Abstract
The robust, reciprocal anatomic connections between the cerebellum and contralateral sensorimotor cerebral hemisphere underscore the strong physiological interdependence between these two regions in relation to human behavior. Previous studies have shown that damage to sensorimotor cortex can result in a lasting reduction of cerebellar metabolism, the magnitude of which has been linked to poor rehabilitative outcomes. A better understanding of movement-related cerebellar physiology as well as cortico-cerebellar coherence (CCC) in the chronic, poststroke state may be key to developing novel neuromodulatory techniques that promote upper limb motor rehabilitation. As a part of the first in-human phase I trial investigating the effects of deep brain stimulation of the cerebellar dentate nucleus (DN) on chronic poststroke motor rehabilitation, we collected invasive recordings from DN and scalp EEG in participants (both sexes) with middle cerebral artery stroke during a visuo-motor tracking task. We investigated the excitability of ipsilesional cortex, DN, and their interaction as a function of motor impairment and performance. Our results indicate the following: (1) event-related oscillations in the ipsilesional cortex and DN were significantly correlated at movement onset in the low beta band, with moderately and severely impaired participants showing desynchronization and synchronization, respectively; and (2) significant CCC was observed during the isometric hold period in the low beta band, which was critical for maintaining task accuracy. Our findings support a strong coupling between ipsilesional cortex and DN in the low beta band during motor control across all impairment levels, which encourages the exploitation of the cerebello–thalamo–cortical pathway as a neuromodulation target to promote rehabilitation.
SIGNIFICANCE STATEMENT Cerebral infarct because of stroke can lead to lasting reduction in cerebellar metabolism, resulting in poor rehabilitative outcomes. Thorough investigation of the cerebellar electrophysiology, as well as cortico-cerebellar connectivity in humans that could provide key insights to facilitate the development of novel neuromodulatory technologies, has been lacking. As a part of the first in-human phase I trial investigating deep brain stimulation of the cerebellar dentate nucleus (DN) for chronic, poststroke motor rehabilitation, we collected invasive recordings from DN and scalp EEG while stroke survivors performed a motor task. Our data indicate strong coupling between ipsilesional sensorimotor cortex and DN in the low beta band across all impairment levels encouraging the exploration of electrical stimulation of the DN.
Introduction
Stroke is a leading global cause of long-term disability among adults and is rapidly becoming a major health burden in developing nations (Virani et al., 2021). Major technological advances for endovascular interventions as well as pharmacological thrombolysis have contributed to acute stroke care aimed at minimizing cerebral tissue loss via early reperfusion. Nevertheless, more than half of stroke victims maintain severe long-term disability, and more than two-thirds of survivors experience lingering paresis of the upper extremity despite intensive rehabilitation, warranting additional strategies to enhance the effects of physical rehabilitation and improve long-term outcomes (Go et al., 2014). Neuromodulation is an important technique with potential to increase the contributions of motor regions to recovery (Elias et al., 2018; Storch et al., 2021; Ting et al., 2021). Better understanding of physiological changes in networks that contribute to poststroke plasticity and rehabilitation is key to identifying novel, more suitable neuromodulation targets and better paradigms of stimulation.
The cerebral motor cortex and the contralateral cerebellar hemisphere are reciprocally connected through the (descending) cerebro–ponto–cerebellar (CPC) and (ascending) cerebello–thalamo–cortical (CTC) pathway loops (Evarts and Thach, 1969; Leiner et al., 1991; Ramnani, 2006; Palesi et al., 2017). CPC and CTC anatomy and physiology have been extensively investigated in healthy animal models using virus tracing and invasive neurophysiological techniques (Llinás et al., 1975; Rispal-Padel et al., 1981; Aumann et al., 1998; Pananceau and Rispal-Padel, 2000; Guder et al., 2020; Metoki et al., 2022). These studies have revealed the contribution of CPC and CTC as a synergistic system in motor planning, execution, and control. In the context of stroke, both clinical and preclinical models have adopted noninvasive imaging and metabolic approaches. Findings have uncovered the phenomenon of crossed cerebellar diaschisis (Baron et al., 1981; Lenzi et al., 1982; Pantano et al., 1986; Nathan et al., 1994; Sommer et al., 2016; Chan et al., 2017), the rapid and lasting reduction in cerebellar metabolism as a consequence of cerebral cortical injury. Clinically, crossed cerebellar diaschisis has been linked to poor rehabilitative outcomes after stroke (Takasawa et al., 2002; Kunz et al., 2017; Sin et al., 2018; Kim et al., 2019), highlighting the importance of interdependence between cerebellum and motor cortex for motor recovery. While the cerebral cortical areas involved in motor processing can be electrophysiologically examined with noninvasive techniques such as electroencephalography (EEG) and magnetoencephalography (MEG), noninvasive assessment of cerebellar electrophysiology in humans is challenged by the limited signal-to-noise ratio and spatial resolution of EEG and MEG (Andersen et al., 2020). The opportunity for direct electrophysiological examination of cerebellar and cortical connectivity online during movement is limited and has not been possible to date in humans.
We present the first in-human examination of the cerebellar dentate nucleus (DN) electrophysiology and cortico-cerebellar connectivity during movement planning and execution in chronic poststroke survivors, enabled using invasive recordings from the DN and EEG. We initiate this examination by characterizing oscillatory activity of the ipsilesional cortex and DN at the onset and offset of movement execution with the paretic extremity, followed by interactions between ipsilesional cortex and DN during isometric motor control. We then evaluate the impact of motor impairment severity on ipsilesional cortex and DN oscillatory power during motor execution as well as on coherence between cerebral cortex and DN during task performance. Last, we examined whether there were relationships between electrophysiological findings and performance on motor tasks. We believe that this knowledge can be relevant to groups developing novel clinical approaches and is important in our own efforts to translate the use of deep brain stimulation (DBS) targeting the cerebello–thalamo–cortical pathway for promoting poststroke rehabilitation (Machado and Baker, 2012; Wathen et al., 2018; Cooperrider et al., 2020).
Materials and Methods
Participants
Data are derived from a single-site, phase I clinical trial (Clinicaltrials.gov ID NCT02835443) investigating the safety and feasibility of DN DBS in chronic, poststroke motor rehabilitation (Machado and Baker, 2012; Wathen et al., 2018; Cooperrider et al., 2020). The study included participants of either sex. All research activities were approved by the Cleveland Clinic Institutional Review Board with participants provided written informed consent. The study was monitored under an Investigational Device Exemption from the US Food and Drug Administration. All recruitment and data collection activities happened between June 2016 and October 2020. Eligibility was determined by the participation in the aforementioned clinical trial. All participants enrolled in the trial underwent surgery for implantation of DBS leads in the DN contralateral to the affected cerebral hemisphere. The study was conducted using the Vercise DBS System manufactured by Boston Scientific. Of the 10 participants, 5 were implanted with Vercise Cartesia eight-contact directional lead that has a distal active tip, a proximal annual contact and two levels of three radially arranged directional contacts in between (contact length, 1.5 mm; contact spacing, 2 mm; contact span, 7.5 mm). The remaining participants had a linear eight-contact lead that consists entirely of annular contacts (contact length, 1.5 mm; contact spacing, 2 mm; contact span, 15.5 mm). Figure 1C shows the localization of the contacts in the DN for each participant in axial, coronal, and sagittal planes.
Data collection
All electrophysiological data were collected during a 5–7 d externalization period immediately following surgical implantation of the DBS lead. As summarized in Figure 1A, local field potential (LFP) data were recorded from the region of the DN by means of a temporary, percutaneous extension attached to the implanted DBS lead. Scalp EEG was acquired using 36 standard, Ag/AgCl cup electrodes placed according to the international 10–10 system. Our montage included all of the channels of the 10–20 system, with additional electrodes placed at AF7, AF3, AFz, AF4, AF8, FT9, FC5, FC1, FC2, FC6, FT10, TP9, CP5, CP1, CP2, CP6, and TP10. Data were referenced online to Pz. The impedance was kept <5 kΩ. Electromyography (EMG) was recorded using a pair of bipolar surface electrodes from the flexor carpi radialis (FCR) muscle during the grip task used in this study. All electrophysiological data were collected at a sampling rate of 5000 samples/s with online high-pass filter of 0.1 Hz using the BrainAmp DC with BrainVision Recorder software (Brain Products). Finally, force exerted during the grip task was recorded separately from a strain gauge-based isometric hand dynamometer (model HD-BTA, Vernier) using a Powerlab interface (model PL3516-0158, ADinstruments) at 1000 samples/s and stored for offline analysis [LabChart (version 8.0), AD Instruments].
Motor function of the upper limb was assessed at baseline before implant surgery by an occupational therapist (Table 1). Upper Extremity Fugl-Meyer (UEFM) assessment was used as a measure of motor impairment. UEFM assessment is a common, reliable, and valid instrument for the assessment of upper limb motor impairment following stroke (Fugl-Meyer et al., 1975). For each participant, we computed the following: (1) the UEFM total score, which is the sum total of upper extremity, wrist, and hand functions (total of subsections A, B, C, and D of the UEFM assessment) with a maximum possible score of 66; (2) the UEFM hand score, which is the sum total of wrist and hand functions, including coordination and speed (total of subsections B, C, and D of the UEFM assessment) with a maximum possible score of 30. The lower the UEFM score, the higher the impairment.
Demographics and task details
Task
Participants performed a block-tracking motor control task with visual force feedback that involved controlling the vertical displacement of a 2-D ball on a computer display to track a scrolling, square-wave trace as accurately as possible. The participants sat upright in a chair with their forearms resting comfortably and were instructed to keep the ball on a scrolling trace by exerting the required grip forces using the dynamometer. The block-tracking paradigm had the following four distinct phases (Fig. 1A): (1) contraction (movement onset), labeled as On in the figure; (2) hold, labeled as H; (3) release (movement offset), labeled as Off; and (4) relaxation, labeled as R. The hold phase was calibrated to 30% of each participant's maximal voluntary contraction. The duration of the hold and relaxation phases lasted for 5 or 2.5 s, depending on each participant's baseline ability to maintain an isometric force. Depending on whether the participant tracked a square wave with 5 or 2.5 s plateau, they tracked 20 or 40 square-wave trials per block, respectively. Participants, on average, performed three blocks of tracking that lasted 100 s each, with a 3–5 min rest period between each block. Task performance was quantified using an accuracy index (Carey et al., 2002), which is the root mean square error between the target force trace and the actual force trace for each block and was averaged across blocks for each participant.
Data preprocessing
The electrophysiological data were analyzed using the MATLAB-based FieldTrip toolbox (Oostenveld et al., 2011) and custom-written scripts. EEG data were first visually inspected and channels with low signal-to-noise ratio were removed, especially along the temporal side because of their susceptibility to muscle artifacts arising from compensatory movements typical of the poststroke state. Continuous EEG, LFP, and EMG data recorded over the whole block were parsed into trials with an interval of −3.5 to 8.5 s relative to movement onset (for 5 s square-wave tracking) or −2.25 to 4.75 s relative to movement onset (for 2.5 s square-wave tracking). The timing of the movement onset and offset were detected from high-pass filtered, Hilbert-transformed, and smoothed EMG muscle activity. Line noise in the data was eliminated using a DFT (Discrete Fourier Transform) filter or spectral interpolation method. All electrophysiological data were then downsampled to 400 samples/s. Further, muscle, EKG, and ocular artifacts in EEG were detected and removed using an independent component analysis (Jung et al., 1998). EEG data were detrended to remove linear low-frequency trends and current source densities computed using a surface Laplacian technique (Hjorth, 1975). EEG data were left–right flipped so that data from all the participants were aligned to a common (right) side reflecting the affected hemisphere. Ipsilesional cortex was defined by the following channels: F4, FC2, FC6, C4, CP2, CP6, and P4. We did not include the central channels (Fz, Cz, or Pz) because of possible contamination of activity from mirror movement of the nonaffected extremity.
To estimate activity in the DN, a bipolar derivative of the contacts of the DBS lead was performed. As a result, participants with an eight-contact linear lead had seven bipolar channels (namely, LFP1-2, LFP2-3, LFP3-4, LFP4-5, LFP5-6, LFP6-7, and LFP7-8). In participants with eight-contact directional leads, the two levels of three radially arranged directional contacts were pooled to compute the average data and mimic an omnidirectional contact, resulting in a total of four contacts: LFP1, LFP234, LFP567, and LFP8. A bipolar derivative of these four contacts led to three channel combinations in these participants (namely, LFP1-234, LFP234-567, and LFP567-8).
Data processing goals and methods
To investigate the excitability of ipsilesional cortex and DN, we performed an event-related desynchronization/synchronization (ERD/S) analysis (Pfurtscheller and Lopes da Silva, 1999). The ERD/S analysis quantifies the frequency-specific changes related to the movement phase that represents the synchrony or desynchrony of the underlying neuronal populations. Previous studies (Pfurtscheller and Lopes da Silva, 1999) have shown movement preparation and execution results in desynchronization (i.e., cortical activation with increased excitability), while movement cessation results in synchronization (i.e., cortical inhibition with decreased excitability).
To investigate the interaction between the ipsilesional cortex and DN, also referred to here as cortico-cerebellar coherence (CCC), we performed a coherence analysis (Bastos and Schoffelen, 2015) during steady-state contraction (i.e., hold period following movement onset). Previous studies (Kilner et al., 1999; Omlor et al., 2007) have shown that phase coherence is most pronounced during steady-state contractions and is abolished during the preceding phasic movements. To investigate the directionality of interaction between the ipsilesional cortex and DN during peak coherence, we further performed a Granger causality (GC) analysis (Bastos and Schoffelen, 2015).
To investigate whether the above electrophysiological measures (ERD/S, CCC, and GC) varied as a function of motor impairment and motor performance, we performed a generalized linear mixed-models (GLMMs) analysis (Frömer et al., 2018) and a correlation analysis, respectively.
ERD/S analysis.
Artifact-free, preprocessed EEG and LFP time series from each trial were decomposed into their time-frequency representations in the 5–45 Hz range with frequency steps of 1 Hz. A 7-cycle Morlet wavelet was used for the continuous wavelet transformation. The raw power spectra were then log-transformed, and event-related power change was calculated by averaging the normalized power of all time-points in each frequency bin relative to the average power of that frequency across the whole experimental session. ERD/S power in the ipsilesional cortex and DN at movement onset and offset was defined as the normalized average relative power over a 2 s window centered at movement onset and offset. For the purpose of GLMMs analysis (described later) for the ERD/S data, we considered the three primary channels over the ipsilesional motor cortex, namely CP2, C4, and FC2, and all LFP contacts in the DN.
Cortico-cerebellar coherence
The hold phase was defined as the steady contraction period following movement onset ramp [i.e., the interval between 2 and 4 s for 5 s block tracking and between 1 and 2 s for 2.5 s block tracking]. This period showed relatively stable force production across participants (Fig. 1A) and outside of power fluctuations caused by phasic movements. The relaxation phase was defined as the period preceding each movement onset ramp [i.e., the interval between −2 and −4 s for 5 s block tracking and between −1 and −2 s for 2.5 s block tracking]. This period showed minimal force production across participants. Although the relaxation phase could be considered as the baseline period, we chose to analyze it with equal importance given the dynamic nature of the tracking task. We computed the CCC (separately for hold and relaxation periods) for each combination of the DN LFP contact and scalp EEG channels using a methodology described previously (Rosenberg et al., 1998), after partializing for the contribution of the EMG. Coherence in the frequency domain quantifies the consistency of the phase difference and amplitude across trials between two signals. CCC was computed across the frequency (5–45 Hz, in steps of 0.5 Hz) coefficients derived using a multitaper method with spectral smoothing set to 3. A coherence value of 0 indicates that the signals being compared have no consistent linear phase relationship, whereas a value of 1 indicates that the two signals are fully phase coherent. The CCC values were further z transformed (Maris et al., 2007) to correct for the variable number of degrees of freedom (trials numbers) across participants.
Significance of CCC was assessed using a nonparametric cluster-based permutation approach (Maris and Oostenveld, 2007) separately for hold and relaxation periods. The main purpose of this permutation analysis was to select one candidate DN LFP contact that exhibited the largest coherence cluster on the ipsilesional cortex EEG channels. We compared the observed CCC on the ipsilesional cortex with a surrogate “null” distribution of maximal CCC values between 5 and 45 Hz. For each participant, we randomly shuffled trials in the LFP data, while retaining the order of trials of the EEG data. Then, we computed the CCC of the shuffled data. After 500 permutations, a distribution of maximum CCC values of shuffled data was created for every LFP–EEG pair. The threshold for every LFP–EEG pair was determined by choosing a p-value at p = 0.01 of this null distribution. After eliminating the LFP–EEG pairs (from the observed data) that did not show CCC exceeding this threshold, we examined spatially the size of the maximal coherence cluster induced by each LFP contact on the ipsilesional cortex. A spatial cluster was defined as a minimum of three neighboring EEG channels on the ipsilesional cortex that showed CCC beyond the significance threshold. In each participant the LFP contact that exhibited the largest coherence cluster between 5 and 45 Hz and the associated three EEG channels during the hold period were chosen for further statistical analysis using GLMMs analysis (described later).
Granger causality.
To investigate causal interactions between the DN LFP contact that exhibited the largest coherence cluster on the ipsilesional cortex EEG channels, we performed a nonparametric spectral GC analysis (Bastos and Schoffelen, 2015) separately for the hold and relaxation phases. To compute GC, transfer functions and noise covariances were estimated through the spectral factorization of Fourier coefficients (Dhamala et al., 2008). A signal, x, is said to “Granger cause” signal y if the prediction of y can be improved by incorporating knowledge of signal x, in addition to the past values of signal y itself.
GLMMs analysis.
We probed the ERD/S and coherence/GC measures to test whether they were influenced by time, frequency, and/or participant motor impairment. This analysis was performed regardless of whether the observed outcome exceeded the significance threshold determined earlier. These were performed using a series of GLMMs. GLMMs (Frömer et al., 2018) are flexible extensions of linear regression that allow for multilevel and non-Gaussian response data. In these models, high-dimensional data do not have to be averaged to a single data point per participant. Instead, all data for all participants are retained, and the clustering inherent in the data are accounted for in the model. For example, observed outcome (e.g., ERD/S power, CCC, GC) clustered within three neighboring EEG channels (e.g., FC2, C4, CP2) clustered within participants. Participants and channels/contacts were treated as random effects, and a compound symmetry covariance structure was applied for each model. Fit was evaluated using Bayes information criterion and visual examination of residuals. For ERD/S power, data were multivariate normal, and effect sizes were the power difference between or among states. For CCC/GC, data were multivariate log-normal, and effect sizes were the percentage differences between or among states.
For time, initial analyses (without any interactions) compared the following two time points: onset versus offset for ERD/S power, and hold versus relaxation for CCC/GC measures in the 5–45 Hz frequency range in all participants. For frequency, initial analyses compared alpha (8–12 Hz), low beta (13–20 Hz), high beta (21–30 Hz), and gamma (31–45 Hz) frequencies while including all time points (onset/offset or hold/relaxation) and all participants. For motor impairment, initial analysis compared severely impaired (UEFM hand score, ≤6) versus moderately impaired (UEFM hand score, >6) participants at all time points and across the entire frequency range of 5–45 Hz.
Secondary models separately tested interactions between time and frequency, time and participant impairment, and frequency and participant impairment. Time, frequency, and each interaction were considered fixed effects in each model. Analyses were conducted using SAS Studio version 3.7 and were two tailed with significance indicated by p < 0.05. Because of the exploratory nature of the study, correction for multiple comparisons was not applied.
Correlation analysis.
While CCC represents the phase synchrony between the ipsilesional cortex and DN during a steady state, it does not account for concurrent excitability between the two areas during phasic movements. Hence, we examined the correlation between ERD/S power in the ipsilesional cortex and ERD/S power in the previously chosen DN LFP contact averaged across periods of movement onset and offset. We further examined whether ERD/S power in the ipsilesional cortex and DN LFP contact during movement correlated with motor performance (indexed by the accuracy of task execution). To identify the spatial extent of correlation on the cortex, a cluster-based permutation technique was used. A simple linear regression was applied to the plots comparing the ERD/S power for each EEG channel and variables of interest (DN ERD/S and accuracy) to determine the gradient slope. The null hypothesis was defined as a zero gradient (i.e., no linear relationship). An empirical null distribution was formed by shuffling the order of the variable of interest, while retaining the order of the EEG ERD/S power. The gradients in the actual data were then compared with this null distribution generated from 2000 permutations, and a p-value was measured by integrating the null distribution between the actual gradient and infinity divided by the total integral. Ipsilesional sensors with gradients at a level of p = 0.01 were considered significantly correlated with the variable of interest. A cluster was defined as a minimum of one EEG channel. Using the same procedure, we examined whether CCC or GC measures (between the chosen LFP contact that exhibited the largest coherence cluster on the ipsilesional cortex and corresponding EEG channels on the ipsilesional cortex) averaged across periods of hold and relaxation phases were significantly correlated with motor performance.
Results
Demographics, motor impairment, and motor performance
A total of 15 participants were screened for eligibility, and only 12 participants were enrolled in the DN DBS clinical trial. However, data presented in this study were not acquired from the first two participants as the investigators were beginning to establish the safety and feasibility of DN DBS and did not externalize the leads. The dataset included 10 participants with severe or moderate-to-severe upper extremity hemiparesis after middle cerebral artery ischemic stroke and their demographics are summarized in Table 1. The extent of stroke lesion in each participant is illustrated in Figure 1B. The mean age was 57.3 ± 7.21 years (age range, 48–70 years). Two participants were female, and six participants had stroke on the right hemisphere affecting their left extremity.
Data collection setup, stroke morphology, and DBS lead localization. A, Continuous electrophysiological data were collected from scalp (EEG), cerebellar DN (LFP), and the FCR (EMG), while the participant performed a visuo-motor block-tracking task using a hand-held dynamometer. All data were parsed offline and time locked to movement onset (red dotted line). EMG and force traces aided in determining movement onset/offset. B, T1-weighted MRI at 3 T with segmentation (in red) of stroke-associated hypointensity. Left hemisphere is shown as the right side of the MRI, and vice versa. C, Implanted DN from each participant (shown in gray dots) were aligned relative to their principal axes (inferior to superior, I to S; posterior to anterior, P to A; lateral to medial, L to M). Each contact (large colored dots) is labeled with the participant ID number. Contacts chosen for the analysis based on surrogate statistics are shown in green, while others are shown in red.
The average UEFM total and UEFM hand score of the total sample were 23.4 ± 6.7 and 7.2 ± 3.4, respectively. Four of the 10 participants had a UEFM assessment hand score of <6, and were categorized as severely impaired, while the rest of the participants were categorized as moderately impaired. We chose to use the UEFM hand score as the basis for participant stratification and further analysis, because it better reflected the distal hand function required for the block-tracking task. Both the mean UEFM total score (17.5 ± 3.1 vs 27 ± 1.4; p = 0.0095; ranksum = 10, Wilcoxon rank-sum test) and the mean UEFM hand score (3.7 ± 1.2 vs 11 ± 1.4; p = 0.0095; ranksum = 10, Wilcoxon rank-sum test) differed significantly between high and moderately impaired participants.
All 10 participants completed the block-tracking task successfully. Six participants completed 5 s block tracking, while four participants completed 2.5 s block tracking. Participants who completed 5 s block tracking performed an average of 63 ± 18 trials. Participants who completed 2.5 s block tracking performed an average 97 ± 49 trials. Participants that tracked a 5 s block had a mean UEFM total score and a mean UEFM hand score of 24.3 ± 2.7 and 8.16 ± 2.7, respectively. Participants that tracked a 2.5 s block had a mean UEFM total and UEFM hand scores of 22 ± 11 and 5.7 ± 4.3, respectively. No significant differences were noted between the groups. The experiments were conducted with tailoring of tasks to participants' level of impairment such that those with moderate impairment favored 5 s tracking and those with severe impairment favored 2.5 s tracking.
The average percentage accuracy for block tracking among all 10 participants was 50.7 ± 17.9%. When stratified based on impairment, the severely impaired participants' accuracy was 42.5 ± 13.2%, while moderately impaired participants' accuracy was 56.2 ± 19.5%. No significant differences were noted between the groups.
ERD/S power
Qualitative analysis of ERD/S power in the ipsilesional cortex
Peak changes in time–frequency maps were observed in channel CP2 over the ipsilesional cortex in the low beta band. Participants who performed a 5 s block tracking (Fig. 2A) showed a mean percentage relative power change (ERD) of −2.13 ± 1.72% and −2.15 ± 1.78% at movement onset and offset, respectively, in the low beta band. Participants who performed a 2.5 s block tracking (Fig. 2B) showed a mean percentage relative power change of 0.62 ± 1.87% and −0.33 ± 0.31% at movement onset and offset, respectively, in the low beta band.
Ipsilesional cortex ERD/S statistics A, B, Time–frequency map from CP2 and line plot derived from low beta (β) band of movement-related activity averaged from six participants who performed a 5 s block-tracking task (A) and four participants who performed a 2.5 s block-tracking task (B) using their affected extremity. Color bar shows z-axis scales from −5% to 5%. Note: five of six participants who performed the 5 s block tracking were moderately impaired participants, while three of four participants who performed the 2.5 s block tracking were severely impaired participants. C, Time × frequency interaction shows that the largest differences in mean ERD/S power between onset and offset occur at high beta and gamma (γ). D, Time × impairment interaction shows ERD/S power was higher among severely impaired participants at onset, whereas at offset ERD/S power was higher among moderately impaired participants. C and D also indicate the main effect of time, such that the mean ERD/S power is consistently higher at onset than offset (i.e., more synchronization at onset than offset). E, Frequency × impairment interaction shows that the ERD/S power was higher in the alpha (α) and low beta band in severely impaired participants. *p < 0.05, **p < 0.01.
GLMMs analysis for ERD/S in the ipsilesional cortex
Time.
Mean ERD/S power was significantly higher at movement onset than at offset (p < 0.001), suggesting that ipsilesional cortical synchronization was greater at movement onset than offset (data not shown). Time × frequency interaction showed that ERD/S power at high beta frequency (p = 0.005) and gamma frequency (p < 0.001, Fig. 2C) were significantly greater at onset than at offset, when compared with other frequency bands. Time × impairment interaction showed that ERD/S power was higher among severely impaired participants at onset only (p < 0.001; Fig. 2D).
Frequency.
Mean ERD/S power was not significantly different across frequency bands. However, frequency × impairment interaction showed that mean ERD/S power was significantly higher in the alpha (p < 0.001) and low beta band in severely impaired participants (p = 0.002; Fig. 2E).
Qualitative analysis of ERD/S power in the DN
Unlike the EEG data, LFP data from the DN had variability in the channels (i.e., contact pairs) because of the different DBS lead types used in the study. Hence, to get a qualitative view we averaged the DN LFP channels in each participant and group averaged the data into 5 and 2.5 s block-tracking datasets. Participants that performed a 5 s block tracking (Fig. 3A) showed a mean percentage relative power change (ERD) of −1.62 ± 1.84% and −1.09 ± 1.27% at movement onset and offset, respectively, in the low beta band. Participants who performed a 2.5 s block tracking (Fig. 3B) showed a mean percentage power change (ERS) of 2.77 ± 2.84% and ERD of −0.27 ± 2.44% at movement onset and offset, respectively, in the low beta band.
Cerebellar DN ERD/S statistics. A, B, Time–frequency map from averaged DN LFP channels and line plot derived from low beta (β) band of movement-related activity averaged from six participants who performed a 5 s block-tracking task (A) and four participants who performed a 2.5 s block-tracking task (B) using their affected extremity. Color bar shows z-axis scales from −3% to 3%. Note: five of six participants who performed the 5 s block tracking were moderately impaired participants, while three of four participants who performed the 2.5 s block tracking were severely impaired participants. C, Significant time × frequency interaction shows the largest difference in ERD/S power between onset and offset in the gamma (γ) band. D, Time × impairment interaction shows that the ERD/S power was significantly greater in severely impaired participants compared with moderately impaired participants at onset only. *p < 0.05, **p < 0.01.
GLMMs analysis for ERD/S in the DN
Time.
Without any interactions, mean ERD/S power was significantly greater at onset than offset (p < 0.001; i.e., more synchronization at onset than offset; data not shown). When applying a time × frequency interaction (Fig. 3C), the difference between onset and offset was more significant within the gamma band than any others (p = 0.001). When applying a time × impairment interaction (Fig. 3D), ERD/S was higher among severely impaired participants at onset (p < 0.001) but no differences were noted at offset.
Frequency.
There were no differences in ERD/S power among frequency bands. There was no frequency × impairment interaction.
Impairment.
Without any interactions (frequency or time), mean ERD/S power was significantly higher in severely impaired participants than in the moderately impaired participants (p = 0.003; data not shown).
Cortico-cerebellar coherence and Granger causality
Qualitative analysis of CCC and GC during hold and relaxation
Hold period.
All participants displayed CCC beyond the significance threshold determined from the surrogate distribution. The largest coherence cluster was observed in the perilesional areas (i.e., EEG channels encompassing frontocentral (FC), central (C), and centro-parietal (CP) areas covering the premotor, motor, and sensorimotor cortex; Fig. 4A). In some participants, the cluster also extended to frontal (F) and parietal (P) regions. Peak CCC between the identified EEG channels on the ipsilesional cortex and DN LFP contact across all 10 participants occurred at 15.83 ± 5.52 Hz (range, 10.15–26.17 Hz). Peak GC directed from ipsilesional cortex to DN occurred at 18.26 ± 9.05 Hz, and from DN to ipsilesional cortex occurred at 15.29 ± 4.66 Hz.
CCC/GC during hold and relaxation states. A, B, CCC/GC derived from a single DN LFP contact that showed maximal coherence cluster on the ipsilesional cortex across all participants during hold (A) and relaxation period (B). Each panel shows the average topographic profile (bottom) derived from the frequency range that exceeded the significance threshold (shown in the gray patch on the top line plot). The line plot in black (left y-axis) shows the CCC between the single DN LFP contact (labeled above the topography) and the highlighted channels (red dots) on the ipsilesional cortex. The blue line (right y-axis) shows the average GC directed from ipsilesional cortex to the DN, and the red line (right y-axis) shows the average GC directed from DN to ipsilesional cortex, indicating the directionality of communication during maximal CCC.
Relaxation period.
Only four participants showed peak coherence above the significant threshold determined from surrogate distribution, and all of them were moderately impaired participants (Fig. 4B). The peak CCC between the ipsilesional cortex and DN across these four participants occurred at 11.62 ± 2.28 Hz (range, 8.98–14.06 Hz). Peak GC directed from ipsilesional cortex to DN occurred at 11.13 ± 2.15 Hz, and DN to ipsilesional cortex occurred at 13.67 ± 2.36 Hz.
GLMMs analysis for CCC
Time.
Without interactions, CCC was not significantly different between the hold and relaxation states. The time × frequency interaction showed that, compared with the relaxation state, CCC was higher in the hold state in the low beta band, but not in other frequency bands (p < 0.027; Fig. 5A).
CCC/GC statistics. A, Time × frequency interaction showed the largest difference in CCC between the hold and relaxation states occurred at low beta (β) band. B, Frequency and impairment interaction showed that severely impaired participants have higher CCC only in the high beta and gamma (γ) bands, although the magnitude of CCC was small. C, Main effect of time showed that GC in both directions was higher during the hold state than the relaxation state. D, Main effect of frequency showed that high beta and gamma GC in both directions was significantly lower than in alpha (α) and low beta bands. *p<0.05, **p<0.01
Frequency.
Without interactions, low beta and alpha bands showed greater CCC than high beta and gamma bands (p < 0.001, data not shown). This is also evident from Figure 4 in individual participants where peak CCC occurs in the alpha/low beta range. The frequency × impairment interaction showed that CCC was higher for severely impaired than moderately impaired participants only in the high beta and gamma bands (p < 0.001; Fig. 5B).
GLMMs analysis for GC
Time.
Ipsilesional cortex → DN GC and DN → cortex GC were both significantly higher in the hold state than the relaxation state (both p = 0.001; Fig. 5C).
Frequency.
Without interactions, ipsilesional cortex → DN and DN → ipsilesional cortex GC was significantly greater in the alpha and low beta bands compared with the high beta and gamma bands (all p < 0.001; Fig. 5D).
Correlation analysis
Ipsilesional cortex ERD/S power versus DN ERD/S power during movement
Significant positive correlation was observed between ERD/S power at channel CP2 on the ipsilesional cortex and DN LFP contact in the alpha (p = 0.0003) and low beta (p = 0.0005) bands. (Fig. 6A).
Correlation statistics. A, Significant ipsilesional ERD/S power correlation with ERD/S power in the chosen DN LFP contact during movement onset and offset occurred at channel CP2 in the alpha (α) and low beta (β) bands. Topography on the top displays the gradient (slope) profile observed between each channel on the scalp and the DN LFP contact in the corresponding frequency bands. Channel CP2 is highlighted in red. ERD/S power at CP2 on the ipsilesional cortex was significantly positively correlated with ERD/S power in the DN LFP contact in the alpha and low beta bands. Each dot in the scatter plot represents a participant. ERD/S power <0 indicates desynchronization (ERD), and >0 indicates synchronization (ERS). B, Significant ipsilesional ERD/S power correlation with task accuracy occurred at channel CP2 in the low and high beta (β) bands. Topography on the top displays the gradient (slope) profile observed between each channel on the scalp and task accuracy in the corresponding frequency bands. Channel CP2 is highlighted in red. ERD/S power at CP2 on the ipsilesional cortex was significantly negatively correlated with task accuracy in the low beta and high beta bands. Each dot in the scatter plot represents a participant ERD/S power <0 indicates desynchronization (ERD), and >0 indicates synchronization (ERS). C, Significant positive correlation was observed between CCC (between the chosen DN LFP contact and corresponding ipsilesional channels) in the low beta (β) band and task accuracy.
ERD/S power versus task accuracy
Ipsilesional cortex.
Significant negative correlation was observed between ERD/S power at sensor CP2 on the ipsilesional cortex and task accuracy in the low beta (p = 0.004) and high beta (p = 0.0045) bands (Fig. 6B).
Dentate.
No significant correlations were found between DN ERD/S power and motor performance.
CCC/GC versus task accuracy
Significant positive correlation was observed between mean CCC in the low beta band (p = 0.01) and task accuracy (Fig. 6C). No significant correlations were observed between GC and motor performance.
Discussion
This is the first in-human investigation of DN physiology and DN–cortical physiological connectivity. It was enabled by invasive DN recordings combined with ipsilesional EEG time locked to visually cued computerized motor tracking task. We systematically examined cortical and cerebellar physiology as well as correlations of ERD/S power and connectivity between ipsilesional cortex and DN. All participants were stroke survivors with moderate to severe motor impairment undergoing a trial (Wathen et al., 2018) of DN-DBS for poststroke recovery. Our data revealed the following: (1) movement-related oscillations in the DN were highly correlated with ipsilesional cortex oscillations during movement, in the alpha and low beta band, with moderately impaired participants showing ERD and severely impaired participants showing ERS; (2) peak coherence between DN and ipsilesional cortex occurred in the low beta band, and this communication was critical for maintaining task accuracy; and (3) bidirectional communication between ipsilesional cortex and DN was evident during significant CCC in the alpha and low beta band.
Movement-related oscillations in the DN and ipsilesional cortex
Movement-related ERD/S phenomena have been widely studied in the sensorimotor cortex (Stancák and Pfurtscheller, 1996; Pfurtscheller et al., 1998; Pfurtscheller and Lopes da Silva, 1999). Increased movement-related ERD, particularly in the beta band, is associated with increased corticospinal and cortical excitability (Pfurtscheller and Lopes da Silva, 1999; Takemi et al., 2013), resulting in facilitation during the generation of movement. In poststroke survivors, a decreased ERD (or ERS) has been reported on the affected hemisphere compared with the nonaffected hemisphere, attributed to an impairment of cortical modulation that results in decreased cortico-spinal output (Rossiter et al., 2014). Consistent with this finding, severely impaired participants in our study showed ipsilesional cortex synchronization (ERS) during movement onset, indicating inhibition, possibly because of deactivation (or “idling”) of neuronal assemblies after severe injury (Pfurtscheller, 1992; Pfurtscheller et al., 1996). Additionally, we found that movement-related oscillations in the DN paralleled those of sensorimotor cortex across alpha and low beta frequency bands, during the phasic movement (i.e., participants with ERD in the ipsilesional cortex displayed concurrent ERD in the DN and vice versa). Our task required participants to exercise greater motor control during movement onset because of the presence of a target (plateau of the block), while offset required them to completely release their grip to fall back at the baseline position. The strong interconnectivity between the cerebral cortex and cerebellum during movement planning and execution, and its preservation in the poststroke state further encourages the exploitation of the CTC pathway loop as a target pathway to promote rehabilitation.
CCC and GC during hold and relaxation
While movement onset and offset involved dynamic motor control, our block-tracking task also had an isometric grip hold phase, in which static motor control was required. The task is widely used to study cortico-muscular coherence (Mima and Hallett, 1999; Larsen et al., 2017) and was implemented here to examine CCC, reflecting the communication between cerebellum and ipsilesional cortex during motor control (Belardinelli et al., 2007). We found CCC to be significantly stronger in the hold state than in the relaxation state. This was aligned with the correlation in ERD/S power between ipsilesional cortex and the DN during movement, reflecting the engagement of this network during volitional effort maintenance, but not during relaxation. The periods of hold and relaxation were picked such that they were outside the movement onset and offset phases to make sure power fluctuations because of phasic movements were minimal and did not impact the CCC estimates.
Peak CCC and GC directed from ipsilesional cortex → DN or DN → ipsilesional cortex occurred predominantly in the low beta band. The findings are in line with those of previous studies investigating the normal state in nonhuman primates (Soteropoulos and Baker, 2006) and MEG/EEG-based human studies (Belardinelli et al., 2007; Mehrkanoon et al., 2016). There was a strong correlation between low beta band CCC and task accuracy, regardless of impairment level, suggesting that beta band mediates compensatory efforts to perform tracking as accurately as possible during task execution. CCC may serve as a surrogate metric for the volume of information exchange between the ipsilesional cortex and DN needed to maintain motor accuracy.
Cerebellar DN has been previously shown to play a major role in motor preparation and ramping activity before movement onset (Kunimatsu et al., 2018; Chabrol et al., 2019). In fact, motor cortex is driven by cerebellum during the initial preparatory phase, and the directionality reverses when motor control is achieved (Belardinelli et al., 2007). Our findings of significantly high directed communication from cortex to DN and DN to cortex during the hold phase corroborate these earlier reports. In a repetitive cyclical tracking task such as ours, the role of cerebellum could be seen as limited, given the predictability and lack of temporal irregularities (Spencer et al., 2003, 2007). While this may be true for healthy cohorts, stroke survivors participating in this study found the task difficult to perform because of their impairment, reflected by the need to tailor task duration according to impairment level, and likely found motor execution to be less predictable than normal counterparts, regardless of the cyclic nature of the task. The high correlation between DN and ipsilesional cortex activity during movement, and high CCC during the hold period may represent the intense flow of information required to compensate for impairment during execution. We speculate that stroke survivors who rehabilitated more extensively than those enrolled in our study may have a lesser role of DN in repetitive tasks. Interestingly, four of six of our moderately impaired participants also displayed CCC in the beta band during relaxation. Given the dynamic nature of the block-tracking task, it is possible that the information flow during relaxation was related to planning for the next cycle. As our clinical trial of DN DBS progresses, we will examine whether these same participants show a reduction of DN involvement during the execution of motor control. Such finding would corroborate that as participants recover, there may be less demand on the DN to support repetitive task planning and execution.
Gamma band and participant impairment
Our results showed a strong interaction between gamma band and severity of impairment. ERS observed in both ipsilesional cortex and DN in the gamma band was significantly greater at movement onset than at offset among severely impaired individuals. Additionally, gamma-band CCC during the hold period of the task correlated with impairment severity. The role of gamma oscillations in movement is still an evolving area of research (Nowak et al., 2018). While some suggest that movement-related changes in the gamma oscillations could be the result of afferent proprioceptive feedback to the motor cortex (Miller et al., 2010), others relate gamma oscillations to active motor control processes (Muthukumaraswamy, 2010). Hence, increased gamma-band synchronization during movement and increased gamma-band CCC during the hold period in severely impaired participants could be because of pathologic (and unsuccessful) compensatory mechanisms resulting from more severe stroke-related injury.
Limitations
We relied on a sample size of 10 participants, the first batch to undergo a novel clinical trial of DN DBS for poststroke rehabilitation. While 12 participants were enrolled in the clinical trial, the current study included only 10 participants who were able to be externalized to collect data from DN. The sample size was dictated by the power required for an early phase I clinical trial to assess the safety of the DN DBS procedure. The size of the sample may be compensated by the rich and novel dataset, consisting of the first ever direct electrophysiological examination of the human DN, combined with EEG, during a motor control task. We acknowledge several variabilities or biases in the dataset. Strokes are always variable in location, size, and level of injury to corticospinal pathways. We limited the variability of the sample by enrolling only stroke survivors with moderate to severe or severe residual hemiparesis. Six of 10 participants performed a 5 s block-tracking task, while others performed a 2.5 s block-tracking task. This customization was necessary to successfully engage participants with lesser or worse impairments in an appropriately challenging task. We adapted the dataset to assimilate findings in a meaningful way. Despite all the limitations, the current study is the first ever to invasively investigate the DN in humans. Hence, we believe that inference will be improved after replication with a larger and potentially more homogeneous sample.
Conclusions
Movement-related DN oscillations largely parallel sensorimotor cortex after stroke, and there is a significant correlation in ERD/S power between DN and ipsilesional cortex across alpha and low beta frequency bands, consistent with robust interconnectivity during movement planning and execution. There were effects related to impairment severity, with moderately impaired participants showing predominantly low beta band ERD and severely impaired participants showing predominantly ERS. Information flow evidenced by CCC in the low beta band was associated with better task accuracy, highlighting the significance of cortico–cerebello–cortical interactions on motor execution and control after stroke. Unlike Parkinson's disease and other movement disorders, where beta-band synchrony between basal ganglia and motor cortical areas are considered to the pathologic (McConnell et al., 2012), the task-related low beta coherence observed between cerebellar DN and ipsilesional motor areas appears to have a functional role in facilitating motor control in stroke survivors. In fact, cerebellar DN has been suggested to entrain beta oscillations throughout the central and peripheral motor systems to facilitate motor activity (Aumann and Fetz, 2004). In support, previous studies (Baker et al., 2010; Machado et al., 2013) have shown that stimulation of cerebellar DN in rodents at 20 Hz (low beta) or 30 Hz (high beta) provided sustained augmentation of excitability of ipsilesional cortex in stroke-induced rodents. The current study further corroborates the beta band as the natural frequency band of coherence between DN and ipsilesional motor regions in humans during task execution and control. These findings support strong opportunities for therapeutic targeting of the cerebello–thalamo–cortical network in the beta band to enhance the excitability of the perilesional cortex. Future stimulation paradigms will explore the possibility of fine-tuning and individualizing the stimulation frequency to each participant's natural frequency of cortico-cerebellar interaction. Further, we anticipate that better outcomes could be achieved if we exploit the cerebellar DN role in error prediction (Cooperrider et al., 2016) toward a closed loop paradigm of DBS that can be used along with motor training.
Footnotes
This project was funded through National Institutes of Health BRAIN Initiative Award 5UH3-NS-10054305. We thank Dr. Jacqueline Chen for providing the details of lead localization in the dentate and MRI images of stroke lesion extent.
A.G.M. and K.B.B. are consultants and have intellectual property licensed to Enspire DBS. Enspire DBS funded part of the clinical trial from which these data are derived but did not fund this research work. The authors declare no other competing financial interests.
- Correspondence should be addressed to Andre G. Machado at MACHADA{at}ccf.org
