Participants were recruited from the greater Madison, Wisconsin, United States area as part of an on-going cross-over controlled prospective randomized rehabilitation study investigating interventional BCI in upper extremity (UE) motor function impairment resulting from stroke. This study is approved by the University of Wisconsin Health Sciences Institutional Review Board (Study ID 20150469). All subjects provided informed written consent upon enrollment.

The BCI system and intervention procedures were consistent with previously published works, Wilson et al. (2009, 2012) and Song et al. (2014) using BCI2000 software (Schalk et al., 2004) version 2 with in-house modifications for input from a 16-channel EEG cap and amplifier (Guger Technologies) and integration with a ball and target gaming visual display (Wilson et al., 2009) and functional electrical stimulation (FES) adjuvant (Popovic et al., 2009; Takahashi et al., 2012; Biasiucci et al., 2018). A more detailed and complete description of the session and run procedures used in data acquisition and analysis reported in this manuscript, as well as a description of the functional electrical stimulation protocol and justification for how this approach further improves the effects of BCI have been published previously (Remsik A. B. et al., 2019).

Sixteen right hemisphere stroke survivors (8 female, age = 62.5 ± 12.7 years [mean ± SD]) (Table 1) participated in 9–15 sessions (13.9 ± 1.28 [mean ± SD]; following removal of artifact EEG data: 11.1 ± 3.87 [mean ± SD]) with the BCI. Participants executed hand movements in response to visual cues displayed on a computer screen concordantly with corresponding audio instructions (e.g., Left, Right, Rest). One session (or run) of the cursor and target task consisted of 10 trials, or attempts, during which the participant attempted to drive a virtual cursor across the display screen into the target space using voluntary modulation of their sensorimotor rhythms (i.e., satisfaction of BCI classifier) before the trial timed out. BCI performance was measured as the number of successful trials (i.e., hits compared to misses) “out of a possible 10” attempts in any one run. This metric in included in Table 1 as “Average BCI Performance.”

The BCI classifier was defined in “screening” sessions (Pre and Post BCI intervention). Screening sessions contained two runs, each consisting of 15 trials for rest, left-hand, and right-hand movements (i.e., 5 trials for each of the three conditions, the order of trials in a run was random). Only the rest and left-hand movement trials are considered here.

From the larger NCT02098265 cohort, only participants with left hand motor impairment were selected for these sub analyses in an effort to control for homogeneity of stroke-related upper extremity impairment, and language or communicative deficits that might interfere with comprehension of the BCI task requirements or task execution.

Statistical differences between post- and pre-intervention cortical estimates were tested using paired-sample t-statistics as part of the statistical non-parametric mapping approach (SnPM) implemented in the s/eLORETA software (voxel-based). Paired-sample t-statistics were also used to test differences in the ROI estimates (power and connectivity). Finally, the Pearson correlation coefficient was used to quantify the correlation between change in hand grip strength and change in Mu rhythm desynchronization (Post vs. Pre BCI intervention). The p values reported are uncorrected for multiple comparisons.

Exact Low Resolution Electromagnetic Brain Tomography (eLORETA) was used to estimate the three-dimensional current density in the brain from the measured scalp EEG spectral data. In the eLORETA implementation, computations are made in a realistic head model using the MNI152 template, with the three-dimensional solution space restricted to cortical gray matter, as determined by the probabilistic Talairach atlas.

The resulting distributions of ERD and ERS heatmaps of Figure 3 illustrate the distribution of ERD and ERS changes from Pre to Post BCI intervention in the n = 16 stroke survivors represented in Figure 4. Considering the distribution of ERS (red) and ERD (blue) voxels in Figure 3, it is apparent that as a result of BCI intervention, participants realized an increase of Mu ERD in ipsilesional hemisphere voxels of sensorimotor cortices during task performance (i.e., hand grasping) with the affected (left) hand compared to rest.

Figure 4 shows pre-BCI intervention (blue) and post-BCI intervention (orange) computed eLORETA cortical spectral power estimates for the Mu [8–12 Hz] band (i.e., 1 average per each of the 16 subjects included in these sub analyses, separately for movement (left) and rest trials) for each of the ten regions of interest and assumed underlying Brodmann’s Areas (Figure 2). ERD increases (Post vs. Pre BCI intervention) were observed in regions of interest ROI #6 through ROI #9. The largest increases in Mu ERD were observed for regions of interest ROI #7 and ROI #8, which correspond to ipsilesional primary motor cortex (BA 4) (paired-samples t-test: p = 0.11) and ipsilesional somatosensory association area (BA 5) (paired-samples t-test: p = 0.056), respectively (Figure 4).

Figure 5A plots change in hand grip strength as a function of change in Mu rhythm desynchronization (Post vs. Pre BCI intervention) in the primary motor cortex of the ipsilesional hemisphere (ROI #7). Each data point represents an individual subject. Please note that negative numbers in the change in Mu rhythm desynchronization (i.e., Post-Pre), represent higher desynchronization in the Post phase. That is, larger negative numbers subtracted from smaller negative numbers will yield negative numbers. Improved hand grip function is positively correlated with increased Mu rhythm desynchronization (Post-Pre) in the ipsilesional primary motor cortex (Pearson’s correlation coefficient r = 0.435, p = 0.046).

A voxel-wise correlation approach (refer to Methods section) was used to represent this relationship in brain atlas space (Figure 5B). The heatmaps show colored r values (e.g., Pearson’s correlation coefficient values) and indicate that the correlation between improvement in hand grip strength and increase in Mu rhythm r values is strongest in voxels representing ipsilesional cerebral cortical sensorimotor areas.

In summary, the above results suggest that for the stroke-lesioned hemisphere, BCI intervention facilitates increased Mu desynchronization associated with movement rehabilitation of the impaired upper extremity.

Direct paths of Pre to post BCI intervention frequency-specific flow of information transmission between cerebral cortical sensorimotor areas were analyzed by computing isolated effective coherence (iCoh) developed by Pascual-Marqui et al. (2014).

Figure 6 plots iCoh values as a function of spectral frequency, Pre (blue) and Post (red) BCI intervention, to and from the 10 regions of interest (ROI #1-10) for the affected (left) hand. Mean values (Pre, Post) were compared via paired-sample t-statistics, thresholded at t = 2.13 [df = 15], p = 0.05 uncorrected).

Figure 7 summarizes t-statistics for the iCoh data shown in Figure 6. Statistically significant increases and decreases in iCoh at relative frequencies from Pre to Post BCI intervention between pairs of ROIs are plotted in red and blue, respectively. Significant increases in iCoh from Pre to post BCI intervention were seen going from ROI #7 to ROI #9, suggesting more causal informational flow is going from ROI #7 (ipsilesional primary motor cortex) to ROI #9 (ipsilesional premotor cortices) Post compared to Pre BCI intervention. Significant decreases in iCoh were seen going from ROI #8 to ROI #5, suggesting less causal informational flow is going from ROI #8 (ipsilesional somatosensory association area) to ROI #5 (contralesional visuomotor coordination area) Post compared to Pre BCI intervention.

Signals in both Mu [8–12 Hz] and Beta [18–26 Hz] frequency ranges served as input commands in the present BCI design (Remsik A. B. et al., 2019). Figures 8, 9 summarize for Mu [8–12 Hz] and Beta [18–26 Hz] signal ranges, respectively, changes in functional connectivity of sensorimotor cerebral cortices during attempted movement of the impaired upper extremity as a result of BCI intervention. Each of the panels of Figures 8, 9 compares [Pre vs Post] iCoh values in Mu (Figure 8) and Beta (Figure 9) frequency ranges and represents a direct path of intracortical causal information flow of oscillatory activity between two ROIs. Statistically significant differences between Pre (blue bars) and Post (red bars) iCoh values between pairs of ROIs are represented by surrounding boxes with blue denoting significant decreases in iCoh, and red boxes signifying significant increases in iCoh.

In the Mu [8–12 Hz] frequency band, iCoh values significantly decreased Pre vs Post BCI intervention from ROI #4 (contralesional premotor area) to ROI #8 (ipsilesional somatosensory association area) (Figure 8). Mu band iCoh values significantly increased Pre vs Post BCI intervention from ROI #7 (ipsilesional primary motor) to ROI #9 (ipsilesional premotor cortices).

In the Beta [18–26 Hz] frequency band, iCoh significantly decreased Pre vs. Post BCI intervention from ROI #8 (ipsilesional somatosensory association area) to ROI #5 (contralesional visuo-motor coordination area) (Figure 9). Beta band iCoh values significantly increased Pre vs. Post BCI intervention from ROI #5 (contralesional visuo-motor coordination area) to ROI #9 (ipsilesional premotor cortices), from ROI #6 (ipsilesional primary somatosensory areas) to ROI #10 (ipsilesional visuo-motor coordination area), and from ROI #7 (ipsilesional primary motor cortex) to ROI #9 (ipsilesional premotor cortices) (Figure 9).

In a less tightly controlled analysis of available NCT02098265 participants at the time, our laboratory demonstrated (Remsik A. et al., 2019) that ipsilesional Mu [8–12 Hz] ERD increases, recorded over cerebral cortical sensorimotor areas by scalp electrodes C3/C4 (Figure 1), were correlated with motor recovery of the affected upper extremity in stroke survivors. Here, we confirm and expand upon our previous findings in three ways. First, we present a more detailed source mapping to cortical space of Mu rhythm desynchronization and show that motor recovery is predominantly associated with Mu ERD changes ipsilesionally, near the primary motor cortex (Brodmann Area (BA) 4) and somatosensory association area (BA 5) (Figure 3). Second, we show that greater Mu desynchronization in ipsilesional primary motor cortex (BA 4) is significantly correlated with improvement in hand grip strength, an objective, quantitative measure of UE function and capacity, following BCI-FES intervention (Figures 4, 5). Third, we show evidence of functional connectivity changes between ipsilesional sensorimotor cortices following BCI-FES intervention, which is consistent with the view that BCI-FES use supports and/or engages fundamental neural mechanics that link volitionally controlled EEG signal changes with changes in objective behavioral measures of motor function and capacity. Such mechanisms may be specific to motor recovery (Biasiucci et al., 2018; Bai et al., 2020) and/or may be signatures of motor learning in general (Yuan et al., 2008; Edelman et al., 2019).

Our results are in line with findings of other studies that have employed similar BCI paradigms in cohorts of stroke survivors with similar demographics (Daly et al., 2009; Ramos-Murguialday et al., 2013; Ang et al., 2014; Ono et al., 2014; Popovic, 2014; Hu et al., 2015; Kim et al., 2016; Biasiucci et al., 2018; Nishimoto et al., 2018; Tabernig et al., 2018; Bai et al., 2020; Simon et al., 2021). For example, Biasiucci et al. (2018) report that only their BCI-FES group exhibited significant functional improvement following BCI intervention. Our result that increased ipsilesional Mu desynchronization is significantly correlated with increased grip strength is in good agreement with the finding of Biasiucci et al. (2018) that only their BCI-FES group exhibited a significant and lasting increase of the strength of the muscle targeted by the FES (in their study: extensor digitorum communis, which elicits full extension of the wrist and fingers). Furthermore, our result that SMR desynchronization is largely ipsilesional, indicating greater activation of the ipsilesional motor system following BCI-FES intervention, is in keeping with previous reports in stroke survivors (Mihara et al., 2013; Li et al., 2014; Biasiucci et al., 2018; Bai et al., 2020). BCI-FES devices may interface with native underlying mechanisms of neuromotor plasticity and control that have been demonstrated to engage and facilitate brain changes and behavior changes indicative of clinical motor recovery.

Importantly, although voxel-wise distribution of eLORETA output revealed a sensorimotor-centered focus of Mu ERD change associated with motor recovery, ipsilesional-focused Mu ERD change may not be the sole driver of motor recovery in stroke survivors using a BCI. Mu signal changes in cortical sensorimotor areas may be part of a broader integration of disparate ROIs in a neural network that includes both intra- and interhemispheric components (Figures 8, 9), although recovery of the primary motor region has been demonstrated to be essential to motor recovery (Grefkes and Fink, 2014). Therefore, we sought to understand whether the underlying neural network is specific to localized neural populations represented by the Brodmann’s area segmentations or, whether there is a larger geography of populations – functionally connected neural networks – that may be influenced by participation in the BCI intervention. We used discrete data in time series to represent directionality or sequencing of activity associated with these changes to test the hypothesis that BCI intervention results in changes in frequency-specific directional flow of information transmission (functional connectivity) in ipsilesional and/or contralesional sensorimotor areas (Ward et al., 2003; Rehme et al., 2011; Dubovik et al., 2012; Várkuti et al., 2013; Nicolo et al., 2015; Pichiorri et al., 2015; Pundik et al., 2015; Wu et al., 2015; Biasiucci et al., 2018).

The changes in functional connectivity reported here, particularly between ipsilesional primary motor cortex and ipsilesional premotor cortices, agree with previous studies that have employed comparable BCI designs (Wu et al., 2015; Biasiucci et al., 2018). Increased interactions between ipsilesional motor brain areas are associated with better recovery and motor performance (Dubovik et al., 2012; Nicolo et al., 2015; Wu et al., 2015). The change in effective connectivity between ipsilesional motor cortex and premotor cortices following BCI intervention was observed for both Mu and Beta frequency ranges in both the present results (Figures 8, 9) and those of Biasiucci et al. (2018), which is consistent with the view that the functional improvement following BCI intervention is due to facilitation and induction of neuroplastic changes associated with motor recovery as well as motor learning more broadly.

BCI-mediated FES leads to the most significant recovery of motor function following stroke (Bai et al., 2020; Simon et al., 2021). The inclusion of FES is thought to contribute to the clinical effect of a BCI not only through the somatosensory contribution of facilitated muscle stimulation but also through pairing of volitionally modified CNS efferent signals with stimulation (i.e., activation) of the impaired distal muscle (Bergquist et al., 2011). Biasiucci et al. (2018) have demonstrated through their BCI-FES versus sham FES experimental design that it is not the FES alone but rather the contingency between rich sensory feedback and suitable activation of cortical motor areas that may drive activity-dependent, Hebbian plasticity that may underlie motor recovery in BCI-FES interventions. Whereas our study design does not allow us to draw the same conclusions as Biasiucci et al. (2018) with respect to the clinical significance of the FES adjuvant, both our results and BCI-FES device are similar to those of Biasiucci et al. (2018). Therefore, it is likely that the clinically relevant functional gains demonstrated here are due to the same strict contingency of BCI-driven FES detailed by Biasiucci et al. (2018).

Future research is needed to better identify and track the genesis and progression of neuroplastic changes, and to determine the relative importance of the various intra- and interhemispheric network connectivity changes presented here. In addition, further research is necessary to discover the mechanistic origins of any such neuroplasticity, and how such mechanism(s) may improve rehabilitation strategies that enable caregivers to provide maximal benefit to patients (Bai et al., 2020; Simon et al., 2021).

Localizing cortical neuronal signatures with EEG scalp electrodes is a process that neuroscience researchers and others have used for some time (Pascual-Marqui et al., 1994; Pascual-Marqui, 2007). It is generally believed that higher density electrode arrays in the EEG cap provide more reliable and accurate localization estimations of the underlying neurophysiological processes. The use of eLORETA in a 16-channel cap arrangement, in non-research settings, likely pushes the boundaries of this method’s practical limits as an estimation tool for source localization of motor-EEG brain signals. Co-registration of spectral EEG data to MNI atlas space is a first step toward addressing this issue. In this work, scalp electrodes, while limited in number (16 channels), specifically covered sensorimotor cortices with greater density than the conventional 10–20 electrode placement system. Arguments can also be made for the use of “individualized” head modeling over “standard” head models. The use and reference to Brodmann’s areas in this work are intended solely as representative labels of the brain segmentations proposed by Brodmann, and we assume that these areas are generally emblematic of the functional cortical brain areas recorded by the scalp EEG electrodes of this cap arrangement (Figure 1).

In future research, repeating or initiating a similar BCI study with high density or 19 channel standard electrode arrays might allow a more strict evaluation of the “boundaries” in channel count and “electrode density as spatial coverage over areas of interest” by comparing the source localization results from these findings to the aforementioned high- and low-density electrode setups (i.e., with “full” head coverage). With any channel count placed “strategically” with higher density than the basic n = 19 (i.e.,10-20 system) over local functional network areas of interest on the scalp, reliable recordings of neuro-electrophysiological functioning might be feasible.