Abstract
As we learn to perform a motor task with novel dynamics, the central nervous system must adapt motor commands and modify sensorimotor transformations. The objective of the current research is to identify the neural mechanisms underlying the adaptive process. It has been shown previously that an increase in muscle co-contraction is frequently associated with the initial phase of adaptation and that co-contraction is gradually reduced as performance improves. Our investigation focused on the neural substrates of muscle co-contraction during the course of motor adaptation using a resting-state fMRI approach in healthy human subjects of both genders. We analyzed the functional connectivity in resting-state networks during three phases of adaptation, corresponding to different muscle co-contraction levels and found that change in the strength of functional connectivity in one brain network was correlated with a metric of co-contraction, and in another with a metric of motor learning. We identified the cerebellum as the key component for regulating muscle co-contraction, especially its connection to the inferior parietal lobule, which was particularly prominent in early stage adaptation. A neural link between cerebellum, superior frontal gyrus and motor cortical regions was associated with reduction of co-contraction during later stages of adaptation. We also found reliable changes in the functional connectivity of a network involving primary motor cortex, superior parietal lobule and cerebellum that were specifically related to the motor learning.
SIGNIFICANCE STATEMENT It is well known that co-contracting muscles is an effective strategy for providing postural stability by modulating mechanical impedance and thereby allowing the central nervous system to compensate for unfamiliar or unexpected physical conditions until motor commands can be appropriately adapted. The present study elucidates the neural substrates underlying the ability to modulate the mechanical impedance of a limb as we learn during motor adaptation. Using resting-state fMRI analysis we demonstrate that a distributed cerebellar-parietal-frontal network functions to regulate muscle co-contraction with the cerebellum as its key component.
Introduction
The central nervous system (CNS) is able to adapt to a novel physical environment by means of two major control mechanisms. One mechanism is through learning an internal model that generates motor commands to create appropriate forces, and the other is impedance control that modulates the impedance of the limbs and joints by regulating muscle co-contraction (Franklin et al., 2003; Osu et al., 2003). In the context of adaptation to novel dynamics, muscle co-contraction provides stability against unpredictable disturbances and is used to compensate for inaccuracy in the internal model. It could also expedite the rate of motor learning by reducing kinematic error with respect to a target position or a desired trajectory while adapting to unfamiliar disturbances (Franklin et al., 2003; Heald et al., 2018). Furthermore, muscle co-contraction plays an important role in everyday activities such as learning a motor skill and manipulating a tool or a handheld object (Milner, 2002). Co-contraction generally decreases as a skill and motor performance improves (Clément and Rézette, 1985). However, it is necessary to maintain sufficient co-contraction to provide mechanical stability depending on the requirements of a given activity (Franklin et al., 2004).
Muscle co-contraction and adaptation of the temporal patterns of muscle activation during the course of motor learning have been investigated through behavioral and psychophysical studies (Milner and Cloutier, 1993; Thoroughman and Shadmehr, 1999; Franklin et al., 2003; Darainy and Ostry, 2008; Franklin et al., 2008). Franklin et al. (2003) identified 3 phases during adaptation to a novel physical environment. When the environmental dynamics change abruptly, there is an initial phase of elevated muscle co-contraction which stiffens the arm and makes it more resistant to unpredictable disturbances. As the learning progresses, the CNS gradually improves the internal model to counteract the disturbance by means of opposing forces during the second phase and reduces reliance on mechanical impedance by reducing the amount of co-contraction. Muscle co-contraction continues to decline after the internal model of the environmental dynamics is accurate, during the final phase of adaptation. It is believed that during this final phase the CNS fine tunes levels of muscle co-contraction to minimize the metabolic cost (Franklin et al., 2004).
In an earlier functional brain imaging study, we found that the cerebellum and inferior parietal cortex were differentially activated in relation to co-contraction of hand and wrist muscles, which was regulated with the complexity of the dynamics of a task (Milner et al., 2006). We, therefore, hypothesized that the functional connectivity between these two regions would be modulated in parallel with the increase and subsequent decrease of contraction during a force field adaptation task. In contrast, we did not expect to find that the functional connectivity between the cerebellum and primary motor cortex would be similarly modulated since we had previously shown that the connectivity between these two regions paralleled the formation of an accurate internal model of the task dynamics which evolves with a different time course than the modulation of generalized muscle co-contraction (Franklin et al., 2003; Vahdat et al., 2011).
We recorded arm muscle activity to quantify co-contraction and used resting-state fMRI during distinct phases of the adaptation to investigate which changes in the resting-state functional connectivity of the sensorimotor network paralleled the regulation of co-contraction (Albert et al., 2009; Vahdat et al., 2011). A significant advantage of this approach over activity-related task based fMRI is that changes in brain activity related to the motor adaptation are not confounded by intense movement (motor execution and sensory feedback) related activity. This approach circumvents masking of salient features related to motor learning, i.e. the persistent changes in connectivity associated with the durable features of motor learning.
Materials and Methods
Participants performed reaching movements in a novel velocity-dependent force field created by a robotic interface. Kinematics and EMG were recorded during four different stages of motor adaptation followed by a resting-state fMRI scan that monitored brain activity immediately after each stage was completed.
Participants
Seventeen right-handed and neurologically normal subjects (10 M, 7 F, age: 25.6 ± 3.8) took part in this study. The study was approved by the McGill University Research Ethics Board.
Experimental setup
Subjects moved a joystick robot (Lai et al., 2003) between two target positions (Fig. 1A). The robot moves along a spherical surface of 41 cm radius and can exert forces in any direction tangent to the sphere. It was programmed to create torques that produced tangential force at the tip of the joystick orthogonal to the direction of motion in a manner that depended on the angular velocity of the robot motors (velocity-dependent force field) according to the following equation:
A, Schematic of the joystick robot. B, A subject with EMG electrodes on ready to perform a reaching movement. C, Subject at the end of a reaching movement trial. Position of the joystick is displayed on the monitor by a square. D, General experimental procedure shows the number of trials in each session which follows by a resting-state scan. NF, null field; FF, force field.
Optical encoders (U.S. Digital, 50000 counts/rev) on the motor shafts were used to provide accurate information about the joystick position. Subjects were seated in front of the robot in a comfortable position with a right angle between forearm and upper arm at the start of the reaching movement (Fig. 1B). In the experimental sessions, subjects moved the robot's handle in the sagittal plane between two target positions (shown on the front screen) located 30 cm and 55 cm anterior to the shoulder along a line perpendicular to the line joining the shoulders, i.e. outward from the shoulder (Fig. 1C). The angular position of the joystick was displayed as a 7 mm square cursor on a monitor placed approximately 1 m in front of the subject. The angular coordinates represented on the display were oriented so that forward motion was represented as vertical motion on the monitor. The targets were displayed as 12.5 mm × 12.5 mm squares. Subjects were instructed to keep the peak hand velocity within the range of 75 ± 5 cm/s which resulted in movement times of approximately 600 ms. They were provided with visual feedback of peak hand velocity range at the end of each trial. If the velocity was too low the target became green, if the velocity was too high it became red and if it was within the desired window it became the color of the cursor. Each trial consisted of an active voluntary movement to the outward target followed by a passive movement (under servo control) of the relaxed arm back to the start position. Prior to starting a trial, subjects were required to move the cursor into the start target window and remain there for 750 ms. Surface EMG from 7 arm and shoulder muscles was acquired using the Delsys Bagnoli system. EMG from antagonistic pairs of elbow muscles (brachioradialis and triceps lateral head), shoulder muscles (anterior deltoid and posterior deltoid, pectoralis clavicular head and posterior deltoid) and biarticular muscles (biceps short head and triceps long head) as well as hand position and handle force were acquired at 2000 Hz. Behavioral experiments took place in the Neuromuscular Control Lab at McGill University after which MRI scans were conducted at the Montreal Neurological Institute (MNI) located approximately 200 m from the lab.
Experimental procedure
Participants took part in four experimental sessions on four different days. In each session, subjects performed a specified number of reaching movement trials between the two target positions. Prior to beginning the movement trials, they performed trials in which a fixed-force was applied in four different directions: 0, 90, 180 and 270 degrees, i.e. rightward, forward, leftward and backward with the arm in the start position. Subjects held the cursor in the start target window while resisting an applied force of 7.5 N. The EMG recorded during these trials was used for EMG normalization. Two trials were performed in each direction. After completing the movement trials, subjects immediately proceeded to the MNI for fMRI scans which were acquired approximately 30 minutes after movement trials had been completed. This time was required to remove the EMG electrodes, walk to the MRI suite, position the subject in the MRI scanner and run preliminary alignment scans.
In the first experimental session, participants performed 200 trials in the null field followed by resting-state fMRI scan 1. The first session familiarized subjects with the experimental procedure and enabled us to compute a muscle co-contraction index after training in the null field. Furthermore, scan 1 aimed to establish a baseline that effectively controlled for changes in functional connectivity associated with the novelty of moving the robot without an associated learning component (without force field adaptation). In the second session, participants performed 6 trials in the force field, where the final trial was a catch trial. During catch trials, the force field was not turned on. Catch trials were used as a way to measure and quantify motor learning. After the catch trial, they underwent resting-state scan 2. This scan was used to examine changes in functional connectivity associated with the first phase of motor adaptation in which muscle co-contraction is substantially increased to reduce performance error (Franklin et al., 2003; Milner and Franklin, 2005). In the third session, subjects performed 50 trials in the force field of which 3 of the last 10 were randomly selected as catch trials. They then underwent resting-state scan 3 which was used to examine changes in functional connectivity associated principally with reduction in co-contraction and improvement in internal model accuracy related to the second phase of adaptation. At this point in learning, it has been found that the internal model is relatively accurate but that co-contraction levels are still declining (Franklin et al., 2003; Franklin et al., 2004). Finally, in the fourth session, subjects performed 200 trials in the force field of which 3 of the last 10 were again randomly selected as catch trials. They underwent resting-state scan 4 which was predominantly associated with additional reduction in co-contraction during the final phase of adaptation. The specific number of trials in each of the three force field sessions was designed to mimic the three distinct phases of dynamic motor adaptation identified from the mean time course of changes in co-contraction of antagonist muscles and the time course of changes in net joint torque in previous studies of force field adaptation (Franklin et al., 2003; Franklin et al., 2008).
The number of days between sessions was selected to isolate three distinct phases of co-contraction modulation and motor learning. In the first session subjects acquired the basic skill to perform the task with only an inertial load (null field). Since the joystick controller compensated for its weight, the remaining dynamics comprised its moment of inertia, which was essentially constant for the straight line trajectory between targets. Given that humans have extensive experience moving objects with constant moments of inertia, subjects immediately acquired the ability to perform the movement and quickly learned to adjust their speed as instructed. Thus, the transformation between arm movement and joystick movement, i.e. the basic representation of the joystick as a tool, was established in the first session. The second experimental session, which consisted of only 6 trials in the force field, was conducted 3.53±1.77 days (mean and standard deviation) after the first session. We performed the third session 6.47±2.53 days after completion of the second session to ensure that there would be little retention of any learning that might have occurred. The fourth session was generally conducted the day after the third session (1.12±0.33 day) to ensure retention of learning and immediately continue with the final stage of adaptation. Figure 1D depicts the experimental protocol.
Experimental design and statistical analysis
The study consisted of 4 sessions on 4 different days and followed a within-subject design. The first session comprised a baseline condition (null field trials) where neither muscle co-contraction nor force field learning was required to execute the task. This served as an internal control whereby changes in functional connectivity in subsequent sessions were determined with respect to the resting-state scan immediately following the first session. All fMRI data were analyzed using FSL software packages (FMRIB Software Library v6.0). Statistical analysis at both subject-level and group-level was performed using a GLM (General Linear Model) approach. Cluster significance threshold was set to p = 0.05 (corrected for multiple comparisons using Gaussian random field theory and corrected for multiple ROIs). We controlled for non-specific changes in functional connectivity by incorporating an index of muscle co-contraction or motor learning as a regressor in the statistical analysis of functional connectivity to ensure that only those connections whose change in functional connectivity was significantly correlated with the co-contraction or motor learning index were retained. Furthermore, linear regression was performed between changes in the co-contraction or motor learning index and changes in the strength of functional connectivity which demonstrated that the degree of co-contraction or motor learning predicted the change in functional connectivity. We also performed paired-sample t-tests to investigate whether changes in the muscle co-contraction index and motor learning index were statistically significant between pairs of sessions with p = 0.05, corrected for multiple comparisons. In our previous study with 13 neurologically healthy subjects (Vahdat et al., 2011), a difference of at least 90% of baseline functional connectivity was detected by the effect of motor learning for the hypothesized links. The standard deviation observed was 60% of baseline functional connectivity. Fourteen subjects would, therefore, provide 90% power at a significance level of 99% (α = 0.01) to detect a change in connectivity from pre- to post-learning equal to baseline connectivity. We expected that the effects of co-contraction on change in functional connectivity would be similar, although we increased our sample size to 17 subjects to account for the possibility of greater variability.
EMG and co-contraction analysis
EMG from 7 muscles, handle force and hand position were recorded from 250 ms before until 1000 ms after the onset of movement and sampled at 2000 Hz. EMG was processed and a muscle co-contraction index for each session was obtained by the following steps (Thoroughman and Shadmehr, 1999; Gribble et al., 2003): (1) raw EMG signals were demeaned and band-pass filtered between 20 and 450 Hz using a 3rd order Butterworth filter with zero-phase lag; (2) root-mean-square (rms) EMG in a 50 ms moving-average window was computed for the entire reaching movement; (3) the maximum rms EMG for each muscle obtained during the 8 fixed-force contractions (2 trials in each force direction) was used to normalize the rms EMG; (4) the minimum normalized rms EMG of the two muscles of an antagonistic pair in each 50 ms window was averaged from 150 ms prior to movement onset a 400 ms after movement onset to obtain the co-contraction index (CI).
The four antagonistic muscle pairs consisted of biceps short head/triceps long head, brachioradialis/triceps lateral head, anterior deltoid/posterior deltoid and pectoralis clavicular/posterior deltoid. This interval for calculating the CI was chosen to exclude EMG related to error correction near the target position.
In order to account for factors related to removal and replacement of the electrodes on different days, for which normalization may not have been adequate, we compared the CI at the end of one session with the CI at the beginning of the next session. For sessions 2 and 3, if the mean CI of trials 3, 4 and 5 at the end of session 2 differed by more than 25% from the mean CI of trials 3, 4 and 5 at the beginning of session 3, the CI for session 3 trials was scaled by the ratio of the means.
A similar procedure was used to adjust the CI of session 4 by scaling the CI of session 4 trials by the ratio of the mean CI of the three consecutive trials immediately preceding the first catch trial at the end of session 3 and the mean CI of trials 3, 4 and 5 at the beginning of session 4. Scaling was applied to the CI for 5 of the 17 participants where the difference in CI between the end of one session and the beginning of the next was greater than 25%.
Co-contraction index
Subsequent analysis was carried out on the mean CI of the final 5 trials in session 1, mean CI of trials 2 through 5 in session 2 and mean CI of the 5 consecutive trials immediately preceding the first catch trial in sessions 3 and 4 (to exclude any possible impact of catch trials on the amount of co-contraction). We compared CIs for all 4 antagonistic muscle pairs and all participants. Muscle pairs brachioradialis/triceps lateral head and anterior deltoid/posterior deltoid had the highest CIs and most closely followed the expected evolution of co-contraction (Bhushan and Shadmehr, 1999; Franklin et al., 2003; Osu et al., 2003; Franklin et al., 2008) over the four sessions for most of the participants (15 out of 17). Therefore, the CI for each participant was defined as the average CI of those two muscle pairs in later functional connectivity analysis. Figure 2A shows mean CI and standard error across all participants for the 4 sessions.
fMRI acquisition
All neuroimaging data were acquired using a 3.0 T Siemens Magnetom Trio MR scanner at the McConnell Brain Imaging Centre of the Montreal Neurological Institute (MNI). Subjects rested in a supine position in the scanner with their head held comfortably in place and restrained by foam pads to minimize head movement. Earplugs were used to reduce scanner noise. Whole-brain functional data were acquired using a regular T2*-weighted EPI sequence (advanced phase-corrected; 32 head-coil channels; axial slices oriented to the AC-PC line; spatial resolution 3×3×3 mm3, isotropic; slice thickness 3 mm, 42 slices; TE = 30 ms; TR = 2.54 s). In each session, two 8-minute resting-state fMRI scans were acquired after short preliminary scans (auto-alignment and sagittal Trufisp). Subjects were instructed to fixate on a cross displayed on a monitor visible through an angled mirror and relax during resting-state scans. A T1-weighted anatomical scan (spatial resolution 1×1×1 mm3, isotropic; slice thickness 1 mm; TE = 2.98 ms; TR = 2.30 s) was acquired at the completion of the resting-state scans. The first scanning session (scan 1) ended with a block-design task consisting of two 5-minute fMRI scans during which subjects performed rhythmic arm reaching movements for 30 s alternating with 30 s of rest. A green circle flashing at 1/3 Hz was used as a cue for movement timing and a red flashing circle was used to signal the resting condition. The subjects practiced the movements prior to the start of the scan to ensure consistency. The movements mimicked those performed during training with the robot, though at a lower speed to minimize movement artifact in the scanner. The task-based fMRI scans were used to identify ROI coordinates in selected brain regions for later functional connectivity analysis.
fMRI data analysis
Image processing was performed using FSL software packages (FMRIB Software Library v6.0), Oxford, UK (Smith et al., 2004; Jenkinson et al., 2012). We used a GLM-based approach to pre-process the resting-state fMRI data similar to Vahdat et al. (2011) but with some modifications. Pre-processing began with the following steps using the first-level FEAT toolbox (Woolrich et al., 2001): (1) two volumes in each scan were removed (to ensure equilibrium magnetization had been reached); (2) non-brain removal using BET (Brain Extraction Tool); (3) slice timing correction; (4) motion correction using the 6 DOF rigid body transformation implemented in FLIRT (FMRIB Linear Image Registration Tool); (5) spatial smoothing (Gaussian kernel of FWHM, 6 mm); (6) temporal high-pass filtering (Gaussian-weighted least-squares straight-line fitting with σ = 100 s, equivalent to a cut-off frequency of 0.01 Hz); (7) global intensity normalization.
Brain extraction for T1-weighted anatomical images was carried out using the optiBET algorithm (Lutkenhoff et al., 2014). Spatial transformations between functional T2*-weighted data and MNI standard space (MNI 152 2 mm brain) were obtained by two separate transformations: Boundary-Based Registration (BBR) (Greve and Fischl, 2009) from functional to anatomical space and nonlinear registration from anatomical to standard space. In order to model and remove physiological artifacts (cardiac and respiration), average BOLD signals in white matter and CSF were calculated and used as confound regressors (Shehzad et al., 2009). For this purpose, we segmented each subject's brain extracted T1-weighted image by FAST (FMRIB's Automated Segmentation Tool) and thresholded the resulting WM and CSF masks to ensure 90% tissue type probability (Zhang et al., 2001). Each thresholded mask was then transferred to the functional space using the transformations already obtained and finally the average time series from all voxels within the mask was calculated. We included confound regressors consisting of motion parameters (3 for translations, 3 for rotations), WM and CSF in the GLM for each subject. The residual signal was further band-pass filtered in the range of 0.01 - 0.10 Hz since the neural activity-related signal in resting-state fMRI is expected to lie within this frequency range (Fox and Raichle, 2007).
ROI identification
Task-based fMRI scans during rhythmic arm reaching movements acquired in the first experimental sessions were analyzed to identify and determine coordinates of seeds (ROIs) for later seed-based analysis. Pre-processing of task scans was performed using the same steps for resting-state data (spatial smoothing with 5 mm kernel, temporal high-pass filter with σ = 85 s), except for removal of physiological artifacts. A regressor for arm movement versus rest condition was modeled with a boxcar function with the temporal derivative of the stimulation timing as an additional regressor. The regressors were convolved with a gamma hemodynamic response function. After subject-level GLM analysis in FEAT, the Z score images were fed to the higher-level FEAT for group-level analysis (Woolrich et al., 2004). A mixed-effects model in FEAT with stringent cluster thresholding (Z > 3, cluster significance threshold of p = 0.01, corrected) was used for the group-level analysis and resulting activation maps in the MNI standard space were used to identify seed coordinates.
We selected ROIs in the areas of the brain within the sensorimotor network based on the prior information about activation and connectivity from human and non-human primate studies of motor learning. These areas include ipsilateral cerebellum, contralateral primary motor cortex (M1), ventral premotor cortex (PMv), dorsal premotor cortex (PMd), supplementary motor area (SMA), primary somatosensory cortex (SI) and posterior parietal cortex (PPC). Each ROI was defined as a sphere with a 6 mm radius centered on the peak activity from the group Z-map task-based scans (for each ROI). Spatial coordinates of the center of each ROI in MNI standard space, Z value of the peak activity and corresponding anatomical labels are listed in Table 1. Two additional ROIs in the secondary somatosensory cortex and basal ganglia were added as those two regions are part of the sensorimotor network (Vahdat et al., 2011), although we did not find activity in these regions in the maps derived from task-based analysis. We used Harvard-Oxford cortical and subcortical (Desikan et al., 2006), Juelich histological (Eickhoff et al., 2007) and probabilistic cerebellar atlases (Diedrichsen et al., 2009) to identify anatomical labels from activation maps.
Selected ROIs from a task fMRI scan during reaching movements inside the scanner
Functional connectivity analysis associated with co-contraction index
We aimed to determine for which neural connections, functional connectivity (FC) evolved in parallel to the pattern of co-contraction during adaptation by determining which changes in functional connectivity were correlated with the CI across different sessions. First, we calculated the average time series of all voxels in each of the seed ROIs from the pre-processed resting-state data and assessed the correlation between this regressor and all other voxels in the brain in a subject-level GLM. We also included the temporal derivative of each ROI's regressor to take into account possible variations in the hemodynamic response function (HRF) delay as the HRF might vary between sessions, subjects and different areas of the brain.
For each subject, the FC maps (coefficients estimates and their variances) were input to an intermediate-level GLM (Woolrich et al., 2004) to identify networks whose FC paralleled the evolution of co-contraction. The GLM model for each subject consisted of a regressor modeling CI values for each of the four sessions and a regressor to model the overall mean across sessions and scans. The CI regressor was orthogonalized with respect to the mean regressor. Finally, a group-level GLM using a mixed-effects model (Z > 2.3, cluster significance threshold p = 0.05, corrected for comparisons using Gaussian random field theory) was performed across subjects to generate thresholded Z score FC maps corresponding to each ROI.
We also examined changes in the strength of FC in relation to CI between consecutive sessions in order to investigate whether distinct neural pathways are involved in modulating muscle co-contraction at certain stages of adaptation in addition to those involved over the entire course of adaptation, e.g. the difference in FC between session 2 and session 3 which is associated with a substantial reduction in muscle co-contraction. For this analysis, a GLM model consisting of a mean regressor and a regressor to model the contrast between the two sessions (-1 and 1) was used after subject-level FC analysis, then group-level analysis was performed using a mixed-effects model (Z > 2.3, cluster significance threshold p = 0.05, corrected). The difference in CI values between the two sessions was included in the group-level GLM as an additional regressor, as described elsewhere (Vahdat et al., 2011; Vahdat et al., 2014). For all FC analysis, multiple comparison correction (multiple ROIs) was performed by considering only clusters with p values less than 0.05/11 = 0.0045 (11 is the number of ROIs) as statistically significant. In addition, we performed linear regression between change in FC (ΔFC) and change in CI (ΔCI) and calculated mean FC values in each experimental session.
Changes in FC during motor learning could be related to more than one process or mechanism. Vahdat et al. (2011) developed a technique that identified networks related to motor aspects of learning and distinguished them from those that were more associated with sensory and perceptual changes. In order to identify brain networks whose functional connectivity is modulated specifically by muscle co-contraction and not by the formation and improvement in internal model accuracy, a motor learning index (MI) was defined as the mean of the maximum perpendicular deviation between the hand trajectory and the straight line connecting the center of the movement start and end targets for all catch trials in a session. A regressor for MI was included in this group-level analysis and CI was orthogonalized with respect to it in order to separate the shared variance attributable to both CI and MI. The mean MI across all subjects is shown in Figure 2B. Finally, to identify specific brain networks involved in motor learning, ΔFC between session 1 (pre-learning) and session 4 (post-learning) was analyzed in relation to the MI.
Results
Evolution of co-contraction index and motor learning index
As expected from previous studies (Franklin et al., 2003; Osu et al., 2003) we found that the CI and MI evolve with different profiles during adaptation to a novel physical environment (Fig. 2). CI increased substantially from session 1 to 2 and then decreased during sessions 3 and 4 whereas MI increased monotonically from one session to the next. We performed paired-sample t-tests (corrected for multiple sessions) to investigate whether CI and MI changed significantly between successive experimental sessions. CI increased significantly from session 1 to session 2 (t = 6.47, p = 2.31×10−5), but was reduced significantly between session 2 and session 3 (t = 5.19, p = 2.67×10−4) and between session 3 and session 4 (t = 4.60, p = 8.79×10−4). MI significantly increased between session 2 and 3 (t = 13.33, p = 8.80×10−10), but did not change significantly between session 3 and 4 (t = 2.15, p = 0.096).
A, Mean CI (filled circles) ± standard error (vertical lines) over all participants for the 4 experimental sessions. CI, co-contraction index. B, Mean MI ± standard error across all subjects. MI, motor learning index.
Figure 3 depicts representative hand paths at various stages of adaptation, including a catch trial near the end of the final training session (session 4).
Hand paths at different stages of force field adaptation for a single participant. Start and end targets are shown as red filled circles. The last trial in the null field (NF S1), the first and the last force field trials in session 2 (FF S2), the last force field trials in sessions 3 and 4 (FF S3, FF S4) and the first catch trial in session 4 (CT S4) are shown. NF, null field; FF, force field; CT, catch trial.
Change in functional connectivity associated with muscle co-contraction throughout the adaptation
We analyzed FC in relation to the CI across all 4 experimental sessions (Fig. 4). This analysis aimed to identify the neural pathways whose FC reliably evolves in parallel to the pattern of muscle co-contraction during the entire adaptation process. Positive FC maps (warm color-coded) followed a pattern which was positively correlated with CI whereas in negative FC maps (cold color-coded), the correlation with CI was negative. Top row shows a neural link between cerebellum (lobule VΙΙΙ) and clusters in inferior parietal lobule (IPL) whose FC was strongly and positively correlated with the CI during the course of adaptation. The mean FC increased in the early phase of adaptation from a value close to zero in the null field, then decreased as adaptation progressed. Functional connectivity between cerebellum (lobule V) and relatively small clusters in the hippocampus was also positively correlated with the CI. For this link, mean FC did not change between sessions 2 and 3 but decreased slightly in the final session. Anatomical location and details of the connectivity are provided in Table 2.
Results summary for the functional connectivity analysis in relation to the CI during the entire adaptation process
Functional connectivity maps in conjunction with co-contraction index throughout the entire motor adaptation. Each row is related to a specific ROI whose FC with some areas in brain significantly changed in relation to the CI. Left column is the location of the ROI displayed in purple dot. Middle panel represents clusters whose FC with the ROI reliably changed parallel to the pattern of the CI. Anatomical labels of the clusters and their displayed coordinates in the MNI space were also mentioned. Mean values of FC averaged across all subjects in the corresponding cluster for each experimental session were shown in the right column. Positive maps were depicted in red to yellow and negative maps were depicted in dark to light blue color-coded maps. IPL, inferior parietal lobule; HP, hippocampus; R, right hemisphere; L, left hemisphere; S1 to S4, sessions 1 to 4.
Change in functional connectivity associated with reduction in muscle co-contraction
Figure 5 shows the links for which ΔFC between session 2 and session 3 (S2 – S3) was significantly correlated with ΔCI on a per subject basis. As seen in the first row, connectivity between cerebellum (lobule VΙ) and clusters located primarily in superior frontal gyrus and partly in middle frontal gyrus decreased in conjunction with CI from session 2 to 3. Subjects who showed greater reduction in co-contraction showed greater decrease in FC (r = 0.59, p = 0.01). On the other hand, we found a map where ΔFC was negatively correlated with the change in CI. There was a negative correlation between ΔFC of primary motor cortex (M1) with clusters in premotor cortex, SMA and SI and ΔCI. In this case, the magnitude of FC was initially positive and became more positive as the CI decreased. Details are listed in Table 3.
Results summary for the functional connectivity analysis in relation to the CI between sessions 2 and 3
Change in functional connectivity in conjunction with co-contraction index between session 2 and session 3. Each row is related to a specific ROI whose FC with some brain areas significantly changed in relation to the CI between these two sessions. In this figure, the positive FC map indicates a decrease in the connectivity from session 2 to 3 and the negative FC map indicates an increase. Left column is the location of the ROI displayed as a purple dot. Two middle panels represent clusters whose FC with the ROI reliably changed in relation to the CI and mean FC values averaged across all subjects in the corresponding cluster for the experimental session. Right column shows linear regression and correlation between change in the FC (ΔFC) and change in the amounts of CI on a per-subject basis. p is the significance level and r is the Pearson correlation coefficient. Positive maps are depicted in red to yellow and negative maps are depicted in dark to light blue color-coded maps. CB, cerebellum; SFG, superior frontal gyrus; MFG, middle frontal gyrus; M1, primary motor cortex; SI, primary sensorymotor cortex; SMA, supplementary motor area; PMC, premotor cortex; R, right hemisphere; L, left hemisphere.
Change in functional connectivity associated with the final stage of co-contraction modulation
ΔFC of cerebellum (lobule V) with motor cortical regions consisting of M1, SMA and premotor areas was positively correlated with the reduction in CI between sessions 3 and 4 (S3 – S4; Fig. 6), i.e. FC became increasingly negative as CI declined (r = 0.64, p = 0.005). Details are listed in Table 4.
Results summary for the functional connectivity analysis in relation to the CI between sessions 3 and 4
Change in functional connectivity in conjunction with co-contraction index between session 3 and session 4. Layout is similar to Figure 4. CB, cerebellum; M1, primary motor cortex; SMA, supplemenary mtor area; PMC, premotor cortex
Functional connectivity analysis in relation to the motor learning
Besides the analysis of functional connectivity associated with co-contraction, we analyzed change in FC in relation to the amount motor learning (MI) between session 1 and session 4 for comparison with the earlier study of Vahdat et al. (2011). Changes in the MI and CI between sessions 1 and 4 were correlated (r = -0.67, p = 0.003), thus functional connectivity analysis in relation to the motor learning included ΔCI as a regressor to dissociate possible co-contraction modulation effects.
Change in FC between primary motor cortex and SPL was positively correlated with the MI as shown in Figure 7 (top row). For this link, ΔFC was greater for the participants with higher MI. Change in connectivity in the link between posterior parietal cortex and SII was also correlated with the MI. The link between cerebellum (lobule V) and SPL showed a negative correlation between ΔFC and MI. The mean FC for this link was initially not significantly different from zero but became negative following adaptation where the magnitude of the negative change was greater for subjects with higher MI. Similar to the previous tables, Table 5 lists details of the correlated clusters within the network.
Results summary for the functional connectivity analysis in relation to the MI
Change in functional connectivity in conjunction with motor learning index between session 1 and session 4. Layout is similar to Figure 4. M1, primary motor cortex; SPL, superior parietal lobule; SII, secondary somatosensory cortex; PPC, posterior parietal cortex; CB, cerebellum.
Discussion
The primary objective of this study was to investigate the neural correlates of muscle co-contraction and brain networks that are involved in modulating co-contraction during dynamic motor adaptation. We analyzed change in the functional connectivity in relation to ΔCI and MI and eliminated the portion of total variance corresponding to both ΔCI and MI such that the residual variance was correlated with only one metric. This procedure was incorporated into the GLM analysis and served as one control for non-specific changes in functional connectivity. More importantly, linear regression between neural measures (ΔFC) and behavioral measures (ΔCI or MI) ensured that only those connections for which there was a statistically significant correlation between the co-contraction index or motor learning index and changes in functional connectivity across participants, were retained. Several studies have reported that resting-state functional connectivity, on the timescale of minutes, is highly reliable on repeated measures from different days (Shehzad et al., 2009; Laumann et al., 2015; Hindriks et al., 2016). Thus, the study design ensured that the identified changes in brain network connectivity were not affected by extraneous factors.
Consistent with our hypotheses, the cerebellum appears to play the predominant role in regulating co-contraction, as it formed the key component of the networks which we found to be correlated with ΔCI. This finding is consistent with results of previous studies involving cerebellar single unit recording (Frysinger et al., 1984) and task related fMRI (Milner et al., 2007). The cerebellum is ideally structured to register errors in motor commands (Imamizu et al., 2000; Ito, 2013; Shadmehr, 2017) and, therefore, to implement countermeasures such as co-contraction to counteract external disturbances. As we predicted, based on our previous study (Milner et al., 2007), FC between cerebellum (lobule VIII) and IPL was correlated with the CI. One of the functions of IPL appears to be the integration of multisensory information, such as vision and proprioception, in the context of spatial attention and guidance of hand movements (Lynch, 1980; Andersen, 1989; Clower et al., 2001; Fogassi and Luppino, 2005; Caspers et al., 2013). Clower et al. (2001) identified a link between the cerebellum and IPL and suggested this projection may play a role in sensory recalibration between vision and proprioception as occurs during adaptation. The need for such adjustments parallels co-contraction levels. Initially, the force field produces large spatial errors which require modifications to both the amplitude and timing of commands to muscles, as well as recalibration of the relation between the visual and proprioceptive representations of the arm in space. Both the cerebellum and IPL can contribute to the incremental adjustment in motor commands during this phase of adaptation, resulting in high functional connectivity between the two regions. Later, as appropriate motor commands are learned, errors are reduced, co-contraction levels gradually decline and the need for strong synchronization between the cerebellum and IPL can be reduced. FC between the hippocampus and the cerebellum also varied in a similar manner to CI. Mean FC between these two regions increased immediately following introduction of the force field, then remained high, decreasing only slightly by the end of training. Activity in the hippocampus during adaptation to such novel dynamics has been shown to be correlated with the predicted amplitude of the perturbing force (Scheidt et al., 2012), likely to keep recent motor commands and the associated errors in memory as they are required for feedback error learning (Kawato et al., 1987). This is consistent with the role of hippocampus in spatial working memory (van Asselen et al., 2006). It has been suggested that the cerebellum plays a critical role in the formation of internal models through feedback error learning (Imamizu et al., 2000) to predict and control the physics of movement (Shadmehr, 2017). We have proposed an augmented model of feedback error learning which incorporates an element for regulating co-contraction (Franklin et al., 2008). If the role of the hippocampus in feedback error learning is to keep recent motor commands in working memory, then our augmented model would predict that it must play a similar role with respect to recent co-contraction levels.
FC between cerebellum and areas predominantly in superior frontal gyrus and marginally in middle frontal gyrus decreased from session 2 to 3 in conjunction with reduction in CI. The change in FC between superior frontal gyrus, which is a part of prefrontal cortex, and the cerebellum suggests that cognitive functions are engaged during the phase of adaptation dominated by feedback error learning (Anguera et al., 2010), but decline as the joint torques are learned. The disappearance of a significant correlation between ΔFC and ΔCI between cognitive regions during the final phase of adaptation suggests that there is no longer a role for these regions once feedforward commands have been perfected. FC between M1 and SMA, premotor cortex and SI increased from session 2 to session 3 which may suggest that the strengthened connectivity within the areas involved in motor execution is associated with learning an internal model of the task dynamics. FC between cerebellum and motor cortex (mainly M1) increased, albeit as negative FC, as co-contraction was gradually reduced from session 3 to 4, suggesting that the CNS fine tunes levels of muscle activation to reduce co-contraction (Franklin et al., 2004). Negative FC implies that the activity in the seed and target regions is reciprocally modulated, implying that if co-contraction is not required, activity in the cerebellar cortex is reduced when motor cortical regions are active. Reduction in activity of the cerebellar cortex is likely associated with increased activity of the dentate nucleus as the result of reduced inhibition from the cerebellar cortex. Projections from the dentate nucleus to motor cortical regions might play a greater role in feedforward (internal model) control in the absence of co-contraction. The increase in negative FC between the cerebellum and the network consisting of M1, premotor cortex and SMA may reflect an increase in feedback gain of this network as feedforward commands are perfected. The reasoning is that the margin for error in motor commands would be reduced as co-contraction is reduced and since the cerebellum serves a prominent role in error detection, its ability to influence the activity of the sensorimotor network would need to be strengthened.
We separately analyzed ΔFC in relation to the MI after learning. Inclusion of ΔCI as a regressor allowed us to account for any effect of change in co-contraction on ΔFC in the motor learning network. We found that ΔFC between M1 and SPL was positively correlated with MI after learning. There is evidence that activity in posterior parietal cortex (PPC) increases during motor learning (Shadmehr and Holcomb, 1997) and TMS of PPC impairs motor learning (Della-Maggiore et al., 2004). Furthermore, M1 seems to rely on inputs from parietal areas rather than feedback from the cerebellum as feedforward commands are learned. A negative correlation between ΔFC and MI for the network connection between the cerebellum and right SPL is consistent with the finding by Vahdat et al. (2011) where ΔFC between cerebellum and SPL was bilaterally correlated with both motor and perceptual learning. Negative connectivity between these two areas following learning may imply different functions according to task and process demands. The cerebellum plays a crucial role in error correction using sensory feedback, in particular from proprioceptors, providing information about the position and displacement of the hand. A postulated primary function of PPC, on the other hand, is in state estimation (Medendorp and Heed, 2019) which is important for feedforward control. The increase in FC between PPC and SII after learning is consistent with the function of state estimation, given that state estimation requires somatosensory inputs related to proprioception and force which could be provided by SII.
In summary, we identified a network comprising cerebellum, inferior parietal lobule and hippocampus whose pattern of change in functional connectivity suggests a role in adjusting limb mechanical impedance to compensate for kinematic errors as motor commands are modified to counteract external disturbances. A network consisting of cerebellum, superior frontal gyrus and motor cortical regions appeared to be specifically engaged in reducing co-contraction as the need for error compensation diminished. The link between cerebellum and superior frontal gyrus contributed to the phase of adaptation where motor commands were being perfected to adjust the net joint torques which likely required cognitive processing, whereas the strengthening of the network connection involving cerebellum and motor cortex more likely reflected a need for sensitivity to small errors as the margin for motor command error decreased in parallel with co-contraction. We identified a network involving cerebellum, SPL, M1 and SII for which change in functional connectivity was correlated with motor learning, although our protocol did not allow us to dissociate changes due to motor learning from those related to perceptual learning (Vahdat et al., 2011).
Footnotes
This work was funded by the Canadian Institutes of Health Research, Grant # 174281.
- Correspondence should be addressed to Theodore E. Milner at theodore.milner{at}mcgill.ca