Effect of head-mounted virtual reality and vibrotactile feedback in ERD during motor imagery Brain–computer interface training

ABSTRACT

Brain–computer interfaces (BCIs) can provide a non-muscular channel of control to stroke patients for motor rehabilitation. This can be achieved through the use of motor imagery (MI) training, involving the modulation of sensorimotor rhythms. The practice of MI has been shown to be able to strengthen key motor pathways when reinforced with rewarding feedback. Recently, there has been a growing evidence of the positive impact of embodied virtual reality (VR) and vibrotactile feedback in MI training. Nonetheless, it is not yet clear what the optimal MI-BCI setup is for evoking stronger sensorimotor rhythms in VR. In this study, we investigate the impact of head-mounted VR, and vibrotactile feedback during MI-BCI training in the induced sensorimotor rhythms. To achieve this, 19 healthy subjects performed MI training with embodied VR between four conditions: head-mounted vs. screen VR, with and without vibrotactile feedback; and two control conditions: abstract MI without embodied feedback, and motor execution. The event-related desynchronization (ERD) and the lateralization indices (LI) of the Alpha and Beta EEG rhythms were analyzed in a within-subject design. Results show that the combination of vibrotactile feedback and embodied VR can induce stronger and more lateralized Alpha ERD; nonetheless, LI was not significantly different across conditions.

KEYWORDS:

• Brain–computer interfaces
• event-related desynchronization
• motor imagery
• virtual reality
• vibrotactile feedback

View correction statement:

Correction

1. Introduction

The implementation of electroencephalography (EEG)-based brain–computer interfaces (BCIs) represents a promising and innovative approach for post-stroke rehabilitation, particularly for patients with a poor prognosis [1]. By decoding motor-related brain signals and utilizing them as control inputs, BCIs offer an alternative communication and control channel for individuals who lack volitional movement. These signals are then reinforced through closed-loop feedback mechanisms to strengthen the critical motor pathways that support post-stroke motor recovery [2,3]. Previous research has demonstrated modest success in utilizing this approach for severe stroke patients [4]. Specifically, motor-imagery-based BCIs (MI-BCIs) have been found to be effective in facilitating recovery from brain lesions, especially in individuals who have suffered from stroke [5,6]. This technique involves the generation of mental images of limb movement, which are then reinforced in a closed-loop into physical movements of exoskeletons [7], virtual avatars [3,6], or activation of functional electrical stimulation (FES) [2]. The importance of the real-time feedback provided in the closed-loop operation of BCI has been shown to be a significant characteristic contributing to motor rehabilitation. Specifically, the decoding of voluntarily modulated EEG sensorimotor rhythms, reinforced by activation of body natural efferent and afferent pathways through closed-loop BCI, may promote purposeful cortical reorganization in relevant cortical areas [2].

Concretely, the use of a MI-BCI can influence EEG rhythms, which are linked to both motor planning and execution. When imagining a motor task, the Alpha (8–12 Hz) and Beta (13–30 Hz) sensorimotor rhythms desynchronize, a phenomenon known as event-related desynchronization (ERD) [8]. ERD, which involves a reduction in EEG power, is attributed to either sensory processing or motor behavior [9].

Presently, there is increasing evidence that utilizing sensorimotor rhythms has the potential to enhance motor function in chronic stroke patients [10]. Specifically, the connection between MI training and changes in ERD during brain–computer interaction is an active field of research [11].

Recent advancements in enhancing MI-BCI control performance have incorporated several approaches such as visual and tactile feedback [12], task gamification [13], visual feedback modality (e.g. screen vs. head-mounted display), and electrode montage optimization [14]. In addition, the utilization of motor-priming preceding the MI-BCI training has been shown to help increase MI-BCI performance [15,16]. Furthermore, accumulated evidence supports the benefits of embodied feedback during MI-BCI training, including the augmentation of MI skills using human-like hands [17,18], and the enhancement of the sense of embodiment in virtual reality (VR) through multisensory feedback related to body movements [19].

In addition, head-mounted virtual reality (HMD-VR) technology can help to deliver realistic first-person perspective embodied feedback, helping to enhance bodily awareness, which has been demonstrated to improve motor learning [20,21].

Moreover, embodied feedback in VR can lead to the activation of the mirror neuron system (MNS) [22,23] by means of motor observation (MO) [24]. Mirror neurons are active not only when performing an action but also when observing another subject performing the same action [23]. Specifically, MO is often used in stroke VR-based rehabilitation [25] since it could provide an additional or alternative source of motor training [26]. Moreover, VR can provide a secure and simulated real-world environment, under controlled conditions, that could potentially help patients transfer the learned skills to real life [27].

Furthermore, although scarce, the use of haptics and vibrotactile stimulation have increasingly shown their potential in BCIs and neurofeedback applications [28]. Combined with embodied visual feedback, haptics are capable of providing additional sensory information and proprioceptive feedback (feeling and seeing hand movements) [29]. Several works demonstrated that the use of vibrotactile feedback enhanced the usability of the BCI systems [30,31], BCI classification performance [12,32–34], but also functional connectivity (FC) of sensorimotor networks [35].

Nonetheless, it is not yet clear what the optimal MI-BCI setup is for inducing stronger sensorimotor rhythms in VR nor what is the contribution of haptic feedback in ERD power within a VR-BCI training paradigm. Given the current limitations, the objective of this paper is to explore which VR – BCI conditions lead to the strongest and most lateralized brain activity modulation.

2. Methodology

In this study, a cohort of healthy subjects had been recruited to perform MI training in a VR-BCI setup. The experimental setup involved six conditions for comparing the effect of head-mounted VR and vibrotactile feedback in ERD power. Finally, this study resulted in the production of more than 100 labeled MI EEG datasets, publicly available [36].

2.1. Participant demographics

A total of 21 healthy volunteers were recruited; however, the data of two subjects were corrupted and had been excluded from the analysis, resulting in 19 subjects in total. The subjects had a mean age of 24.79 years (SD = 3.54 years), with the youngest being 21 years old and the oldest 36. The cohort consisted of 68% male and 32% female participants. In terms of education, 16% only had a high-school diploma, while 32% had a bachelor’s degree, 42% a master’s degree, and 11% a doctorate. All participants signed an informed consent before participating in the study in accordance with the 1964 Declaration of Helsinki, and the protocol was approved by the Ethics Committee of CHULN and CAML (Faculty of Medicine, University of Lisbon) with reference number 245/19.

2.2. Experimental procedure

2.2.1. Conditions

The experiment consisted of having the subjects perform MI training under four experimental conditions and two control conditions. Specifically, the experimental conditions used NeuRow [6] – a VR-BCI training paradigm that renders a set of virtual arms from an avatar in a first-person perspective – while the control conditions used non-VR and non-embodied abstract feedback, based on the BCI-Graz paradigm [9], and involved either MI or motor execution. Further, two of the NeuRow conditions involved vibrotactile stimulation, activated on each of the virtual hand that was involved in MI after the cue and for the duration of the trial (Figure 1).

Figure 1. a. Experimental setup. (i) 32 active electrodes EEG system; (ii) HMD VR; (iii) controllers for vibro-tactile feedback; b. MI trial of the training protocol. 5-seconds of baseline, followed by a cue (directional arrow) for left or right-hand MI and 5 seconds of motor imagery with (i) black screen for MI and ME conditions, and (ii) the movement of the virtual arm for NeuRow conditions: MIMO, MIMOHP, MIMOVR, and MIMOVRHP.

All six conditions and their acronyms are described below, with conditions 1 and 6 being the control:

1. Motor Imagery (MI): The standard MI training, screen based, with a fixation cross and directional arrows on a black background.

2. Motor Imagery/Motor Observation (MIMO): A MI training paradigm using NeuRow, with a fixation cross and directional arrows overlaid on a screen-based VR environment.

3. Motor Imagery/Motor Observation with Haptics (MIMOHP): A MI training paradigm using NeuRow, with a fixation cross and directional arrows overlaid on the VR environment, which was displayed through a computer monitor. Hand controllers also provided haptic feedback through vibrotactile stimulation during the MI trial on each of the virtual hand that was involved in MI.

4. Motor Imagery/Motor Observation with VR HMD (MIMOVR): A MI training paradigm using NeuRow, with a fixation cross and directional arrows overlaid on the VR environment, which was rendered through a VR HMD.

5. Motor Imagery/Motor Observation with VR HMD and Haptics (MIMOVRHP): A MI training paradigm using NeuRow, with a fixation cross and directional arrows overlaid on the VR environment, which was rendered through a VR HMD. Hand controllers also provided haptic feedback through vibrotactile stimulation for each hand during the MI trial, similarly to MIMOHP condition.

6. Motor Execution (ME): A fixation cross and directional arrows were displayed on a black background through a computer monitor (similarly as in MI), and guided the subjects through the experiment by having them tap their fingers accordingly. Data from this condition were available only after subject 7, so only 10 subjects have performed ME.

2.2.2. Protocol

All subjects participated in all conditions and in a randomized order, except the ME condition, which was always performed last. All conditions took place on the same day, in a within-subject design. Every condition involved one session with 42 trials, 21 for left- and 21 for right-hand imagery, except for the ME condition, which involved 18 trials. Each trial had an overall duration of 10 s, with 5 s of baseline and 5 s in the MI task, including an average inter-trial gap of 1.25 s (Figure 1b). This resulted in 8 min and 19 s per condition, and 3 min, and 49 s for ME. In between conditions, there was a gap of approximately 5 min, where the configuration of the setup was adapted for the next condition (e.g. replacing VR-HMD with screen, or adding haptics), and the signal quality was inspected.

2.2.3. Visual feedback

The visual instructions remained consistent across all conditions, with a fixation cross appearing on the screen, followed by a directional arrow indicating the hand side. Next, and during the MI trial, in all conditions – except ME – subjects were instructed to perform motor imagery only of the rotation of each corresponding hand after the cue from a first-person perspective, and across the whole duration. A black screen was displaying a fixation cross during the trial of MI and ME conditions (Figure 1b(i)), and the movement of the virtual arm for NeuRow Conditions: MIMO, MIMOHP, MIMOVR, MIMOVRHP (Figure 1b(ii)). An example of the NeuRow visual feedback in a closed loop can be found online.¹

2.3. Experimental setup

2.3.1. Hardware

To record the EEG signals, a wireless EEG amplifier was used (LiveAmp; Brain Products GmbH, Gilching, Germany), with 32 active electrodes and 3 accelerometer channels at a sampling rate of 500 Hz. The spatial distribution of the electrodes followed the 10–20 systems (Figure 1a(i)). The visual feedback was provided through a $24{ }^{\mathrm{\prime }\mathrm{\prime }}$ LCD PC monitor in all conditions except in MIMOVR and MIMOVRHP, in which an Oculus Rift CV1 headset (Facebook Reality Labs, formerly Oculus VR, CA, USA) was used instead (Figure 1a(ii)). Lastly, the vibrotactile feedback was delivered through the Oculus Rift hand controllers (Figure 1a(iii)).

2.3.2. Software

The experimental pipeline and stimulation generation was implemented through NeuXus,² an open-source Python toolbox for real-time biosignal processing [37]. During the experiment, NeuXus was sending the stimulus markers (e.g. start of trial, left, right, end of session, etc.) both to VR via the Lab Streaming Layer (LSL)³ protocol, and via a serial-port to the TriggerBox (Brain Products GmbH, Gilching, Germany) for a reliable synchronization between signals and markers (or triggers). Next, the raw data were visualized and logged in the BrainVision Recorder (v1.22; Brain Products GmbH, Gilching, Germany).

2.4. Data analysis

2.4.1. Data pre-processing

The data were analyzed in MATLAB (R2021b and R2022a; The MathWorks, Inc., Natick, MA, USA) with the EEGLAB toolbox (v2022.0; Swartz Center for Computational Neuroscience, San Diego, CA, USA).

Initially, the EEG was re-referenced to common average (CAR) and downsampled to 125 Hz. Next, a band-pass filter between 1 and 40 Hz was applied, and the trials were epoched between −5 and 5 seconds. For removing artifacts from the signals, Independent Component Analysis (ICA) was performed. In addition, we utilized the ICLabel tool (Swartz Center for Computational Neuroscience, San Diego, CA, USA), which helps distinguish independent components (ICs) as brain or non-brain sources. Finally, bad trials were inspected and removed manually from the data.

2.4.2. ERD measures

ERD of Alpha (8–12 Hz) and Beta (13–30 Hz) was computed across all conditions over the epoched data as a percentage change of the power spectral density (PSD) of the signal during MI from baseline (eq.1), according to [38], and the median ERD (mERD) was extracted from each electrode during the MI trial between 0 and 5 s.

(1) $ERD(\mathrm{\%})=\left(\frac{\mathrm{B}\mathrm{a}\mathrm{s}\mathrm{e}\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{w}\mathrm{e}\mathrm{r}-\mathrm{M}\mathrm{I}\mathrm{P}\mathrm{o}\mathrm{w}\mathrm{e}\mathrm{r}}{\mathrm{B}\mathrm{a}\mathrm{s}\mathrm{e}\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{P}\mathrm{o}\mathrm{w}\mathrm{e}\mathrm{r}}\right)\times 100$(1)

Furthermore, in order to quantify the lateralization of Alpha and Beta ERD between hemispheres, we computed the lateralization index (LI). Lateralization between hemispheres is commonly used to quantify the asymmetry of neural modulation intensity in brain imaging studies [39]. In this study, LI was computed on mERD over the C3 and C4 electrodes (eq.2), since they are considered the target electrodes for capturing the maximal ERD close to the sensorimotor area [38].

(2) $LI=\frac{\left[\begin{array}{c}\left(mER{D}_{\mathrm{l}\mathrm{e}\mathrm{f}\mathrm{t},\mathrm{C}3}-mER{D}_{\mathrm{l}\mathrm{e}\mathrm{f}\mathrm{t},\mathrm{C}4}\right)\\ +\left(mER{D}_{\mathrm{r}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t},\mathrm{C}4}-mER{D}_{\mathrm{r}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t},\mathrm{C}3}\right)\end{array}\right]}{2}$(2)

which is positive if the ERD power is mostly contralateral to the side of the arm movement during MI, or negative if it is ipsilateral.

2.4.3. Statistical tests

In order to determine statistically significant differences between the conditions in the mERD and LI, the Kruskal – Wallis test was used. Further, the LI was also compared to a null lateralization index (LI = 0) to determine if the brain modulation was significantly lateralized in any conditions. Due to the relatively small sample size (N = 19), a non-parametric test was utilized with a significance level of 0.05 (p < 0.05). Finally, whenever the null hypothesis was rejected, a post-hoc test was performed. The analysis consisted of pairwise comparisons using Dunn’s test, as it typically follows the Kruskal – Wallis test due to computing the same ranks.

3. Results

The analysis of the results aimed to determine which condition is capable of eliciting the most pronounced and predominantly lateralized Alpha and Beta ERD during training for MI.

3.1. ERD differences

From the ipsilateral Alpha mERD, ME shows the strongest ERD (Mdn = −37.45%), followed by MIMOHP (Mdn = −12.12%), MIMOVRHP (Mdn = −8.69%), MIMO (Mdn = −1.79%), MIMOVR (Mdn = −1.48%), and MI (Mdn = 3.57%). Between the contralateral medians, ME has similarly the highest median compared to the rest of the conditions (Mdn = −37.27%), followed by MIMOHP (Mdn = −33.30%), MIMOVRHP (Mdn = −22.86%), MIMOVR (Mdn = −20.16%), MIMO (Mdn = −12.92%), and MI (Mdn = −5.08%) (Figure 2(i)).

Figure 2. Contralateral and ipsilateral mERD (%) for all conditions and between (i) Alpha and (ii) Beta bands. * indicates statistically significant differences $(p<0.05)$.

From the ipsilateral Beta mERD, ME has the strongest ERD (Mdn = −17.30%), followed by MIMOHP (Mdn = −11.08%), MIMOVR (Mdn = −8.70%), MIMOVRHP (Mdn = −8.05%), MIMO (Mdn = −7.50%), and MI (Mdn = −5.47%). Between the contralateral mERD, ME has the strongest ERD (Mdn = −24.34%), followed by MIMOVR (Mdn = −17.07%), MIMO (Mdn = −15.39%), MIMOHP (Mdn = −14.29%), MIMOVRHP (Mdn = −13.76%), and MI (Mdn = −8.69%) (Figure 2(ii)).

According to the Kruskal – Wallis test, the Alpha mERD was significantly different across all groups (ipsilateral: ${\chi }^2$ = 33.92, p < 0.001; contralateral: ${\chi }^2$ = 27.54, p < 0.001), but there were no significant differences between the Beta mERD (ipsilateral: ${\chi }^2$ = 10.75, p = 0.057; contralateral: ${\chi }^2$ = 10.30, p = 0.067). Specifically, according to Dunn’s test for the post-hoc pairwise comparisons (p < 0.05), MI was significantly different from all the other conditions for contralateral electrodes (MIMO: p = 0.043; MIMOHP: p < 0.001; MIMOVR: p = 0.012; MIMOVRHP: p < 0.001; ME: p < 0.001), while ME was significantly different from MI (p < 0.001) and the NeuRow conditions without haptic feedback, MIMO (p = 0.009) and MIMOVR (p = 0.033). On the other hand, the NeuRow conditions with haptic feedback were not significantly different from ME for the contralateral electrodes (MIMOHP: p = 0.480; MIMOVRHP: p = 0.153). The condition ME also had significantly different ipsilateral mERD from all the other conditions (MIMOHP: p = 0.021; the others: p < 0.001).

Overall, all conditions were able to induce ERD, although MI exhibited the lowest power as shown by the median values in Figure 2 Nonetheless, by looking at the temporal evolution of the mERD%, Alpha rebounds earlier compared to the rest of the conditions, which might have resulted into decreased ERD values when averaged across the whole trial (Figure 3). Further, and as anticipated, the strongest ERD was observed during ME [40]. The conditions MIMOHP and MIMOVRHP showed similar ERDs to ME for Alpha band and surpassed it momentarily in the right-hand trials Figure 3(ii)

Figure 3. Temporal evolution of average Alpha and Beta ERD (%) between the contralateral electrodes (C3 or C4) for left (L) and right (R) trials.

3.2. Spatial distribution of ERD

In order to understand better the location and spread of ERD across the different conditions, the spatial distributions of the mERDs have been extracted and are shown in Figure 4 Specifically, the ERD peaks were located in the region posterior to the central locations (C3, C4) in all MI conditions, while ME’s peaks were slightly more anterior and closer to the sensorimotor cortex. The MI control condition had a weak ERD spread in the posterior region of the sensorimotor cortex, while the NeuRow conditions had clearly defined ERD clusters around the contralateral electrodes. Conditions MIMOHP and MIMOVRHP had the most prominent ERD peaks of all MI conditions, particularly for Alpha band. The conditions MIMO and MIMOVR showed similar Alpha ERDs, with intensities in-between those in MI and the NeuRow conditions with haptic feedback. Condition ME had the strongest, most evident ERD. The MI conditions showed more lateralized Alpha ERDs than ME, albeit not as prominent.

Figure 4. Topographic plots of the average (i) Alpha and (ii) Beta ERDs (%) for all subjects, across all conditions.

3.3. Lateralization indices

In terms of ERD lateralization, the median Alpha LI of MIMOVR shows the highest (Mdn = 15.2), followed by MIMOHP (Mdn = 14.1), MIMOVRHP (Mdn = 13.7), and MIMO (Mdn = 10.8). The conditions MI and ME had the lowest medians (Mdn = 5.3 and Mdn = 7.3, respectively) (Figure 5(i).

Figure 5. Lateralization indices for (i) Alpha and (ii) Beta ERD. *indicates statistically significant differences from null (LI = 0) $p<0.05$.

In terms of Beta LI, the condition ME had the highest median (Mdn = 8.6) but also the broadest distribution. The MI conditions had similar distributions between themselves, with the conditions MIMOHP and MIMOVRHP having the highest medians (Mdn = 5.1 and Mdn = 5.4, respectively). However, all conditions included negative indices in their sample groups, which are found up to the lower quartile of the distributions (Figure 5(ii)).

According to the Kruskal–Wallis test (p < 0.05), none of the sample groups were significantly different (Alpha: ${\chi }^2$ = 6.06, p = 0.300); Beta: ${\chi }^2$ = 3.77, p = 0.582). However, there was a significant difference between the sample groups and a null LI, LI = 0 (Alpha: ${\chi }^2$ = 28.17, p < 0.001; Beta: ${\chi }^2$ = 20.03, p = 0.003). Dunn’s test for pairwise comparisons found MI and the NeuRow conditions to have significantly different Alpha LI sample groups from the null LI (MI: p = 0.010; NeuRow conditions: p < 0.001; ME: p = 0.057), while all the Beta LI sample groups were significantly different (MI: p = 0.006; MIMO and MIMOHP: p = 0.005; MIMOVR: p = 0.003; MIMOVRHP and ME: p < 0.001).

4. Discussion

4.1. Strongest ERD

From all the MI conditions, the ones that use NeuRow (MIMO, MIMOHP, MIMOVR, and MIMOVRHP) led to significantly stronger contralateral Alpha ERD. However, the NeuRow conditions that use vibrotactile stimulation as haptic feedback (MIMOHP and MIMOVRHP) produced the strongest Alpha ERD, which are comparable with motor execution’s equivalent (ME). The abstract-feedback control condition (MI) performed significantly worse than the NeuRow conditions, thus suggesting that NeuRow and haptic feedback, together, lead to stronger Alpha ERD. Furthermore, there were no significant differences between MIMOHP and MIMOVRHP, suggesting that haptic feedback is possibly a more important factor than VR for evoking stronger ERD since it has been shown that it can increase functional connectivity (FC) of sensorimotor networks [35].

4.2. Temporal evolution of ERD

Interestingly, Alpha ERD rebounds earlier during MI, compared to MIMO conditions. Our hypothesis is that because all MIMO conditions were using NeuRow, showing a repetitive movement of the virtual hand rotating across all 5 s of the trial, it might have helped the participants to sustain a stable and vivid imagery across time. Although this has also been observed in a cohort of older-adult group from a previous study with NeuRow [18], older research has shown that prolonged movements do not necessarily produce prolonged ERD throughout the whole movement [41]. Nonetheless, this ERD behavior could be related more to the type of MI strategy used and not trial duration, since prior research has shown that target-directed motor acts, can produce stronger ERD with prolonged rebound [42]; therefore, more research is necessary before concluding on the temporal impact of MIMO in ERD.

In terms of sensorimotor rhythm modulation, the Alpha band was much more reactive to the different conditions than the Beta band, as expected, being in line with prior research [38]. Nevertheless, motor execution led to a slightly, though not significantly, stronger Beta ERD than MI. ERD in Beta band has been observed shortly after performing both MI or motor execution.

Finally, every MIMO condition, except the control MI, was able to sustain, on average, their ERD until the end of the trials. Not only is it important to induce a strong ERD but to also be able to sustain it, as the brain modulation is consequently longer and the cortical reorganization more thorough. ERD in MIMOHP and MIMOVRHP could surpass ME’s ERD toward the end of the trials, on average, which suggests that these conditions are closer to ME at inducing a strong brain modulation.

4.3. Most lateralized ERD

All MI conditions led to contralateral brain modulation but not significantly different between them; thus, all the conditions may be capable of inducing lateralized ERD. However, only ME’s Alpha LI was not significantly lateralized. This is in line with prior research, and specifically, it is well known that ERD in the contralateral hemisphere becomes bilateral symmetrical during execution of movement [38]. Further, the laterality in the MIMO conditions could be an important component for rehabilitative BCIs, since motor commands for the arm and hand generally arise from the contralateral motor cortex, where most of the relevant corticospinal tract originates [43]. Specifically, latest results in MI-BCI interventions with stroke patients have shown a relationship between ERD lateralization with the functional improvement [44,45].

5. Conclusions

This paper aimed to find the VR–BCI condition that leads to the strongest and most lateralized ERD during MI, while adding additional evidence concerning the impact of haptics in MI-BCI training by contributing also with more than 100 labeled EEG datasets [36]. Based on the acquisition of EEG signals and an analysis of the Alpha and Beta ERD and LI, the use of a virtual environment, NeuRow, and haptic feedback – implemented as vibrotactile stimulation in this study – led to significantly stronger contralateral ERD, which were comparable to motor execution. Furthermore, VR HMD did not lead to such results by itself, being comparable to just using a computer monitor without haptic feedback. Lastly, all MI conditions invoked contralateral ERD as anticipated.

In terms of limitations, the current study is limited by its sample size, and also by the fact that ME was available only for 10 subjects. Hence, conclusions obtained from the statistical tests need to be translated cautiously; however, they are able to illustrate important trends, and in-alignment with prior literature.

Finally, in terms of future work, more studies should assess the sense of embodiment through a questionnaire, which is not normally employed in studies that make use of immersive VR technology. Particularly, this could also help answer whether a VR headset is warranted in getting a stronger ERD for stroke rehabilitation, given its additional cost and complexity over a standard monitor. Moreover, the comparison of these conditions needs to be extended in a closed-loop BCI, assessing feature discriminability and the online performance of the classifier. Ultimately, future work needs to investigate the long-term effect of VR-BCI training through the practice of embodied and haptic feedback through a longitudinal study. Specifically, the impact on the activity of brain patterns of chronic stroke patients, measured not only by neuroimaging (e.g. EEG, fMRI), but also clinical scales (e.g. Fugl-Meyer Assessment scale).

Acknowledgements

This research was funded by the Fundação para a Ciência e Tecnologia (FCT) through CEECIND/01073/2018, the NeurAugVR (PTDC/CCI-COM/31485/2017), the LARSyS—FCT (UIDB/50009/2020), and NOISyS (2022.02283.PTDC) Projects.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

The work was supported by the Fundação para a Ciência e a Tecnologia.

Notes

1. https://www.youtube.com/watch?v=zbAKJiHulIw

2. https://github.com/LaSEEB/NeuXus

3. https://github.com/sccn/labstreaminglayer