Abstract
Dynamic adaptation is an error-driven process of adjusting planned motor actions to changes in task dynamics (Shadmehr, 2017). Adapted motor plans are consolidated into memories that contribute to better performance on re-exposure. Consolidation begins within 15 min following training (Criscimagna-Hemminger and Shadmehr, 2008), and can be measured via changes in resting state functional connectivity (rsFC). For dynamic adaptation, rsFC has not been quantified on this timescale, nor has its relationship to adaptative behavior been established. We used a functional magnetic resonance imaging (fMRI)-compatible robot, the MR-SoftWrist (Erwin et al., 2017), to quantify rsFC specific to dynamic adaptation of wrist movements and subsequent memory formation in a mixed-sex cohort of human participants. We acquired fMRI during a motor execution and a dynamic adaptation task to localize brain networks of interest, and quantified rsFC within these networks in three 10-min windows occurring immediately before and after each task. The next day, we assessed behavioral retention. We used a mixed model of rsFC measured in each time window to identify changes in rsFC with task performance, and linear regression to identify the relationship between rsFC and behavior. Following the dynamic adaptation task, rsFC increased within the cortico-cerebellar network and decreased interhemispherically within the cortical sensorimotor network. Increases within the cortico-cerebellar network were specific to dynamic adaptation, as they were associated with behavioral measures of adaptation and retention, indicating that this network has a functional role in consolidation. Instead, decreases in rsFC within the cortical sensorimotor network were associated with motor control processes independent from adaptation and retention.
SIGNIFICANCE STATEMENT Motor memory consolidation processes have been studied via functional magnetic resonance imaging (fMRI) by analyzing changes in resting state functional connectivity (rsFC) occurring more than 30 min after adaptation. However, it is unknown whether consolidation processes are detectable immediately (<15 min) following dynamic adaptation. We used an fMRI-compatible wrist robot to localize brain regions involved in dynamic adaptation in the cortico-thalamic-cerebellar (CTC) and cortical sensorimotor networks and quantified changes in rsFC within each network immediately after adaptation. Different patterns of change in rsFC were observed compared with studies conducted at longer latencies. Increases in rsFC in the cortico-cerebellar network were specific to adaptation and retention, while interhemispheric decreases in the cortical sensorimotor network were associated with alternate motor control processes but not with memory formation.
Introduction
Learning to control our movements in novel dynamic environments relies, in part, on motor control processes that use error feedback to adjust our motor plans from movement to movement (Shadmehr, 2017). Errors between predicted and experienced sensory feedback are used to recalibrate our “internal model” of expected task dynamics, that is used to generate motor commands for the next movement. Following exposure to a given condition, these adapted motor plans are consolidated into motor memories that contribute to improved performance (behavioral “retention”) on re-exposure to the same condition. Behavioral studies suggest that consolidation processes begin within 15 min following task execution, after which point initially fragile motor memories become more stable and resistant to interference (Miall et al., 2004; Krakauer et al., 2005; Criscimagna-Hemminger and Shadmehr, 2008). Currently, the neural correlates of these early consolidation processes are not well understood.
TMS and lesion studies have identified both the motor cortex (Herzfeld et al., 2014; Faiman et al., 2018), and the cerebellum (Donchin et al., 2012; Izawa et al., 2012) as necessary for motor memory formation following dynamic adaptation tasks. However, because of methodological limitations, these studies could not investigate how network level interactions between these regions are associated with consolidation processes. Functional magnetic resonance imaging (fMRI) can localize brain regions associated with task performance (Milner, 2002; Diedrichsen et al., 2005, 2007), while resting state fMRI-acquired after task performance can measure resting state functional connectivity (rsFC), reflective of motor memory formation processes, within and between these brain regions (Albert et al., 2009).
Changes in rsFC associated with adaptation and retention, as well as the concurrent contributions of perceptual learning and impedance control, have primarily been investigated >30 min after task execution (Vahdat et al., 2011, 2014; Babadi et al., 2021). On this timescale, decreases in rsFC within the trained cortico-cerebellar network and increases within the sensorimotor network have been attributed to adaptation. However, different patterns of functional connectivity have been observed during early consolidation of similar tasks, including increased connectivity in the cortico-cerebellar network in visuomotor adaptation (Tzvi et al., 2020), and interhemispheric decreases in the sensorimotor network following robotic training (Sergi et al., 2011). Moreover, longitudinal studies show that patterns of rsFC change with latency following learning, and hold different information about behavioral retention (Sami et al., 2014; Della-Maggiore et al., 2017), making the expected patterns of rsFC during early consolidation of dynamic adaptation unclear.
To address this knowledge gap, we quantified changes in rsFC immediately following a dynamic adaptation wrist-pointing task and investigated their specificity to behavioral features of dynamic adaptation and retention. We used a novel fMRI-compatible robot, the MR-SoftWrist (Erwin et al., 2017), to measure rsFC immediately before and after a motor execution and a dynamic adaptation task executed during fMRI (N = 30), and assessed their behavioral retention the next day. A second group of participants (N = 13) was included as a negative control group, that performed a single day protocol in an unlearnable dynamic perturbation condition to control for changes in rsFC associated with dynamic motor control processes independent from dynamic adaptation.
To identify changes in rsFC with task performance, we compared rsFC measured at baseline to rsFC measured immediately following the motor execution task, the dynamic adaptation task and the unlearnable dynamic perturbation task. We hypothesized that (1) rsFC would change within the trained cortico-thalamic-cerebellar (CTC) network and between cortical sensorimotor regions following the dynamic adaptation task, but show no change following nonadaptation conditions. To identify rsFC associated with behavioral features of adaptation, we performed a linear regression analysis between rsFC and behavior. Adaptation specific features of behavior were identified via variance decomposition analysis for further dissociation of effects in rsFC specific to adaptation. We hypothesized that (2) rsFC measured in both the cortico-thalamic-cerebellar and cortical sensorimotor networks would be predictive of behavioral measures of adaptation and its retention, indicative of their role in motor memory formation.
Materials and Methods
Participants used an fMRI compatible wrist robot, the MR-SoftWrist, to perform wrist pointing tasks in different dynamic conditions. In our main cohort (group 1), the dynamic conditions included a zero force condition that was used as a control condition for motor execution, followed by a constant velocity-dependent lateral force condition that was used to study dynamic adaptation of the wrist (Fig. 1). Subjects returned ∼24 h later to assess behavioral retention of training. A second, negative control group (group 2) was recruited to perform the same wrist pointing task in the zero force condition followed by a randomly alternating lateral force task to quantify effects associated with alternate error reduction strategies (AERS), independent from adaptation. The MR-SoftWrist recorded participants kinetics and kinematics during task performance. In both groups, fMRI was used to measure neural activity associated with active task performance and resting-state BOLD signal before and after each task.
A, Experimental protocol. On day 1, following informed consent (IC) and robotic familiarization, subjects completed the fMRI protocol, a series of resting state functional scans interleaved with motor tasks performed with the MR-SoftWrist in the scanner. Motor tasks included a motor execution task (ME), executed in a zero-force environment, a dynamic adaptation task (DA), executed in a constant velocity-dependent lateral force environment, a random perturbation task (RP), executed in a randomly alternating lateral force environment, and a resistive force task (RF), executed in a velocity-dependent resistive force environment. Resting state scans were acquired to assess functional connectivity at baseline (REST1), following motor execution (REST2) and following either the dynamic adaptation task (group 1, REST3) or the random perturbation task (group 2, REST3). On day 2, group 1 participants performed a behavioral follow-up task (BF) to assess change in behavior in the lateral force field environment that would indicate motor memory formation following day 1 training. The composition of each motor task, including number of active blocks and corresponding force field condition is reported below the experimental protocol diagram. B, A CAD rendering of the MR-SoftWrist (pictured on the right) in its operating condition in the fMRI environment. Red arrows depict the rotational axes used in task execution that correspond to flexion-extension (FE) and radial ulnar deviation (RUD) of the wrist. C, Force field conditions. Gray arrows depict exemplary cursor velocity in task execution, while red arrows depict the corresponding direction of robot applied forces.
Participants
Inclusion criteria required participants to be free from neurologic or musculoskeletal injury, right hand dominant, and free from contraindications to MRI. Group 1: 34 healthy young adults [21 Male (M), 13 Female (F), age: 24 ± 4 years], were recruited to participate in the main cohort. Two of these participants were excluded due to hardware failures (F, 28 years old; F, 30 years old), and two participants were excluded for abnormal task performance, observed as errors that increased across the dynamic adaptation task (M, 21 years old; M, 25 years old). Group 2: an additional 17 healthy young adults (7 M, 10 F, age: 23 ± 4 years) were recruited for the control group. Three of these participants were excluded due to hardware failures (F, 19 years old; F, 21 years old; M, 28 years old), and one participant was excluded for a benign structural abnormality in the cerebellum (M, 28 years old). The study was approved by the Institutional Review Board of the University of Delaware, IRB no. 906215-10.
Experimental procedure
The experimental protocol is shown in Figure 1A. On day 1, all participants first performed a “familiarization” session with the MR-SoftWrist outside of the scanner to familiarize them with task instructions and interaction with the robot. The familiarization task was equivalent to the motor execution task, detailed in the Task Design section below. Following familiarization, participant placement and device set-up in the scanner took roughly 30 min.
The fMRI protocol began with 10 min of structural scans, comprising a localizer, T1 weighted and GRE-field map sequence, followed by three resting states scans interleaved by each motor task. All motor tasks followed a blocked design (Fig. 1A, bottom), with 15-s passive blocks interleaved between blocks of 24 active trials. During these passive blocks, the robot locked in position and a virtual cursor was displayed performing straight point-to-point movements between the targets to match the visual stimulus seen during task performance.
For both groups, the first resting state scan (REST1) was taken to assess baseline resting state functional connectivity. During the resting state scans, participants were instructed to fixate on a white cross displayed on a black background, and to relax but not to close their eyes or fall asleep. Participants then performed the motor execution task, which consisted of 144 trials (six blocks) performed in a zero force condition. REST2 was acquired immediately following task completion to measure neural activity at rest following motor execution of wrist pointing.
Following REST2, participants in group 1 performed a dynamic adaptation task that consisted of 168 trials divided into seven blocks. The first six blocks were performed in a constant velocity-dependent lateral force condition (Shadmehr and Mussa-Ivaldi, 1994), and the last block was performed in the zero force condition to assess any after effects indicative of adaptation. REST3 was acquired immediately following the dynamic adaptation task to measure neural activity at rest during early stages of consolidation following dynamic adaptation. After REST 3, subjects performed a resistive force task that consisted of seven blocks, and 168 trials. The first block was performed in the zero force condition, followed by six blocks in a resistive force condition (Diedrichsen et al., 2005; Suminski et al., 2007). This task was included to assess task-related activations unrelated to the current aims of this study. As such, we list it here for completeness of the experimental protocol executed by participants in group 1, but will not present results related to this task.
Following REST 2, participants in group 2 performed a random perturbation task, that consisted of 168 trials divided into seven blocks, to match the structure of the dynamic adaptation task (Fig. 1A, top right). The first six blocks were performed in a velocity dependent lateral force condition that alternated between clockwise (CW) and counterclockwise (CCW) directions (Shadmehr and Holcomb, 1997; Gupta and Ashe, 2007), and the last block was performed in the zero force condition to confirm the absence of after effects, as no significant adaptation is expected to occur in this task. Changes in force-field direction occurred with a 50% probability between each trial, and logic was applied to prevent more than four consecutive trials occurring in a constant field (defined as either a constant field direction, i.e., CW-CW-CW-CW, or constant alternation between field directions, i.e., CW-CCW-CW-CCW). Clockwise and counterclockwise perturbations were evenly balanced across this task, such that the average lateral force experienced across the task was 0. REST3 was acquired immediately following the random perturbation task to measure neural activity at rest following a dynamic perturbation task that elicits alternate error reduction strategies, but no significant adaptation.
On day 2, participants in group 1 performed a behavioral follow-up task to assess behavioral retention of day 1 training. Behavioral follow-up sessions were conducted ∼24 h after the initial session (mean ± SD: 23 ± 1.7 h, range: [21 29] h). Participants performed the task in the supine position in a mock scanner built to the same dimensions as the MRI scanner. The follow-up task consisted of 240 trials including 10 active blocks interleaved with 15-s passive blocks. The first two blocks were performed in the zero force condition, blocks 3–8 were performed in a constant lateral force condition, and blocks 9–10 were performed in the zero force condition.
MR-SoftWrist
The MR-SoftWrist is an fMRI-compatible robotic exoskeleton (Fig. 1B) that supports wrist pointing in flexion-extension (FE) and radial-ulnar deviation (RUD) in a circular workspace with a radius of 20°. The device has a maximum output torque of 1.5 N·m about each axis and implements gravity compensation and force feedback at 1000 Hz. The MR-SoftWrist can display a variety of kinesthetic environments to the user, ranging from a zero force mode that minimally perturbs the user's movements (trial average resistive torque = 0.05 N·m; peak resistive torque <0.17 N·m), to a stiffness control mode that displays a high stiffness environment (max virtual stiffness, kv = 0.45 N·m/°; Erwin et al., 2017). Interaction force and velocity were logged at a rate of 1000 Hz, while joint trajectories used for behavioral data were logged at a rate of 100 Hz. The MR-SoftWrist introduces no change in the noise characteristics of functional images or measures of neural activity acquired during operation of the robot (Farrens et al., 2018). For more on the control and design of the MR-SoftWrist, see Erwin et al. (2017).
Task design
For all motor tasks, subjects grasped the handle of the MR-SoftWrist with their right hand and moved their wrist to control a cursor displayed on a monitor. The cursor was displayed continuously throughout all tasks as a gray circle (radius 1°). Flexion-extension of the wrist moved the cursor horizontally, while radial-ulnar deviation moved the cursor vertically. Pronation-supination was prevented by a forearm support. Subjects were cued to make alternating flexion-extension rotations to move a cursor in a straight line to one of two targets (radii 1.25°) located at (±10, 0) degrees in flexion-extension, radial-ulnar deviation, respectively (Fig. 1B,C). Trial onset was cued by a change in target color from black to blue. Trial completion was achieved when the error between the cursor and the target was <2° for >250 ms. After trial completion, the reached target provided timing feedback for 0.5 s by turning red if the movement duration was >650 ms or green if it was <300 ms. Otherwise, the target remained black. Duration of the inter-trial interval between feedback and the cueing of the next trial was chosen at random from a truncated normal distribution of with a mean of 0.75 s, a SD of 0.25 s, and bounds of [0.25 1.75] s.
In our experimental protocol, the robot operated in one of four control modes: a zero force mode, a lateral force mode, an error clamp mode and a resistive force mode (Fig. 1C). In the zero force mode, the desired value of the interaction torque is set to zero (
fMRI acquisition
All fMRI data were acquired at the Center for Biological and Biomedical Imaging at the University of Delaware, on a 3.0 T Siemens Prisma MR scanner using a 64-channel head coil. Participants laid in the scanner in the supine position, with their head immobilized by foam padding. Earplugs were provided to reduce scanner noise, as were headphones for communication with experimenters in between imaging sequences. Following a brief initial localizer scan (Siemens AA Head Scout, spatial resolution: 1.6 × 1.6 × 1.6 mm3, isotropic; TR = 3.15 s, TE = 1.37 ms), we performed a T1-weighted anatomic scan (spatial resolution 1 × 1×1 mm3, isotropic; TE = 3.02 ms; TR = 2.30 s), followed by a GRE field map (spatial resolution 3 × 3 × 3 mm3, TR = 400 ms, TE = 4.92 ms). All subsequent functional scans (task-based and resting state) were acquired using a multiband T2*-weighted EPI sequence (axial slices oriented to the AC-PC line, spatial resolution 2 × 2 × 2 mm3, 60 slices, TE = 30 ms, TR = 1 s, acceleration factor = 4). For all functional scans, subjects viewed a monitor displaying visual stimuli through an angled mirror attached to the head coil.
Experimental design and statistical analyses
This study included two groups. Our main cohort (group 1) performed two sessions, conducted ∼24 h apart. The first session, conducted on day 1, included a motor execution task (zero force condition) to assess behavior and neural activations associated with task execution in the absence of motor learning. This served as a control condition to identify change in behavior and neural activity associated with dynamic adaptation in the subsequent dynamic adaptation task (constant lateral force condition). Resting state scans were acquired before and after each motor task, and behavioral retention of training was assessed the following day in a matched dynamic adaptation task. A second group of participants (group 2) was recruited to perform a single session experiment. These participants performed the same day 1 protocol as group 1, with the exception that group 2 was exposed to an unlearnable perturbation schedule designed to elicit alternate error reduction strategies without adaptation instead of the dynamic adaptation task. The subsequent REST3 condition was used as a negative control for changes in rsFC associated with alternate error reduction strategies independent from dynamic adaptation.
Resting state functional connectivity (rsFC) was defined between regions of interest (ROIs) within two task-localized networks, a bilateral cortical sensorimotor network and a cortico-thalamic-cerebellar network for each resting state condition. A mixed model analysis was used to estimate rsFC between each edge (ROI pair) included in this network, and model parameter estimates were used to determine significant change in functional connectivity following motor execution, dynamic adaptation, and unlearnable dynamic perturbations. For group 1, one-way repeated measures ANOVA of behavioral kinematics and kinetics was used to establish behavioral evidence of adaptation and its retention at the group level. To further investigate the association between rsFC and behavior, we performed separate linear regression analyses between all edges within each network to behavior measured on day 1, and behavior on day 2, indicative of motor memory formation, correcting for multiple comparisons. A variance decomposition analysis was used to identify behavioral features of adaptation independent from alternate motor control strategies, and used in a post hoc analysis of edges with significant associations to behavior to further dissociate between adaptation-specific effects and alternate error reduction strategies.
With 30 subjects, the experiment executed with our main cohort is powered (β > 0.90) to detect an effect size of 0.83 for paired tests between REST conditions at the significance threshold α < 0.0033, corrected for multiple comparisons via Bonferroni correction. Sample size was informed by previous studies identifying changes in resting state functional connectivity associated with motor adaptation at longer timescales [Mawase et al. (2017) reported an effect size of 0.86; Vahdat et al. (2011) reported an effect size of ∼1]. To control for multiple comparisons in our regression analysis between rsFC and behavior, we used a Bonferroni correction to control for the number of edges within each network of interest. Because of the difference in sample size from group 1, for group 2 (negative control group), changes in rsFC were compared in terms their Cohen's d effect size coefficient, and significance threshold was established using the coefficient yielding the pFWE < 0.05 for the initial group of participants (|dmin| = 0.324).
Behavioral data analysis and measures
Data preprocessing
Processing of behavioral data were conducted using MATLAB (The MathWorks, version 2020b). For data processing, position data were up-sampled to a 1000-Hz sampling rate to match velocity and force data using the MATLAB resample function. All position, velocity, and force data were then low-pass filtered with a 4th order Butterworth filter with a 25-Hz cut off frequency using MATLAB's filtfilt function. For trial-by-trial data analysis, trial onset was defined as the instant the absolute cursor velocity exceeded 20°/s (2.5 times greater than the minimum measurable velocity given encoder resolution and sampling rate). Trial end was defined as the instant the cursor was within 3° of the target in the flexion-extension direction. Trials were excluded if they matched any of the following conditions: (1) a trial duration of <200 ms, or >700 ms; (2) a maximum velocity below 40°/s; (3) a reversal in goal directed velocity that occurred before the trial max velocity, indicative of false starts. At the group level, this resulted in an exclusion of <3.1% of all field trials (zero-force, lateral force, random force) and <2.0% of error clamp trials on day 1, and <2.8% of all field trials and <2.1% of error clamp trials on day 2.
For each trial, position and force data from trial onset to trial end were resampled into 1000 data points, and were divided between extension and flexion movements. Within each day, direction-specific baseline profiles were defined by averaging across all successful zero-force trials for force and position data. Direction-specific baseline profiles were then subtracted off all measured profiles, such that all behavioral metrics reported reflect a change relative to typical, nonperturbed wrist pointing.
Behavioral metrics
All behavioral metrics were calculated in the first 150 ms after movement onset to capture behavior associated with feed-forward motor control processes that most directly reflect participants preplanned motor actions based on an internal model of expected task dynamics (Crevecoeur and Scott, 2014; Heald et al., 2018). To quantify wrist pointing kinematics in all conditions, we calculated the angular trajectory error as the internal angle between the cursor and the start and end targets, taken at the cursor's maximum lateral deviation within the first 150 ms after trial onset (Diedrichsen et al., 2005). To quantify wrist kinetics during error clamp trials, we calculated the adaptation index as the area under the measured force profile divided by the area under the ideal force profile necessary to compensate for the lateral force field within the first 150 ms after trial onset (Izawa et al., 2012; Farrens, 2022). As error clamp trials occur unexpectedly, the lateral force profiles measured in these trials reflect participants expectation of the required task dynamics, that is modulated by the exposure and removal of the lateral force field. An adaptation index of one represents perfect adaptation to the lateral force field, while an adaptation index of zero represents no adaptation.
Group level behavioral analysis
We used repeated measures one-way ANOVAs to test for evidence of adaptation and its retention in our behavioral measures, for interpretation of changes in resting state functional connectivity observed across each task at the group level. Adaptation is measurable as increases in adaptation index during exposure to a lateral force field, that persist transiently and cause after-effects in trajectory errors when the perturbation is removed (Smith et al., 2006; Shadmehr et al., 2010). Gains in adaptation index should coincide with reductions in trajectory error, although trajectory errors are reflective both of adaptation and alternate motor control strategies. Retention is measurable as increases in adaptation index on day 2 compared with day 1, paired with greater decreases in trajectory errors. For trajectory errors, specific features of retention include increases in the rate of error reduction (i.e., learning rate) between days, and decreases in initial trajectory errors between days, termed “recall” (Mawase et al., 2017)
To test for the presence of these behavioral effects in group 1, we performed two separate ANOVAs to test for an effect of experimental phase on trajectory error and adaptation index measured across tasks performed on both days. For adaptation index, the day 1 experimental phases included Baseline, defined as the average adaptation index measured in the last block (block 6) of the motor execution task, Early lateral force and Late lateral force defined as the average adaptation index measured in block 1 and block 6 of the dynamic adaptation task, respectively, and After Effects, defined as the average adaptation index measured in the first block of zero-force trials following the lateral force condition in the dynamic adaptation task (block 7). For day 2, experimental phases were defined similarly, such that day 2 Baseline, Early lateral force, Late lateral force and After Effects phases corresponded to the average adaptation index measured in blocks 2, 3, 8, and 9 of the behavioral follow-up task, respectively (Fig. 1, bottom left).
For trajectory error, we defined experimental phases Baseline and Late lateral force in the same way as for adaptation index data for day 1 and day 2. We defined an Initial Errors phase as the average trajectory error measured in trials 2–4 of block 1 of the dynamic adaptation task for day 1, and in block three of the behavioral follow-up task for day 2. The first trial was excluded as it was assumed to act as the initial cue to participants to recall any previously learned response to the lateral force field (Mawase et al., 2017). We then defined a Learning rate phase for day 1 and day 2 as the average trajectory error measured in the subsequent trials (5–24) in block 1 and in block 3, respectively. Finally, the After Effects phase was defined as the average trajectory error measured in trails 1–4 of the first zero force block following the lateral force condition on both days (block 7 and 9, respectively), to capture the more transient nature of after effects in trajectory error data.
When the ANOVA returned a significant effect, post hoc Tukey's test was used to quantify the effect of experimental phase on adaptation index and trajectory error. For adaptation index data, increases in the Early lateral force, Late lateral force and After Effects phases compared with Baseline would all be indicative of adaptation. Expected increases in adaptation index between the Early and Late lateral force phases should be paired with significant decreases in trajectory errors (i.e., error reduction) across this same experimental phase. For trajectory error data, decreases in the After Effects phases from the Baseline phase, indicative of errors in the opposite direction of the applied perturbation, would provide evidence of significant adaptation. Between days, evidence of retention would be supported by a significant decrease in trajectory error, or increase in adaptation index, across any paired experimental phase (i.e., day 2 Late Lateral Force vs day 1 Late Lateral Force).
For group 2, we used the random force task to elicit alternate error reduction strategies that are independent from adaptation, as alternations occur too rapidly for significant adaptation to the lateral force field to occur (Takahashi et al., 2001; Gupta and Ashe, 2007). Resting state functional connectivity measured following this task was used to identify changes associated with alternate dynamic control processes contributing to error reduction, independent from effects specific to dynamic adaptation. As such, we performed the same repeated measures one-way ANOVA on group 2 adaptation index data using the experimental phases defined for the day 1 tasks in group 1, to confirm the absence of significant evidence of adaptation in the random force task. We additionally performed a paired t test of average trajectory error data measured in the after-effects phase to the baseline phase, to confirm the absence of after effects following the random force condition.
Subject-specific behavioral summary metrics
To investigate whether resting state functional connectivity is associated with variability in adaptation behavior and its retention, we defined subject-specific metrics of behavior measured on day 1 and day 2 for group 1 data. Data measured at the individual subject level is highly variable, and easily influenced by behavioral and measurement noise. To mitigate effects of noise, we used a double-exponential model of the form:
To quantify the degree to which each subject adapted to the lateral force field we used two primary metrics: magnitude of adaptation index and magnitude of after effects. To quantify individuals' magnitude of adaptation index, we fit a model of the form
To quantify effects associated with error reduction, which is driven by adaptation as well as alternate error reduction strategies, we fit a double exponential model of the form: ER
All model fits were performed using a nonlinear least squares fit to subject-level data using MATLAB's fit function. To avoid local minima, we performed each fit 500 times starting with a different initial value for each parameter and selected the resultant best fit for each participant. All participants' behavioral data were well fit by the double exponential model [Trajectory error data day 1 (mean ± SEM): R2 = 0.300 ± 0.0055; day 2: R2 = 0.303 ± 0.0048; Adaptation index data day 1: R2 = 0.386 ± 0.0058; day 2: R2 = 0.397 ± 0.0065; After-effects day 1: R2 = 0.464 ± 0.0068; day 2: R2 = 0.463 ± 0.0071]. To validate that behavioral associations established using model-based measures were robust to arbitrary definitions of subject-specific behavior, we additionally performed all rsFC analyses (detailed below) with behavioral metrics based on average behavior measured in the same experimental phases as our model-based metrics. All average-based metrics showed a strong correlation (all R > 0.92) to our model-based definitions.
We used all three model-based behavioral metrics to investigate rsFC associated with adaptation, as each measure suffers from unique measurement and behavioral noise. The magnitude of adaptation index relies on a small number (18 trials) of infrequently sampled data points (1/8th frequency) that are highly susceptible to measurement error and motor variability, while both relative error reduction and magnitude of after effects use a greater number of samples (126 and 42 trials, respectively), but may include components of alternate control strategies. As such, all metrics contain information about adaptation learning as well as variability because of motor execution, measurement error, and utilization of alternate control strategies.
Variance decomposition
As a goal of this study is to identify changes in resting state functional connectivity specific to adaptation, we sought to further decompose the variance in our behavioral metrics into a part that is correlated predominantly with adaptation-specific behavior and a residual uncorrelated component, associated with alternate error reduction strategies (AERS).
To achieve this dissociation, we used a previously-developed variance decomposition method (Vahdat et al., 2011). Using the same methodology, we aimed to identify a mutual component of variance between two metrics (X and Y), such that:
Next, we aimed to isolate variance within one behavioral metric that is independent from another behavioral metric, such that:
The solution to both equations is shown in Equation 3:
Before variance decomposition all metrics were rescaled between 0 and 1, using a min-max normalization.
To quantify adaptation-specific behavior, we used the magnitude of adaptation index and magnitude of after effects metrics, which most directly reflect recalibration of the internal model resulting from adaptation. Using Equations 1 and 3, we set X = magnitude of adaptation index and Y = magnitude of after effects to define a mutual component of adaptation (M = MA),
Next, we used Equations 2 and 3 to identify variance in the relative error reduction metric that is independent from adaptation-specific behavior (MA), and instead attributable to alternate error-reduction strategies (AERS). For this decomposition, we set X = MA and Y = Relative Error Reduction, such that:
The resulting AERS metric quantifies the variance in the relative error reduction metric that is independent from adaptation, and instead attributable to alternate error reduction strategies. We expect that the AERS metric will predominantly capture behavior associated with co-contraction associated with impedance control; however, it can also be influenced by explicit aiming based strategies and other model-free learning strategies (Franklin et al., 2003; Diedrichsen et al., 2010; Huang et al., 2011; Heald et al., 2018; Farrens et al., 2022).
Importantly, while our original behavioral metrics shared mutual information about the adaptation process, these two metrics are independent from each other. As such, these two metrics can be used in a post hoc exploratory analysis to determine whether associations identified between rsFC and our primary behavioral metrics (magnitude of adaptation index, magnitude of after effects, and relative error reduction) were driven by adaptation-specific processes (MA), or alternate error reduction strategies (AERS).
fMRI data analysis and measures
fMRI data preprocessing
Preprocessing of fMRI data were performed using the SPM12 software (Ashburner et al., 2016). All task-based and resting state functional images were unwarped and field map corrected, realigned, normalized into standard (MNI) space for group analysis, and smoothed with a 6 mm Gaussian kernel. We used the artifact detection tools (ART) toolbox to identify outlier volumes, defined as volumes with framewise displacement above 0.9 mm or global BOLD signal changes above 5 SDs (Whitfield-Gabrieli, 2009). Outlier scans, along with rotational and translation head movement parameters were included as nuisance regressors of noninterest in our tasked based analysis. For resting-state data, we used a multiple regression analysis to remove effects of noninterest that included the same nuisance regressors as our task-based data as well as the first five principal components extracted from the white matter and CSF tissue maps to model physiological noise (Shehzad et al., 2009). Additionally, resting state data were bandpass filtered between 0.008 and 0.09 Hz to isolate the expected frequency range for signal related to neural activity (Fox and Raichle, 2007).
ROI identification
Task-based fMRI scans were used to determine relevant regions of interest (ROIs) for use in a ROI-based analysis of our resting state data. For all motor tasks, we used a boxcar function to model active task blocks versus passive blocks that was then convolved with the standard hemodynamic response function in SPM12 to create task-related regressors. For the motor execution task, this regressor was used to identify neural activation associated with unperturbed wrist pointing. For the dynamic perturbation tasks (group 1: Lateral force, group 2: Random force), this regressor was used to identify regions associated with dynamic motor control. Regressors were entered into a first-level general linear model (GLM) analysis, and contrasts between coefficients associated with active and passive task regressors were used to produce subject-specific t-maps of task-related activation.
For the dynamic adaptation task, subject level contrasts were entered into a second level analysis, and a one-sample t test was used to identify group-level activation associated with dynamic motor control. From this activation map, we selected a subset of regions with significant task-related activation based on prior knowledge of brain networks associated with adaptation (Sergi et al., 2011; Vahdat et al., 2011; Mawase et al., 2017) for use in a restricted ROI-based analysis to limit the number of multiple comparisons in our analysis. These ROIs included the bilateral primary motor cortices (M1, BA4), premotor cortices (PM, BA6), and primary sensory cortices (S1, BA1-3), defined using the Juelich histologic atlas for cortical regions (Eickhoff et al., 2007), the left thalamus, defined using the Harvard-Oxford atlas for subcortical regions (Desikan et al., 2006), and the right cerebellar lobules 6 and 8 (CB6, CB8) defined using the MNI FLIRT atlas (Diedrichsen et al., 2009), all thresholded at 50% probability.
For the left (contralateral) cortical areas in the sensorimotor network, which are highly engaged during task execution, we defined subject-specific ROIs based on task related activation measured within each anatomic region of interest (M1, PM, S1) for both group 1 and group 2 data. Subject specific t-maps corresponding to the contrast of Dynamic Perturbation > Passive Display were thresholded at t > 0 and weighted by the contrast of (Dynamic Perturbation > Passive Display) – (Zero Force > Passive Display), where Dynamic Perturbation was either the lateral force or random force condition, and Passive Display corresponded to the passive blocks within each task. From these maps, we determined the center of mass of task related activation within each anatomic ROI and constructed a spheroid centered on the center of mass using the fslmaths function (https://fsl.fmrib.ox.ac.uk/). ROIs were created with a target volume of 180 voxels, which corresponds to a spheroid with a radius of 7 mm (Mawase et al., 2017). However, as all ROIs were bounded by the borders of the anatomic region of interest, spheroids were iteratively increased in radius to reach the desired volume. The spheroid radius was increased by 1 mm up to a limit of 14 mm, so long as the absolute error between the desired target volume and achieved volume decreased and the activation cluster within the ROI remained contiguous. We used this process to create subject-specific ROI within the left primary motor, premotor and primary sensory regions for both group 1 and group 2 (Resultant ROI volumes: L-M1: 184.3 ± 15.8 [136, 222] voxels (mean ± SEM [range]); L-PM: 192.9 ± 18.1, [155, 254] voxels; L-S1: 193.4 ± 31.0 [155, 254] voxels).
The cortico-thalamic-cerebellar (CTC) network of the trained wrist was then defined as the left cortical motor areas, left thalamus, and right cerebellar ROIs (L-M1, L-PM, L-S1, L-Thalamus, R-CB6, R-CB8), while the cortical sensorimotor network comprised the six bilateral cortical motor ROIs (L-M1, L-S1, L-PM, R-M1, R-S1, R-PM). Both networks were formed as a complete graph with 6 nodes and 15 edges.
Resting-state functional connectivity analysis
We first aimed to determine whether functional connectivity was modulated by task performance in any edge of the networks of interest. Based on previous works, we hypothesized that there would be no significant change following motor execution, as limited learning occurs within this task, and that there would be significant change following dynamic adaptation resulting from adaptation learning and subsequent memory formation (Albert et al., 2009). To determine whether these changes were unique to dynamic adaptation, or general across dynamic motor control processes, we compared changes following the random perturbation task observed in group 2 with those observed across the dynamic adaptation task in group 1.
Resting state fMRI analysis was performed using the CONN toolbox (Nieto-Castanon, 2020). First, for all subject-level data in each resting state condition, the average timeseries from all voxels within each ROI was calculated. For each condition, resting state functional connectivity (rsFC) between each ROI pair was then quantified as the Fisher-transformed correlation coefficient between the average timeseries for each ROI. As such, each ROI pair (network edge) was assigned a scalar representing the functional connectivity between the two ROIs (nodes) for all participants in each condition.
For our restricted ROI-based analysis, we performed a mixed model analysis to test for the effect of condition (REST1, REST2 and REST3) on measured rsFC for each edge in our two networks of interest [e.g.,
We performed the restricted ROI-based analysis separately in our two networks of interest (i.e., a 15-edge cortical sensorimotor network, and a 15-edge cortico-thalamic cerebellar network, as defined above). For group 1 analysis, we corrected for multiple comparisons associated with the multiple edges within each network using a Bonferroni correction for the model significance (pFWE < 0.05), such that model results were deemed significant at p < 0.0033 (0.05/15 edges). Because of the conservative nature of this correction that assumes independence between each edge of the network, edges that achieved a reduced model significance of punc < 0.01 are also reported within the results section. To account for the smaller sample size of group 2, group 2 results are reported by their Cohen's d effect size coefficient, and thresholded using the coefficient yielding the pFWE < 0.05 for group 1 participants (|dmin| > 0.324).
Finally, to confirm that our restricted ROI-based analysis included relevant brain regions for investigation of resting state functional connectivity, we performed an unrestricted GLM analysis using our three subject-specific ROIs in the left sensorimotor network (L-M1, L-S1, L-PM) as seed ROIs for investigation of resting state activity measured across the whole brain. In this analysis, the seed ROI's average timeseries is regressed onto each voxel timeseries measured across the whole brain, to create subject level β-maps of functional connectivity across the whole brain associated with the seed ROI. To quantify baseline rsFC associated with each seed ROI, we performed a group-level one-sample t test on REST1 data. To identify clusters with significant change in rsFC between rest conditions associated with each seed ROI, we performed group-level paired t tests between conditions. As in the previous analysis, comparison of REST3 and REST1 was used to test for effects of dynamic adaptation and other dynamic motor control processes and comparison of REST2 and REST1 was used to test for effects of motor execution. We chose a standard threshold of punc < 0.001 (at the voxel-level), pFDR < 0.05 (at the cluster level) to test for significant effects for each map.
Association between functional connectivity and adaptation behavior
Primary behavioral measures associated with rsFC
To investigate associations between variations in adaptation behavior measured on day 1 and rsFC measured in different resting state conditions, we performed a linear regression between day 1 behavior (magnitude of after effects, magnitude of adaptation index, and relative error reduction) and rsFC measured within each restricted network of interest (cortical sensorimotor, cortico-thalamic-cerebellar). We used the general model (Eq. 4):
To account for multiple comparisons associated with the multiple edges within each network, we used a Bonferroni correction (pFWE < 0.05), such that model results were deemed significant if the reduced model reached a significance of p < 0.0033 (0.05/15 edges). Within the reduced model, we then considered model terms significant after correcting for the number of terms within the reduced model (
For investigation of associations to behavior on day 2, we performed the same analysis, but using a general model that includes a term for day 1 behavior (Eq. 5):
For this analysis, the day 1 behavior term was not removed regardless of its significance (i.e., the backwards removal process only applied to parameters
To confirm that model terms included in our linear regression analysis do not violate assumptions of independence, we calculated the variance inflation factor (VIF) for all potential models included in the analysis pipeline, which determines the severity of multicollinearity in an ordinary least squares regression analysis. All models had a VIF below a conservative threshold factor of 5 (max VIF = 4.2, average VIF = 1.1).
Post hoc rsFC variance decomposition
For edges with resting state functional connectivity that was significantly associated to our primary behavioral measures in the linear regression analysis, we performed a post hoc analysis using the mutual adaptation (MA) and alternate error reduction strategies (AERS) metrics to determine whether observed effects in our rsFC conditions (REST1, REST2-REST1, or REST3-REST1) were driven by adaptation specific processes (MA), or alternate control processes (AERS). For day 1 associations, we used a linear model of the form (Eq. 6):
For day 2 associations, we sought to capture effects associated with motor memory formation after controlling for day 1 behavior. As such, for edges identified in the linear regression analysis with significant associations between rsFC and behavior, we performed a direct analysis between significant rsFC terms and day 2 MA and AERS metrics independently, including a day 1 term to account for effects of day 1 behavior (e.g.,
Results
Behavioral analysis
Analysis of group 1 trajectory error and adaptation index data showed evidence of adaptation and retention, indicating that changes in resting state functional connectivity following the dynamic adaptation task may be influenced both by adaptation behavior and its subsequent consolidation into motor memory. Analysis of the group 2 behavioral data showed no significant evidence of adaptation, indicating that changes across the random force task should reflect processing of independent dynamic control processes.
Dynamic adaptation task
Adaptation Index
The ANOVA fit to the group 1 adaptation index data had an R2 adj = 0.639, and identified a significant effect of experimental phase (DF: 7. F Ratio: 49.54, p < 0.001). Post hoc Tukey's testing identified significant increases in all experimental phases compared with Baseline (p < 0.001), confirming that significant adaptation occurred on both days (Fig. 2, bottom left). On both day 1 and day 2, adaptation index significantly increased over the course of the lateral force task (Late LF > Early LF, p < 0.001), as expected. Contrasts performed between parameter estimates showed that adaptation index in the late lateral force phase was larger on day 2 compared with day 1, but not significantly different (
Top, Group 1 average behavior across all tasks (left, adaptation index; right, trajectory error). Black dots denote the average behavior measured on each trial and gray shaded regions represent the standard error. Experimental phases of interest included in the mixed model analysis are indicated at the bottom of each figure. Bottom, Bar plots of behavior measured within experimental phases included in the mixed model analysis. Error bars depict standard error within each phase, and greyscale dots depict individual subject data measured within each phase. Dashed vertical lines delineate day 1 and day 2 behavior. Red asterisks indicate phases that are significantly different from baseline behavior, while magenta asterisks and lines denote significant changes in behavior measured between days (p < 0.05). Left, Average adaptation index. For space, experimental phases have been abbreviated as follows: Baseline as BL, Early Lateral Force as ELF, Late Lateral Force as LLF, and After Effects as AE. Right, Average trajectory error. For space, experimental phases have been abbreviated as follows: Baseline as BL, Initial Errors as IE, Learning Rate as LR, Late Lateral Force as LLF, and After Effects as AE. Day 1 and day 2 conditions are denoted by a 1 or 2, respectively.
Trajectory error
The ANOVA fit to the group 1 trajectory error data had an R2adj = 0.930, and identified a significant effect of experimental phase on trajectory errors (DF: 9. F Ratio: 397.36, p < 0.001). Post hoc Tukey's testing showed significant After Effects compared with Baseline on both days (p < 0.001). On both days, there were significant increases in Initial Errors over Baseline, that significantly decreased across the lateral force condition (all experimental phase comparisons p < 0.001; Fig. 2, bottom right). Between days, contrasts of parameter estimates showed there were significantly smaller initial errors on day 2 compared with day 1, consistent with “recall” (
Random force task
The ANOVA fit to the group 2 adaptation index data showed no significant effect of experimental phase (DF: 3. F Ratio: 1.236, p = 0.31), in line with our expectations that no significant adaptation occurred in the random force task. Paired t test of trajectory error data similarly showed no significant difference in after effects from baseline errors (BL: −0.65 ±1.46, after effects: 0.19 ± 1.83, p = 0.23).
Variance decomposition results
By definition, the variance decomposition analysis resulted in metrics for mutual adaptation (MA) and alternate error reduction strategies (AERS) that were independent from each other on both days. To confirm the expected associations between these metrics and our primary behavioral metrics, we performed correlational analyses between the MA and AERS metrics to our primary behavioral metrics of relative error reduction, magnitude of adaptation index, and magnitude of after effects.
We expected that variance in magnitude of adaptation index and magnitude of after effects would be largely explained by our measure of mutual adaptation for both days, with no association to alternate error reduction strategies. For day 1 behavior, magnitude of adaptation index and magnitude of after effects were significantly correlated with MA (R2 = 0.841, and R2 = 0.647, respectively, both p < 0.001), with no association to AERS (R2 = 0.003, p = 0.79 and R2 = 0.006, p = 0.68, respectively). The same pattern of results was observed on day 2, for MA (magnitude of adaptation index R2 = 0.683; magnitude of after effects R2 = 0.977, both p < 0.001), and AERS (magnitude of adaptation index R2 = 0.003, p = 0.81; magnitude of after effects R2 = 0.006, p = 0.95).
For relative error reduction, which captures behavior associated with adaptation as well as alternate error reduction strategies, we expected to see a significant association to both MA and AERS. For day 1, relative error reduction showed a significant association to both MA (R2 = 0.282, p = 0.0025) and AERS (R2 = 0. 718, p < 0.001), as expected. For day 2, AERS explained a significant portion of the variance observed in relative error reduction (R2 = 0.718, p < 0.001), while MA did not (R2 = 0.119, p = 0.062). The lack of significant association between MA and relative error reduction on day 2 could be because of individuals using a different mix of control strategies in response to dynamic perturbations, or influenced by retention of day 1 learning. Retention would result in a lower relative error reduction due to recall, paired with greater adaptation resulting from storage of the previously learned internal model, decreasing a direct correlation between the two metrics.
Representative subjects with different magnitudes of MA and AERS are provided in Figure 3. It is important to note that while these metrics are independent, they are not mutually exclusive, meaning that individuals can use both adaptation and alternate error reduction strategies during task execution. Indeed, all subjects in our study used some combination of AERS and MA processes during dynamic adaptation task performance, with average AERS (0.3486 ± 0.037 n.u.) comparable to average MA (0.3301 ± 0.031 n.u.). As the intrinsic dynamics of the wrist are dominated by stiffness (Charles and Hogan, 2011; Formica et al., 2012), and alternations in perturbation direction between right and left directed trials likely mediate effects of model-free learning strategies (Diedrichsen et al., 2010; Huang et al., 2011), these contribution are most likely due to concurrent impedance control (Franklin et al., 2003; Farrens et al., 2022).
Adaptation index data and normalized trajectory errors from representative participants. Red lines depict exponential model fits to adaptation index data, magenta lines depict exponential model fits to after effects in trajectory errors, and blue lines depict exponential model fits to trajectory errors in the lateral force condition. Trajectory errors are normalized by the model estimated initial trajectory error at the onset of the lateral force condition. Subject 021 is an example of an individual with a large end magnitude of adaptation and after effects (∼65% of the perturbation) that mostly explain the percent error reduction observed (∼75%). This subject represents an individual classified by our variance decomposition as having a large MA and low AERS. Subject 018 exhibited roughly the same level of error reduction as subject 021 (∼75%) but lower levels of adaptation (∼40%). This subject is representative of an individual classified as having roughly equivalent use of MA and AERS. Finally, Subject 005 is an example of an individual with low end magnitude of adaptation and after effects (∼25% of the perturbation) with much larger error reduction (∼60%). This subject represents an individual classified as having a large AERS and low MA.
Given the relative contributions of both control strategies to task execution, it is likely that rsFC following the dynamic adaptation task will reflect activity resulting from adaptation as well as alternate error reduction strategies. As such, these independent measures of MA and AERS can be used to identify adaptation specific effects from alternate control processes in rsFC associated to our primary behavioral metrics.
Change in rsFC with task execution
We used a mixed model analysis to identify changes in rsFC measured in each condition (REST1, REST2, REST3) within the sensorimotor and cortico-thalamic-cerebellar networks. Comparison of REST2-REST1 was used to identify changes associated with motor execution, while REST3-REST1, and REST3-REST2 were used to identify changes associated with adaptation and other forms of dynamic motor control. Following the motor execution task, in which we hypothesized no changes would occur, our mixed model analysis identified interhemispheric decreases in rsFC within the cortical sensorimotor network. Following the dynamic adaptation task, in which we expect to see changes associated with adaptation as well as other dynamic motor control processes, both the mixed model and seed-to-voxel analysis identified decreases in interhemispheric rsFC within the sensorimotor network, and significant increases within the cortico-thalamic-cerebellar network. Following the random perturbation task, only decreases in interhemispheric rsFC within the sensorimotor network were observed. A detailed breakdown of each analysis is provided below.
Restricted network analysis
The mixed model identified significant changes in rsFC following motor execution and both dynamic perturbation tasks (Figs. 4, 5). Following motor execution, the model identified decreases in rsFC between the right and left motor cortices and an increase in rsFC between the left primary motor cortex and the left thalamus (Model pFWE ≤ 0.05: L-PM to R-M1: p = 0.014, effect size: −0.430; L-M1 to L-Thalamus, p = 0.0032, effect size: 0.595; Model punc ≤ 0.01: L-PM to R-S1: p = 0.005, effect size: −0.466).
Results of the restricted ROI model based analysis for group 1 participants. Top, Brain schematic shows the location of the restricted sensorimotor (short dashes) and cortico-thalamic-cerebellar (long dashes) ROI networks. Differences between conditions within each network are reported in dark blue and red lines at the pFWE < 0.05 threshold, and in faded blue and red lines at the p < 0.01 threshold. Middle, Model results with pFWE < 0.05 are reported in the top three rows, while results with p < 0.01 uncorrected are reported in the bottom three rows. Bottom, Barplot of rsFC measured in each condition for all edges reported in the table. Dots represent subject-level data, and error bars depict the mean and standard error within each condition. Red asterisks denote significant differences from REST1 (R1) and black asterisks denote significant differences from REST 2 (R2). REST3 condition is abbreviated as R3, and thalamus has been abbreviated as “Thal”.
Results of the restricted ROI model based analysis for group 2 participants who performed the random force task in place of the dynamic adaptation task. Top, Brain schematic shows the location of the restricted sensorimotor (short dashes) and cortico-thalamic-cerebellar (long dashes) ROI networks. Differences between conditions within each network are reported with a significance, thresholded based on Cohen's d (|d| > 0.324). Middle, The table reports model results for group 2 rsFC in ROI pairs that showed significant effects in our primary analysis (group 1). Results previously observed at pFWE < 0.05 are reported in the top three rows, while results observed at punc < 0.01 are reported in the bottom three rows. Bottom, Barplot of rsFC measured in each condition for all edges reported in the table. Dots represent subject-level data, and error bars depict the mean and standard error within each condition. Red asterisks denote significant differences from REST1 (R1) and black asterisks denote significant differences from REST 2 (R2). REST3 condition is abbreviated as R3 and, thalamus has been abbreviated as “Thal”.
Following the dynamic adaptation task (REST3-REST1), there were increases in rsFC within the trained cortico-thalamic-cerebellar network, between the left cortical regions to both the thalamus and the cerebellum (model pFWE ≤ 0.05: L-M1 to L-Thalamus: p = 0.0032, effect size: 0.678; L-S1 to R-CB6: p = 0.0006, effect size: 0.711; model punc ≤ 0.01: L-M1 to R-CB6: p = 0.001, effect size: 0.615), as well as persistent decreases in connectivity between the sensorimotor cortices (Model pFWE ≤ 0.05: L-PM to R-M1: p = 0.0009, effect size: −0.579; Model punc ≤ 0.01: L-PM to R-S1: p = 0.002, effect size: −0.504; L-M1 to R-M1: p = 0.001, effect size: −0.541). In a direct comparison of REST3-REST2, only increases between the sensorimotor cortex and cerebellum approached significance (Model pFWE ≤ 0.05: L-S1 to R-CB6, p = 0.096, effect size: 0.338; Model punc ≤ 0.01: L-M1 to R-CB6, p = 0.039, effect size: 0.376).
For group 2, the model again identified decreases in rsFC between the right and left motor cortices following the motor execution task at the corrected threshold (|d| ≥ 0.324: L-M1 to R-M1: effect size: −0.575). Following the random perturbation task, there were no changes in rsFC within the trained cortico-thalamic-cerebellar network, but persistent, larger decreases in interhemispheric rsFC between the sensorimotor cortices (|d| ≥ 0.324: L-M1 to R-M1: effect size: −0.963, L-PM to R-M1: effect size: −0.566).
These results suggest that interhemispheric decreases observed between the sensorimotor cortices (L-PM to R-M1, L-M1 to R-M1) are associated with nonadapatation-specific features of task performance, as changes were observed in groups 1 and 2 following the motor execution task, and following both dynamic perturbation conditions. Instead, changes within the cortico-cerebellar pathway may be due to adaptation, as effects observed in group 1 reached significance only in REST3-REST1, but not following motor execution (L-M1 to R-CB6, L-S1 to R-CB6), and exhibited greater change in REST3-REST 2 compared with REST2-REST1 (L-M1 to R-CB6). Moreover, no changes within this network were observed following exposure to the unlearnable dynamic perturbation task.
Seed-to-voxel analysis
Our seed-to-voxel analysis identified significant positive baseline rsFC between the left sensorimotor ROIs (L-M1, L-PM, L-S1) and bilateral motor areas, the frontal cortex, the basal ganglia and the cerebellum (regions V–VI, VIII). For the REST2-REST1 contrast, no significant changes were returned for the left M1 or left PM ROIs. The left SI ROI had significant changes in rsFC to the visual cortex. For the REST3-REST1 contrast, the left M1 ROI showed significant decreases to the right primary motor and sensory motor cortices (BA4, BA1, BA3), increases to the left thalamus and left caudate, and increases to the right cerebellum lobules 6 and 8 (Fig. 6). The left S1 ROI showed significant decreases to the right superior parietal lobe (SPL7, IPL), and significant increases to the left thalamus and caudate, and the right cerebellum lobule 6. Both the left M1 and S1 ROIs showed significant increases to the visual cortex. The left PM ROI showed significant decreases to the right primary and sensory motor cortex (BA4, BA3, BA1). Brain regions identified in this analysis largely overlap with those included in our restricted network analysis, and confirm the apropriateness of the ROIs included in our restricted network analysis based on task-related activations.
Group level results of each individual left motor seed (M1, PM, S1) to whole brain resting state functional connectivity following dynamic adaptation task performance (REST3-REST1). All statistical parametric maps are overlaid on axial slices of the standard Montreal Neurologic Institute 152 template, with reported z coordinates in mm. For each ROI, the seed-voxel level T-maps were thresholded at p-voxel-uncorrected < 0.001, p-cluster-FDR < 0.05, resulting in |T(29)| > 3.65, Kcluster≥ 134.
Interestingly, this analysis showed no significant effects in the REST2-REST1 condition, in line with results from previous rsFC analyses that used similar methodology (Albert et al., 2009, Sami and Miall, 2013). For the REST3-REST1 comparison, the seed-to-voxel analysis results largely supported effects observed in the restricted network analysis, and further implicate the cortico-thalamic-cerebellar network as significantly modulated following adapation.
Associations between rsFC and behavior
We hypothesized that changes within the sensorimotor network and the cortico-thalamic- cerebellar network would explain variance in adaptation behavior and its retention. To test this hypothesis, we performed a linear regression analysis between our primary behavioral metrics (relative error reduction, magnitude of adaptation index, and magnitude of after effects) and rsFC, using a model with baseline rsFC (REST1) and change in rsFC with each task from baseline (REST2-REST1 and REST3-REST1) as explanatory variables. For all associations to day 2 behavior, we included day 1 behavior as an additional explanatory variable to isolate effects associated with change in behavior between days, indicative of retention. Stepwise backwards removal of nonsignificant terms was used to identify rsFC conditions that showed the greatest association to behavior, while controlling for possible contributions from all other rsFC conditions, as no terms that contributed to model performance were removed. Subsequent results are reported based on the reduced models resulting from our analysis pipeline. Post hoc analysis to independent measures of MA and AERS were used to determine whether associations to our primary behavioral metrics were driven by adaptation or alternate error reduction processes. A network level breakdown of these results by resting state condition is provided below (Table 1; Table 2; Fig. 7).
Results of the linear regression analysis between rsFC and behavior (RER: relative error reduction, MAE: magnitude of after effects, MAI: magnitude of adaptation index) measured on day 1 and day 2
Results for the post hoc analysis between rsFC associated to adaptation behavior (RER, MAE, MAI) to independent measures of adaptation (MA) and alternate error reduction strategies (AERS)
Association between rsFC and adaptation behavior. A, Brain schematic showing the location of each network of interest. B, Association between rsFC and the primary outcomes of day 1 behavior (dashed lines) and between rsFC and the primary outcomes of day 2 behavior (solid lines). For space reasons, magnitude of adaptation index has been abbreviated as MAI, magnitude of after effects as MAE, and relative error reduction as RER. Edges with significant association (pFWE < 0.05) as determined in the reduced model analysis are depicted as opaque lines, while edges with associations at pUNC < 0.01 are shown as faded lines with 50% opacity. For both thresholds, only edges with model terms with
Associations between behavior and rsFC within the cortico-thalamic-cerebellar network
Linear regression analysis identified multiple associations between baseline rsFC measured within the cortico-cerebellar network and gains in task performance measured on day 2 (Table 1). Within the cortico-cerebellar network, the Left M1-Right CB6 edge returned a positive association between increases in magnitude of adaptation index on day 2 and greater baseline rsFC (pMODEL = 0.0030, pREST1 = 0.0027), that post hoc analysis attributed to adaptation-specific processes (D2 MA ∼ 0.28 + 0.41(REST1) + 0.31(D1 MA), pMODEL = 0.024, pREST1 = 0.007; Table 2). In the Left S1-Right CB6 edge, the linear regression identified a positive association between increases in magnitude of adaptation index and rsFC measured at baseline and across the motor execution task [D2 MAI ∼ 0.15 + 0.47(REST1) + 0.69(REST2-REST1)+ 0.53(D1 MAI); pMODEL = 0.0013, pREST1 = 0.0117, pREST2-REST1 = 0.0014; Table 1]. Post hoc analysis suggested that these changes were more associated with adaptation than alternate error reduction strategies, but no significant associations were identified (Table 1). Conversely, lower baseline rsFC between the left PM and right CB8 was significantly associated with greater relative error reduction [D2 RER, 0.34–0.42(REST1) + 0.49(D1 RER), pMODEL = 0.0008, pREST1 = 0.002], that was associated with a greater reliance on alternate error reduction strategies on day 2 compared with day 1 [D2 AERS ∼ 0.32–0.47(REST1) + 0.24(D1 AERS), pMODEL = 0.022, pREST1 = 0.01; Table 2]. Together, these results suggest that greater pre-task rsFC (primarily at baseline) between cortico-cerebellar nodes contributes to greater consolidation of adaptation learning, while lower rsFC within this network may lead to the utilization of alternate control strategies.
The linear regression analysis returned the largest set of associations between change in rsFC following the dynamic adaptation task and behavior. In the cortico-cerebellar network, gains in adaptation on day 2 were positively associated with increases in rsFC measured following the dynamic adaptation task. At the pFWE < 0.05 level, the Left M1-Right CB8 edge was positively associated with greater relative error reduction [D2 RER∼ 0.2 + 0.42(REST3-REST1) +0.64(D1 RER), pMODEL = 0 0.0005, pREST3-REST1 = 0.0012]. At the pUNC < 0.01 level, the Left PM-Right CB8 edge showed a similar relationship to increases in magnitude of after effects [D2 MAE∼ 14.32 + 17.1(REST3-REST1) +0.19(D1 MAE), pMODEL = 0.0083, pREST3-REST1 = 0.0037]. Post hoc analysis attributed effects in both edges to adaptation specific processes (Fig. 8; Table 2), with no significant association to alternate error reduction strategies [Left M1-Right CB8, D2 MA ∼ 0.34 + 0.42(REST3-REST1) + 0.19(D1), pMODEL = 0.016, pREST3-REST1 = 0.0098; Left PM-Right CB8, D2 MA ∼ 0.32 + 0.43(REST3-REST1) + 0.28(D1), pMODEL = 0.008, pREST3-REST1 = 0.0047]. These results suggest that increases in rsFC between cortical motor regions and the posterior cerebellum following dynamic adaptation reflect consolidation of adaptation, that contributes to increases in adaptation on day 2 over day 1 behavior.
Relationship between day 2 behavior and changes in resting state functional connectivity following the dynamic adaptation task (REST3-REST1) in the cortico-cerebellar pathway, as identified by linear regression analysis. Magnitude of after effects is abbreviated as MAE, and relative error reduction is abbreviated as RER. The x-axis reports measured change in rsFC, while the y-axis of each subplot reports the residual portion of day 2 behavior that is not explained by other model terms [i.e., day 2
Within the cerebellum (Right CB6-Right CB8), there was a positive assocation (pUNC < 0.01) between increases in rsFC and greater magnitude of adaptation index measured on day 1 [D1 MAI ∼ 0.35 + 0.48(R3-1), p = 0.0089], that post hoc analysis attributed to adaptation specifc processes (D1 MA ∼ 0.25 + 0.43(R3-1), p = 0.016). There was additionally an association (pUNC < 0.01) between gains in magnitude of adaptation index on day 2 and both baseline rsFC and change in rsFC following the dynamic adaptation task (Table 1). However, this result was confounded by the significant association between two explanatory variables included in the model (day 1 MAI and change in rsFC across the REST3−REST1 condition, p = 0.0089). To control for these effects, we performed a post hoc analysis of the association between change in magnitude of adaptation index and rsFC. Results showed that neither baseline rsFC nor increases in rsFC were significantly associated to changes in behavior [ΔMAI = −0.10 + 0.37(REST1), p = 0.045; ΔMAI =0.03 + 0.09(REST3−REST1), p = 0.719]. As such, associations between rsFC in this edge and day 2 gains in adaptation were considered to be spurious. These results are in line with previous works that show the cerebellum is engaged in processing kinematic errors that contribute to adaptation, which may explain the association to day 1 adaptation measures, but may not contribute to motor memory formation, indicated by the lack of association to retention (Galea et al., 2011; Herzfeld et al., 2014).
Association between behavior and rsFC within the cortical sensorimotor network
Linear regresson analysis returned no significant associations between behavior and baseline rsFC within the cortical sensorimotor network (pFWE < 0.05). At the pUNC < 0.01 level, a positive association between relative error reduction and baseline rsFC was identified within the trained left sensorimotor cortex [Left M1-Left S1, D1 RER ∼ 0.51 + 0.24(REST1), p < 0.0088; Table 1], that post hoc analysis attributed to alternate error reduction strategies [D1 AERS ∼ 3.62 + 2.78(REST1), p < 0.0097; Table 2], with no association to adaptation.
There were no significant associations between changes in rsFC following the motor execution task and behavior at either threshold (pFWE < 0.05, pUNC < 0.01).
Linear regresson analysis returned significant (pFWE < 0.05) associations between changes in rsFC following the dynamic adaptation task and day 1 behavior. Decreases in interhemispheric rsFC between cortical sensorimotor areas were associated with greater relative error reduction [Left S1-Right M1, day 1 RER ∼ 0.59 − 0.34(REST3-REST1), p = 0.0007], and lower magnitude of after effects [Left PM-Right M1, D1 MAE ∼13.44 + 9.85(REST1) −10.68(REST2-REST1) + 19.82 (REST3-REST1), pMODEL = 0.0033, pREST3-REST1 = 0.0006; Fig. 9]. In line with these results, post hoc analysis showed that decreases in rsFC across both edges were preferentially associated with greater alternate error reduction strategies [Left S1-Right M1, D1 AERS ∼ 0.28–0.42(REST3-REST1), p < 0.0007; Left PM-Right M1, D1 AERS ∼ 0.29 − 0.27(REST3-REST1), p < 0.029].
Relationship between day 1 behavior and changes in resting state functional connectivity following the dynamic adaptation task (REST3-REST1) in the cortical sensorimotor network, as identified by linear regression analysis. Magnitude of after effects is abbreviated as MAE, and relative error reduction is abbreviated as RER. The x-axis reports the measured change in rsFC, while the y-axis of reports the residual portion of day 1 behavior that is not explained by other model terms [i.e., day 1 MÂE = day 1 MAE –
These results suggest that decreases in interhemipheric functional connectivity between the trained and untrained motor cortex are associated with independent motor control processes that occur in parallel to adaptation, but contribute to behavior via alternate error reduction strategies.
Supplementary analysis showed the same pattern of results at similar levels of significance for linear regression analysis using alternate definitions of subject-level behavior based on averages. Additionally, using a model constructed with the term REST3-REST2 rather than REST3-REST1 (i.e., changes following adaptation that were greater than those observed following motor execution), we observed highly overlapping patterns of associations within the cortico-cerebellar network (significant positive associations between postadaptation change in Left M1–Right CB8 and day 2 gains in relative error reduction, and in Left PM–Right CB8 and day 2 gains in magnitude of after effects), indicating that these effects are attributable to consolidation of adaptation (Extended Data Table 1-1).
Table 1-1
: Results of the linear regression analysis using the alternate model construction (terms R1, R2-R2, R3-R2) between rsFC and behavior (RER: relative error reduction, MAE: magnitude of after effects, MAI: magnitude of adaptation index) measured on day 1 and day 2 for the edges identified in our main analysis (terms R1, R2-R2, R3-R1). The table shows parameter estimates and their significance for the most reduced model returned by our analysis pipeline. Model significance is bolded for pFWE < 0.05, and non-bolded for pUNC < 0.01. Individual model terms are bolded for pα,corrected < 0.05. Model significance highlighted in yellow identify edges whose significance falls below our threshold in the main paper, although all remain below p = 0.05. Download Table 1-1, DOCX file.
The results from our linear regression analysis partially support our second hypothesis. Only changes in rsFC within the cortico-cerebellar network were associated with behavior on day 1 and retention of training on day 2, indicative of memory formation. Post hoc analysis showed that variance in rsFC measured in this network was specific to dynamic adaptation. Instead, changes in rsFC within the sensorimotor network were associated to day 1 behavior measured in the dynamic adaptation task but not retention of training on day 2. Post hoc analysis determined that variance in rsFC measured in this network was associated with alternate error reduction processes and not with adaptation.
Discussion
We aimed to identify the neural processes associated with motor memory formation following dynamic adaptation of the wrist. Using an fMRI-compatible robot, the MR-SoftWrist, participants performed wrist pointing during stable and unstable dynamic perturbations. We quantified changes in resting-state functional connectivity (rsFC) within task-localized brain regions in the cortical sensorimotor and cortico-thalamic-cerebellar network of the trained right wrist immediately following task execution and adaptation. To determine associations between rsFC measured in these networks, dynamic adaptation, and motor memory formation, we performed a linear regression between rsFC and day 1 and day 2 behavior. Post hoc analyses of rsFC to metrics of adaptation identified via variance decomposition, and comparison to changes in rsFC measured following unlearnable dynamic perturbations were used to identify networks specific to adaptation. Critically, our work provides new insight on how rsFC in the cortical sensorimotor and cortico-thalamic-cerebellar networks is modulated immediately after exposure to dynamic perturbations, and how these modulations are associated to adaptation-specific behavior and its retention.
Behavioral evidence of adaptation and retention
Group-level behavioral analysis showed evidence of adaptation, including significant after effects and increases in adaptation index. Between days, there was evidence of retention, including significant decreases in initial errors, consistent with recall processes (Mawase et al., 2017) and significantly lower final errors. There was no significant change in relearning (Joiner and Smith, 2008; Mawase et al., 2017), possibly due to behavioral washout from the performance of the resistive force task between days (here not analyzed). However, similar tasks with nonconflicting task dynamics or performed at a latency of >10 min following training have been shown to cause limited-to-no interference of retention for dynamic adaptation (Bock et al., 2001; Miall et al., 2004; Krakauer et al., 2005; Criscimagna-Hemminger and Shadmehr, 2008). Therefore, we consider these results confirm that significant retention occurred between days, and that participants formed motor memories following training that may influence rsFC.
To investigate the association between rsFC and behavior, we defined subject specific measures of magnitude of adaptation index, after effects and relative error reduction. Using these metrics, we performed a variance decomposition analysis to create independent measures of adaptation-specific behavior (MA) and alternate error reduction strategies (AERS) to distinguish between variance attributable to both processes in our primary behavioral metrics. For relative error reduction, driven by adaptation and alternate error reduction strategies, the AERS metric accounted for a greater percent of the variance in behavior than MA (AERS day 1–day 2: 72–88%; MA: 26–12%). Moreover, the magnitude of adaptation achieved in our study (mean ± SEM: 0.355 ± 0.0257), while significant, was lower than magnitudes reported in other paradigms following a similar number of trials (typically magnitude > 0.5; Donchin and Shadmehr, 2002; Smith et al., 2006; Lee and Schweighofer, 2009). These results suggest that alternate error reduction strategies (most likely co-contraction) played a significant role in task execution, and motivate differentiation between effects attributable to each control process in measures of rsFC.
Changes in rsFC following task performance
We hypothesized that there would be significant change in rsFC following the dynamic adaptation task, but no change following the motor execution task, as no significant learning should occur (Albert et al., 2009). However, across all tasks we identified significant decreases in interhemispheric functional connectivity between the trained and untrained sensorimotor cortices. In both groups, interhemispheric decreases following the dynamic perturbation task (REST3-REST1) were more widespread compared with motor execution alone (REST2-REST1), and reached significance in the supplemental voxel-based analysis only in the REST3-REST1 condition. These changes in rsFC may be result from increased excitation of the trained motor cortex, previously observed in unilateral force production tasks and repetitive TMS (Peltier et al., 2005). Unilateral excitation of the motor cortex has been shown to decrease interhemispheric rsFC by increasing interhemispheric inhibition from the stimulated cortex to the contralateral hemisphere (Watanabe et al., 2014). To our knowledge, these results have not been measured at longer timescales following motor tasks (Albert et al., 2009; Sami and Miall, 2013), and may represent transient changes in rsFC following task execution that appear to increase with increased force production.
Within the cortico-thalamic-cerebellar network, changes in rsFC between the primary motor cortex and the left thalamus were observed following the motor execution and the dynamic adaptation task, but showed no difference between REST3-REST2. Similar to interhemispheric decreases in the sensorimotor cortex, changes in this edge may reflect transient nonspecific task-effects due to increased activity in the trained motor cortex, which has significant projections to the thalamus (Shadmehr and Holcomb, 1997). Instead, increases in rsFC between the left cortical sensorimotor regions to the right cerebellum were observed in both the mixed model and seed-to-voxel analyses following the dynamic adaptation task (REST3-REST1), but were not present following the motor execution task (REST2-REST1). Although direct comparison between REST3 and REST2 showed nonsignificant increases when applying a Bonferroni correction, this contrast may be affected by subthreshold changes in REST2 associated with adaptation-based learning of the dynamics of task execution with the device (i.e., recalibration of a known behavior in a new environment). Moreover, the same changes in rsFC REST3-REST1 were not detected in group 2, exposed to a nonlearnable dynamic perturbation. Consequently, the increases in rsFC in the cortico-cerebellar network appear to be primarily driven by dynamic adaptation. These results are in line with previous works that show the cortico-thalamic-cerebellar network is integral to dynamic adaptation (Chen et al., 2006; Criscimagna-Hemminger et al., 2010; Izawa et al., 2012), and exhibits changes in rsFC following adaptation (Vahdat et al., 2011; Mawase et al., 2017; Babadi et al., 2021).
Associations between rsFC and behavior
We hypothesized that changes in both networks would explain variance in dynamic adaptation behavior and its retention. However, our analysis showed that rsFC measured in the cortical sensorimotor and cortico-thalamic-cerebellar networks are associated with different aspects of dynamic motor control.
RsFC between cortical sensorimotor areas and the cerebellum was associated with motor memory formation processes of adaptation. Greater baseline and pre-task rsFC (REST1, REST2-REST1) between the left sensorimotor cortex (M1, S1) and right anterior cerebellum (CB6) were associated with greater gains in adaptation, suggestive of a priming effect. Following the dynamic adaptation task (REST3-REST1), increases in rsFC between regions in the left sensorimotor network (M1, PM) and right posterior cerebellum (CB8) were positively associated with greater error reduction and after effects on day 2; gains that were attributed to adaptation-specific processes in post hoc analysis. Conversely, lower baseline rsFC between the left premotor cortex and the posterior cerebellum was associated with greater use of alternate error reduction strategies on day 2. As such, individuals with lower baseline rsFC in this edge may have been less likely to form adaptation specific memories and rely more on alternate error reduction strategies on day 2. Instead, individuals with greater gains in adaptive behavior had greater baseline rsFC in this network, that increased following task performance during early consolidation, implicating this pathway in adaptation and its retention.
Both the group-level change in rsFC and linear regression analyses indicate that increases in rsFC within the cortico-cerebellar network are associated with consolidation of dynamic adaptation. However, previous rsFC studies measured at longer timescales (>30 min) after adaptation report a different pattern of change, namely an inhibitory relationship (decreased rsFC) between the left motor cortex and right cerebellum (Vahdat et al., 2011; Mawase et al., 2017). TMS studies conducted during visuomotor adaptation have shown a faciliatory relationship between the cerebellum and the motor cortex early in the adaptation process, when errors are large, and an inhibitory pattern that returns following acquisition of the learned motor pattern when errors are small (Schlerf et al., 2012). Given the immediate acquisition of our resting state scans, our effects may capture this faciliatory (increased rsFC) relationship between the cerebellum and the motor cortex that has not been observed at longer timescales. Moreover, previous studies that report an inhibitory relationship achieved a greater degree of error reduction at the group level that may explain the inhibitory effects observed (Vahdat et al., 2011).
Finally, rsFC within the cortical sensorimotor network was associated with day 1 behavior only. Baseline rsFC within the left sensorimotor cortex was positively associated with greater relative error reduction, and interhemispheric decreases in rsFC following the dynamic adaptation task were associated with greater relative error reduction and lower magnitude of after effects, attributable to alternate error reduction processes. Following the unlearnable perturbation task, known to elicit impedance control (co-contraction) for error reduction, similar interhemispheric decreases were observed. Recently, increases in co-contraction early in a dynamic adaptation task have been associated to decreases in rsFC between the left M1 and right sensorimotor areas (Babadi et al., 2021). Together, these results suggest that the sensorimotor network contributes to error reduction in dynamic adaptation tasks by independent dynamic control strategies.
In summary, this study supports the hypothesis that changes in rsFC measured immediately after adaptation are reflective of motor memory formation processes, and provides evidence that different brain networks, the cortico-cerebellar network and the cortical sensorimotor network, are involved in different motor control processes that improve performance during dynamic adaptation of wrist movements.
Footnotes
This work was supported by the National Science Foundation Grant 1943712 and the American Heart Association Award SDG 17SDG33690002.
The authors declare no competing financial interests.
- Correspondence should be addressed to Fabrizio Sergi at fabs{at}udel.edu
