitivity [40], so our intensity levels would not be a problem for future application of the method in these patients. In this sense, studies involving stroke patients reported stimulation intensities of 31.7 ± 2.9 mA to evoke NWR and 22.1 ±2.2 mA for gait re-education based on NWR [20].

In this study, it was observed that volunteers achieved ERD even after receiving an unpleasant stimulus 15 s earlier. In other words, they did not suppress the oscillations of their sensorimotor rhythms (at alpha and beta frequencies), perhaps because the electrical stimuli were annoying but not painful. Phasic pain stimuli transiently suppress oscillations at alpha and beta frequencies [24]. Nociceptive stimulation produces transient and strong alpha-ERD at latencies between 150 and 1000 ms after stimulation with ITI ranging from 8 to 12 s [23]. It was also reported that this short latency nociceptive-induced alpha-ERD reflects the summation of sensory-related and task-related cortical processes [21]; but the effects on ERD induced by unpleasant stimuli have not yet been reported.

In the BCI context, feedback can be given intrinsically or extrinsically, as a response to the MI. The feedback informs the person about their performance while the motor task was being performed. Extrinsic feedback is defined as the information provided by an external source, for example a therapist or a display [41–43]. Intrinsic feedback is related to sensory information from changes in the body, given for example by orthotic or FES [8, 10, 11]. Intrinsic feedback activates visual, auditory, proprioceptive and tactile pathways and enables an appropriate sensory context to facilitate cortical reorganization. For BCIs in rehabilitation, this feedback is generally richer than extrinsic feedback [44, 45], particularly with FES-evoked NWR since afferent sensory signals, descending modulating signals and motor commands are integrated by the central nervous system into a motor output. Moreover, the electrical stimulation needs to be combined with simultaneous voluntary effort to potentially provide an artificial way that ensures synchronized activation of the residual corticospinal tract [18]. This voluntary effort involves the sensorimotor rhythms modulation as a response to the motor task, i.e. ERD which was detected by the BCI. No reports about ERD-BCI with FES-evoked NWR have been found.

In this study a main effect of the feedback type was found in the subsequent ${\bar{A}}_{MI},$ with intrinsic feedback resulting in a larger ${\bar{A}}_{MI}$ (figure 6(a)), but it was not found for the subsequent $ERD \%$ (figure 6(b)). Attention to unpleasant stimuli could have influenced the ${\bar{A}}_{MI}$ [46], but it is not clear. On the other hand, $ERD \%$ takes into account the resting condition (${\bar{A}}_{ITI}$). The 15 s duration ITI avoided the habituation of NWR and allowed the volunteers to prepare for a new ERD but, perhaps, it was not long enough to reach the resting state during the use of BCI_NWR. The role of mental states in determining individual post-stimulus brain activity, perception and behaviour is a key in this type of studies [24]. Future studies are necessary to explore the influence of the attention to unpleasant stimuli and to determine whether the volunteers went back easier into rest with RV feedback than with NWR feedback.

The mean BCI accuracies (74.4% for the BCI_VR and 75.0% for BCI_NWR) showed that both of the BCI were higher than the chance level of 63.3% [47] and they were in the range reported by other authors for lower limb ERD_based BCI [6].

Although the development of the BCI_NWR was not the aim of this study, the concept of eliciting the NWR by a BCI was also explored using an ERD-based BCI with FES-evoked NWR in the Feedback stage.

Regarding the applied ERD-based BCI, the Acquisition, Filtering, Segmentation, Spatial Filtering, Feature Extraction and Feature Selection blocks were based on those proposed by Schalk and Mellinger [27] and used by several authors [48–50]. A 16th-order autoregressive model was used to estimate power spectral density because it performs better than Fast Fourier Transform in short data segments [51].The order of the model was selected as it is suggested in the literature for scalp EEG sampled at 256 Hz [27, 52]. In future studies, it could be interesting to compare the advantages and disadvantages of different feature extraction methods.

The BCI recorded the EEG in monopolar configuration, filtered it, and then used a Laplacian spatial filter to generate a single signal that highlights the signal coming from the electrode Cz (Cz' signal). Other authors recorded the EEG in bipolar configuration and chose the most convenient pair of electrodes for the best BCI control, placed in different areas of the cerebral cortex [15]. However, in the context of neuro-rehabilitation, the aim is to achieve motor recovery. Then, it is essential to record the EEG signal originated in the injured motor cortical area which is expected to recover through training (in our study, the lower limb and therefore Cz). Although in stroke people the EEG signal to noise ratio is likely reduced, there are studies that report that it is possible to identify ERD and control a BCI with signals from the affected hemisphere [6, 11, 53, 54].

Regarding spatial filters, in a review reported by Bashashati et al [55], the authors reported that 32% of the studies used Laplacian filters, 22% used principal component analysis (PCA) or independent component analysis (ICA), 14% common spatial patterns (CSP) and 11% used common average reference (CAR) filters. In filters based on PCA, ICA or CPS the weights are adapted from data of all the EEG channels [56], which may not be those corresponding to the cerebral cortex to recover. On the other hand, the use of CAR filters tends to reduce the impact of artefacts common to all channels, but they do not fulfil the objective of enhancing the signal coming from a specific EEG channel. For these reasons, and given that the filter seeks to highlight spatially localized brain activity [31], it is considered that the choice of the Laplacian filter is appropriate for the application in neuro-rehabilitation.

The used BCI detected the MI estimating the average energy of the EEG signal at ${f}_{ERD}$ in Cz' in each MI and then compared it with an adaptive threshold established as the lowest of the last two ${\overline{A}}_{MI}.$ This strategy was similar to those used by other authors who established an adaptive threshold as the mean of the means of the populations of the ${\bar{A}}_{{f}_{ERD}}$ during the MI and during rest [57]. The used BCI generated a binary output to command the actuator, for each detected ERD during a MI. Similarly, others studies proposed a BCI that provided a binary command signal for the control of a FES device. These BCIs were based on ERD with a double threshold configured for each subject with an empirical procedure [15].

All of the modules of the BCI used in this study could be modified or optimized by using other machine learning strategies [58] but always taking into account the feasibility of use in a clinical rehabilitation setting, that is, being of rapid configuration and calibration and easy-to-use by the therapist/operator.

Seven of the eight healthy naïve volunteers showed ${f}_{ERD}$ (table 1) in beta rhythm, while only one participant evidenced ERD in mu rhythm. These results coincide with the bibliography where it was reported that there is a desynchronization of the both sensorimotor rhythms [2, 4, 5, 32, 33].

After each activation of the electric stimulator by the BCI, somatosensory evoked potentials were registered, as shown in figure 3. When using electrical stimulation at high-intensity levels just like in this study (13.6 ± 6.4 mA; needed to overcome the threshold for NWR evocation and to perform joint movement), somatosensory evoked potentials reflect the simultaneous activations of large-diameter, lower threshold non-nociceptive fibers (type Aβ) and small-diameter, higher threshold nociceptive fibers (Aδ) [59].

In the example of the evolution of the RMS_norm during 3 series of a session shown in figure 5, the intra-subject variability is observed during a session and at a constant stimulation intensity. Likewise, there is a tendency to decrease the RMS_norm (relative to the maximum of the series) during the third series. This could be due to the habituation of the NWR that, although the ITI of 15 seconds was established to avoid it, might not be enough for this subject. These considerations should be taken into account at the time of the implementation of a rehabilitation therapy. Also it would be interesting to estimate the ERD latency both as a biomarker and as a detector of the onset of ERD in order to give feedback to the user at the earliest possible time [60].