Use-dependent plasticity explains aftereffects in visually guided locomotor learning of a novel step length asymmetry

Abstract

Studies of upper extremity reaching show that use-dependent plasticity, or learning from repetition, plays an important role in shaping motor behaviors. Yet the impact of repetition on locomotor learning is unclear, despite the fact that gait is developed and practiced over millions of repetitions. To test whether repetition alone can induce storage of a novel walking pattern, we instructed two groups of young healthy subjects to learn an asymmetric walking pattern through two distinct learning paradigms. The first group learned a new pattern through an established visual distortion paradigm, which provided both sensory prediction error and repetition of movement patterns to induce walking aftereffects, and the second received veridical feedback with a target change, which provided only repetition (use-dependent plasticity) to induce aftereffects. When feedback was removed, both groups demonstrated aftereffects in the primary outcome, step asymmetry index. Surprisingly, despite the different task demands, both groups produced similar aftereffect magnitudes, which also had similar rates of decay, suggesting that the addition of sensory prediction errors did not improve storage of learning beyond that induced by the use-dependent process alone. To further characterize the use-dependent process, we conducted a second experiment to quantify aftereffect size in a third group who practiced double the asymmetry magnitude. This new group showed a proportionately greater magnitude of the use-dependent aftereffect. Together, these findings show that the primary driver of storage of a new step length asymmetry during visually guided locomotor learning is repetition, not sensory prediction error, and this effect scales with the learning magnitude.

NEW & NOTEWORTHY Use-dependent plasticity, or learning from repetition, is an important process for upper extremity reaching tasks, but its contribution to walking is not well established. Here, we demonstrate the existence of a dose-dependent, use-dependent process during visually guided treadmill walking. We also show that sensory prediction errors, previously thought to drive aftereffects in similar locomotor learning paradigms, do not appear to play a significant role in visually driven learning of a novel step asymmetry during treadmill walking.

INTRODUCTION

Human locomotor patterns must be routinely adjusted for both short- and long-term needs, for example, to transition from walking on concrete to sand or to cope with a chronic, painful musculoskeletal condition. It is therefore critical that walking be responsive and adaptable to environmental constraints and behavioral goals.

There are several mechanisms of motor learning that enable the acquisition and adjustment of new movement patterns, including both implicit and explicit learning processes (French et al. 2018; Mazzoni and Krakauer 2006; Reisman et al. 2005). Perhaps the most well-studied form of implicit motor learning is sensorimotor adaptation (SMA), in which a systematic perturbation to a movement is delivered, creating a sensory prediction error (Tseng et al. 2007) or a mismatch between the predicted and actual sensory feedback. This error signal drives changes to the feedforward movement plan to reduce error size on subsequent trials (Krakauer et al. 1999; Martin et al. 1996; Morehead et al. 2017; Reisman et al. 2005; Shadmehr and Mussa-Ivaldi 1994). Another form of learning is use-dependent plasticity (UDP), which is characterized by the development of movement biases favoring the direction of previously repeated movements (Diedrichsen et al. 2010; Verstynen and Sabes 2011). Both SMA and UDP are implicit processes, but, unlike SMA, UDP is not driven by error. Rather, it is believed to be a form of Hebbian learning generated purely by repetition of a specific pattern of movement kinematics (Classen et al. 1998; Orban de Xivry et al. 2011). A golfer who repeats several practice swings before hitting the ball is attempting to take advantage of this process.

Learning of novel walking patterns in humans is often studied with SMA paradigms such as split-belt adaptation (Iturralde and Torres-Oviedo 2019; Lam et al. 2006; Mawase et al. 2013; Morton and Bastian 2006; Reisman et al. 2005). Although neuro-rehabilitation targeting the learning or relearning of locomotor patterns relies heavily on repetition-based interventions rather than learning through sensory prediction errors (Hornby et al. 2015, 2016, 2020), few studies have examined mechanisms of UDP-based locomotor learning. Some locomotor learning studies have provided evidence for the contribution of repetition to savings and aftereffects (Huynh et al. 2014; Long et al. 2016; Malone et al. 2011; Ochoa et al. 2017; Roemmich and Bastian 2015; Yen et al. 2015), but none has directly compared effects of SMA- versus UDP-based learning of locomotor patterns.

In recent years, several groups have developed locomotor learning paradigms using real-time visual feedback to induce changes in gait while walking on a treadmill (French et al. 2018; Hussain et al. 2013; Kim and Krebs 2012; Kim and Mugisha 2014). We have shown that even though these paradigms appear to provide sensory prediction errors during learning, the newly learned walking patterns are acquired, in large part, because of participants’ development of explicit strategies (French et al. 2018; Hussain et al. 2013). However, aftereffects observed when participants are told to return to walking “normally” without visual feedback indicate that at least some of the learning occurs implicitly (French et al. 2018). A weakness of these paradigms is their inability to disentangle whether the implicit learning is derived from sensory prediction errors (i.e., SMA) or repetition of movement kinematics (i.e., UDP).

Therefore, in the present study (experiment 1), we set out to tackle these two questions: 1) Can repetition, and thus UDP, be utilized to induce step asymmetry aftereffects following visually driven locomotor learning on a treadmill? and 2) Is the aftereffect previously observed after visually driven locomotor learning induced by SMA, UDP, or some combination of the two? We hypothesized that both SMA and UDP were involved in the implicit component of learning, such that providing repetition of movement kinematics alone would be sufficient to induce aftereffects, but that providing both sensory prediction errors and repetition would result in significantly greater aftereffects. Based on the results of experiment 1, we conducted a second experiment to further examine the dose-response relationship between the magnitude of learned walking asymmetry and the size of the aftereffect induced by UDP.

METHODS

Participants

A total of 36 young, healthy individuals between the ages of 18 and 40 yr participated in this study. Subjects were excluded if they had any chronic or recent orthopedic, neurological, or psychiatric conditions or any other health condition that would affect their ability to walk or learn a new walking pattern. Subjects were also excluded if they had previously participated in a locomotor learning study. For experiment 1, 20 individuals were recruited and randomized into one of two groups, a group who learned from sensory prediction error and repeated the same pattern of movements (SPE+REP; n = 10) or a group who learned from veridical feedback (without a sensory prediction error) and repeated the same movement pattern (REP; n = 10). Based on our findings from experiment 1, an additional 16 individuals were later recruited to participate in a third group, repetition-dose response (REP-DR). All procedures were approved by the University of Delaware Institutional Review Board, and all subjects provided written informed consent.

Paradigm

Subjects participated in a single session of treadmill walking at their comfortable speed, defined as a “comfortable but brisk pace” between 1.0 and 1.2 m/s. A ceiling-mounted harness (providing no body weight support) was donned, and subjects were instructed to lightly hold onto a front handrail during walking for safety. During some portions of walking, real-time visual feedback was displayed (with software by The Motion Monitor Toolbox; Innovative Sports Training Inc., Chicago, IL) on a computer monitor (51 cm) located ~60 cm in front of the subject and at eye level (Fig. 1A). Subjects learned a novel step length asymmetry by responding to different forms of visual feedback of step lengths depending on the group assignment.

Fig. 1.

Experimental paradigm. A: experimental setup. Subjects walked on a treadmill with the belts tied while watching feedback of their step length during the Learning phase. The green bar represents the right step length, and the blue bar represents the left step length. B: learning schedule. Subjects walked in 3 phases: Baseline, Learning and Washout. Gray areas indicate phases during which no visual feedback was provided and subjects were told to “look forward and walk normally.” REP, repetition; REP-DR, repetition-dose response; SAI, step asymmetry index; SPE+REP, sensory prediction error and repetition. During Learning, subjects were taught to change their step length asymmetry based on the visual feedback. C and D: rendition of the visual feedback for the SPE+REP (C) and REP and REP-DR (D) groups during Learning. The solid bars represent the step length observed by the subjects on the screen, and the dotted bars represent the actual step length (which subjects could not see). There was also a horizontal line for each leg that was derived from each subject’s baseline step length and served as the target step length during Learning. A gain was introduced into the SPE+REP group’s feedback, causing a sensory prediction error (C), whereas an adjustment to the target line was introduced into the REP group’s feedback, resulting in no sensory prediction error (D).

Experimental timeline.

The walking paradigm was divided into 3 phases (Fig. 1B): Baseline (6 min), Learning (22 min), and Washout (22 min). Baseline was subdivided into Baseline 1, 2, and 3 (2 min each). In Baseline 1, subjects walked normally without visual feedback in order to acclimate to the treadmill and to obtain baseline step length measurements. In Baseline 2, subjects were oriented to veridical visual feedback through verbal instructions and a practice period of guided modification of step lengths so subjects could experience how changing step lengths was reflected in the visual feedback display. Finally, in Baseline 3, subjects were asked to return to walking normally without visual feedback. Learning was subdivided into Learning 1 (12 min), 2 (5 min), and 3 (5 min). Before the beginning of Learning 1, all subjects were told of the upcoming change in step length symmetry required and what they would need to do to be successful at the task. Learning 1 consisted of a gradual, stepwise introduction of a step length asymmetry during walking, induced through the visual feedback. The visual feedback promoted a 1% greater asymmetry every minute for the first 10 min and then was held constant at the full 11% asymmetry for an additional 2 min. During Learning 2 and Learning 3, subjects continued walking with the visual feedback promoting the full asymmetry. Just before the start of Learning 2 and Learning 3, 30-s “catch” trials were performed (Catch 1 and Catch 2), during which the visual feedback was removed and subjects were instructed to “look forward and walk normally.” After Learning 3 was completed, subjects performed a Washout phase, in which the visual feedback was removed and subjects were instructed to “look forward and walk normally,” exactly as instructed in Catch 1 and 2. Between phases, the treadmill was stopped for 30–60 s so instructions for the next phase could be provided. Also, all subjects were encouraged to stay focused on the task by looking forward at the screen.

Visual feedback for SPE+REP, REP, and REP-DR groups.

The visual display consisted of two different-colored bars that represented step lengths on the right and left legs. The bars grew vertically during the swing phase of each leg, providing a visual representation of that leg’s step length that was synchronized in real time with the subject’s stepping movement during swing phase. The height of each bar was frozen on the screen from the moment of heel strike until the next swing phase began for that leg. There was also a horizontal line for each leg that was derived from each subject’s baseline step length and served as the target step length during Learning. Subjects were instructed to hit the target line exactly with each step. The target line was set with a margin of error of ±2% (Fig. 1, C and D).

During the Learning phases, all groups were driven by the visual feedback to produce a novel step length asymmetry, but the visual feedback that caused this behavioral change differed based on group assignment. Subjects in the SPE+REP group were provided erroneous visual feedback during Learning (Fig. 1C). That is, a gain was added to the height of the vertical bars representing each step length. The left bar showed subjects taking 11% shorter steps than their true left step length, and the right bar showed subjects taking 11% longer steps than their true right step length. The horizontal target lines for each leg remained at their baseline step length. The erroneous visual feedback induced a sensory prediction error (i.e., a mismatch between predicted and actual visual feedback of step lengths). Subjects in the REP group were provided with veridical feedback of step lengths throughout Learning. In contrast to the SPE+REP group, step asymmetries were induced by shifting the horizontal target line by 11% (Fig. 1D) rather than manipulating the gain on visual feedback. To reach the target, individuals were required to increase their left step length by 11% and simultaneously decrease their right step length by 11%. Therefore, the REP group did not experience a sensory prediction error, as there was no mismatch between predicted and actual visual feedback; instead they were provided a visual cue to change their step lengths over the course of learning, resulting in repetitive practice of a new asymmetry. Both groups in experiment 1 (SMA+REP and REP) experienced the same step asymmetry magnitude adjustment in the Learning phase, the same initial degree of target error, and the same amount of practice time. Figure 1, C and D, depict the visual feedback seen at the beginning and end of Learning 2 (i.e., when the maximum step asymmetry is being learned). In experiment 2, a third group, REP-DR, experienced the exact same schedule, visual feedback, and instructions as the REP group, except that the total target shift was 22% instead of 11%. This was gradually induced with a 2% change of target location every minute during the first 10 min of Learning and then held constant for an additional 2 min at the end of Learning 1 (Fig. 1B).

Data Collection

Subjects walked on a dual-belt treadmill (with belts tied) instrumented with two force plates, one under each belt (Bertec, Columbus, OH). Kinematic data were collected with a Vicon MX40 motion capture system with eight cameras and Nexus software (Vicon Motion Systems, Inc., London, United Kingdom). Seven retroreflective markers were placed on bilateral 5th metatarsal heads, lateral malleoli, calcanei, and the left 1st metatarsal head and recorded at a frequency of 100 Hz. Ground reaction forces were recorded from the force plates at a frequency of 1,000 Hz and time-synchronized with the kinematic data in Nexus.

Data Analysis

Data were analyzed with custom-written MATLAB code (MathWorks, Natick, MA). Kinematic and kinetic data were low-pass filtered at 10 Hz with a fourth-order Butterworth filter. The force plate data were used to identify initial contact (heel strike) as the time when the force reading first increased over 20 N and liftoff (toe off) as the time when the force reading first decreased below 20 N for each step. Erroneous force plate events were removed and replaced with kinematic events (2.0% of events). Step lengths were calculated as the distance between the leading and trailing calcaneus markers at the time of heel strike.

The primary outcome measure was step length asymmetry during walking, as this was the gait feature targeted during Learning and because step asymmetry has been assessed in numerous previous locomotor learning paradigms (French et al. 2018; Kim and Krebs 2012; Leech et al. 2018; Reisman et al. 2005). Step length asymmetry was quantified with a step asymmetry index (SAI):

$$\text{SAI}\left(\text{\%}\right)=\frac{\left({\text{Step\hspace{0.17em}Length}}_{\text{LONG}}-{\text{Step\hspace{0.17em}Length}}_{\text{SHORT}}\right)}{\left({\text{Step\hspace{0.17em}Length}}_{\text{LONG}}+{\text{Step\hspace{0.17em}Length}}_{\text{SHORT}}\right)}\times 100$$

where LONG (left leg) and SHORT (right leg) refer to the legs that were driven to increase and decrease their step lengths, respectively. SAIs were analyzed at the following epochs of interest: Late Baseline (last 50 strides of Baseline 3), Early Catch 1 (first 10 strides), Early Catch 2 (first 10 strides), Late Learning (last 50 strides of Learning 3), and Early Washout (first 10 strides). Comparisons of SAIs during these epochs allowed us to test the main hypothesis related to aftereffect sizes between groups (SPE+REP vs. REP in experiment 1; REP vs. REP-DR in experiment 2).

We also compared the rate of decay between groups during Washout. Previous work in the upper extremity has suggested that aftereffects induced by recalibration of an internal model (i.e., SMA) decay more rapidly than aftereffects induced by repetition (i.e., UDP) (Diedrichsen et al. 2010). Here, decay rates served as a secondary outcome to assess potential differences between SPE+REP and REP groups. We quantified decay rate by regressing consecutive strides during the Washout such that strides “n” were regressed over strides “n + 1”; 1 minus the slope of this regression line served as the decay rate for each subject (Kitago et al. 2013). This measure of rate provides an estimate for the amount of step asymmetry forgotten from one stride to the next during Washout.

Statistical Analysis

Statistical analysis was performed in SPSS (Microsoft, Chicago, IL). We first compared demographic data among all three groups with one-way ANOVAs. For experiment 1, SAIs were compared between SPE+REP and REP groups with a 2 × 3 mixed-models ANOVA with repeated measures on the factor epoch (for SAI, epochs were Late Baseline, Late Learning, and Early Washout). We also assessed the emergence of aftereffects by comparing Late Baseline, Early Catch 1, Early Catch 2, and Early Washout between groups (2 × 4 mixed-models ANOVA). Finally, decay rates were compared between SPE+REP and REP groups with an independent Student’s t test. For experiment 2, SAIs were compared between REP and REP-DR groups in the same fashion as experiment 1. If group SAI values differed significantly at Late Baseline (Student’s t test), we used baseline-corrected SAI values for all measures and removed the Baseline epoch from the ANOVA. When any ANOVA was significant, post hoc comparisons were performed with Bonferroni corrected t tests for interaction effects and pairwise comparisons for main effects. The α value was set at 0.05 for all comparisons. Assumptions of normality and sphericity were confirmed with Shapiro–Wilk’s and Mauchly’s tests, respectively.

RESULTS

When comparing demographic data across the SPE+REP (8 women, 2 men), REP (8 women, 2 men), and REP-DR (9 women, 7 men) groups, there were no differences with respect to age (mean ± SD; SPE+REP: 21.8 ± 3.5 yr, REP: 20.5 ± 1.7 yr, REP-DR: 21.0 ± 2.3 yr; P = 0.48), treadmill speed (SPE+REP: 1.14 ± 0.07 m/s, REP: 1.13 ± 0.07 m/s, REP-DR: 1.11 ± 0.06 m/s; P = 0.50), or number of strides performed during Learning (SPE+REP: 1,166.2 ± 66.2 strides, REP: 1,157.9 ± 30.5 strides, REP-DR: 1,127.3 ± 63.6 strides; P = 0.57).

Experiment 1

Figure 2 shows the key findings from experiment 1. Group-averaged SAI data are shown on a stride-by-stride basis for Baseline and Learning phases and the initial Washout phase in Fig. 2A. Here it can be seen that the rate of gradual learning and total amount of asymmetry learned were similar between the SPE+REP and REP groups. Likewise, the aftereffects observed during Washout were also similar. Figure 2B shows group-averaged SAI values for the key epochs of Late Baseline, Late Learning, and Early Washout. The ANOVA revealed no main effect of group (F_1,18 = 0.154, P = 0.699, ${\eta }_{\text{p}}^2$ = 0.009). There was, however, a main effect of epoch (F_2,17 = 389.116, P < 0.001, ${\eta }_{\text{p}}^2$ = 0.979). Pairwise comparisons indicated that SAIs were greater at Late Learning compared with Late Baseline (P < 0.001), greater at Late Learning compared with Early Washout (P < 0.001), and greater at Early Washout compared with Late Baseline (P < 0.001). The difference between Late Baseline and Early Washout indicated the presence of a significant aftereffect induced in both SPE+REP and REP groups. No significant interaction effect was observed (F_2,17 = 0.435, P = 0.654, ${\eta }_{\text{p}}^2$ = 0.049), indicating that the two groups learned a similar amount and showed a similar aftereffect size.

Fig. 2.

Experiment 1 results. A: group mean ± 1 SE of step asymmetry indexes (SAIs) shown in bins of 5 strides for Baseline, Learning, and Early Washout phases and for the 2 Catch trials for sensory prediction error and repetition (SPE+REP) and repetition (REP) groups. Data for each phase are truncated to the length of the subject with the least number of strides for that phase. Gray areas indicate phases during which no visual feedback was provided. B: group mean ± 1 SE of SAIs for both groups during Late Baseline, Late Learning, and Early Washout epochs. A significant effect of epoch was observed, with differences between each phase and crucially between Late Baseline and Early Washout. Horizontal bars represent group means; open circles show each individual’s performance. C: group mean ± 1 SE of SAIs during Late Baseline, Early Catch 1 and 2, and Early Washout epochs. A significant effect of epoch was observed, with differences between each phase and Late Baseline and differences between Early Washout and Early Catch 1. Horizontal bars represent group means; open circles show each individual’s performance.

The assessment of the temporal emergence of aftereffects in SPE+REP and REP groups is illustrated in Fig. 2C, which shows group mean SAIs during Early Catch 1 and Early Catch 2 compared with Late Baseline and Early Washout epochs. The ANOVA showed no main effect of group (F_1,18 = 0.003, P = 0.957, ${\eta }_{\text{p}}^2$ < 0.001) but did show a significant effect of epoch (F_3,16 = 11.498, P < 0.001, ${\eta }_{\text{p}}^2$ = 0.683). Post hoc analyses revealed that all three aftereffects were significantly greater than Baseline (all P ≤ 0.001). In addition, Early Catch 1 was significantly less than Early Washout (P = 0.009). No significant interaction effect was observed (F_3,16 = 0.727, P = 0.551, ${\eta }_{\text{p}}^2$ = 0.120), indicating that the aftereffects emerged over a similar time frame for SPE+REP and REP groups.

Figure 3A shows group-averaged SAI data on a stride-by-stride basis for each group during the Washout phase. As stated above, the ANOVA comparing initial aftereffect sizes showed no difference between groups at Early Washout (Fig. 2B). However, we also assessed whether learning via SPE+REP versus REP could result in different aftereffect decay rates. Mean decay rates are displayed in Fig. 3B. The REP (0.782 ± 0.130) and SPE+REP (0.712 ± 0.150) decay rates were not significantly different (P = 0.28).

Fig. 3.

A: group mean ± 1 SE of step asymmetry indexes (SAIs) shown in bins of 5 strides for entire Washout phase for sensory prediction error and repetition (SPE+REP) and repetition (REP) groups. Data are truncated to the length of the subject with the least number of strides. B: group mean ± 1 SE for decay rates during the Washout phase, calculated as a proportion, between 0 and 1, of SAI forgotten from one stride to the next. Open circles show each individual’s performance.

Experiment 2

Results from experiment 1 strongly suggested that step asymmetry aftereffects in this treadmill-based visual locomotor learning paradigm were due solely to a use-dependent process. In experiment 2, we sought to characterize the dose-response relationship between the size of the asymmetry induced via use-dependent locomotor learning and the size of the aftereffect at Early Washout. To accomplish this, we compared a new REP-DR group to the REP group from experiment 1.

Figure 4 shows the key findings from experiment 2. Group-averaged SAI values are shown on a stride-by-stride basis for Baseline and Learning phases and the Washout phase in Fig. 4A. Figure 4B shows the results of the ANOVA for the epochs of interest. Because SAIs were different between groups at Late Baseline (P = 0.041), we baseline-corrected all SAIs and performed a 2 × 2 ANOVA comparing Late Learning and Early Washout between groups. There was a main effect of epoch (F_1,24 = 567.772, P < 0.001, ${\eta }_{\text{p}}^2$ = 0.959), a main effect of group (F_1,24 = 82.859, P < 0.001, ${\eta }_{\text{p}}^2$ = 0.775), and a significant group × epoch interaction effect (F_1,24 = 90.260, P < 0.001, ${\eta }_{\text{p}}^2$ = 0.790). Post hoc testing of the interaction effect revealed between-group differences at Late Learning (P < 0.001) and at Early Washout (P = 0.020), with the size of the aftereffect being proportional to the size of learned asymmetry.

Fig. 4.

Experiment 2 results. A: mean ± 1 SE of step asymmetry indexes (SAIs) shown in bins of 5 strides for Baseline (Bsl), Learning, and Washout phases and for the 2 Catch trials for repetition (REP) and repetition-dose response (REP-DR) groups. Data shown are baseline-corrected and truncated to the length of the subject with the least number of strides for that phase. Gray areas indicate phases during which no visual feedback was provided. B: group mean ± 1 SE of baseline-corrected SAIs for both groups during Late Learning and Early Washout epochs. A significant group × epoch interaction effect was observed, with post hoc testing showing significant between-group differences at both time points. Horizontal bars represent group means; open circles show each individual’s performance.

DISCUSSION

In the present study, we observed that repetition of a novel step asymmetry caused a use-dependent bias toward the practiced direction. To our knowledge, this is the first such demonstration in walking in which UDP was isolated from SMA, and a use-dependent bias was observed. In addition, the comparison between a group that received both repetition and sensory prediction errors (SPE+REP group) and a repetition-only group (REP group) revealed that the implicit aftereffects can be explained solely by the use-dependent bias in the repeated direction and not the recalibration of an internal model due to a sensory prediction error. Specifically, and counter to our hypothesis, the aftereffects experienced by the group who learned from a visual gain sensory prediction error as well as repetition (SPE+REP group) were not significantly different from the group who received repetition only and no sensory prediction error (REP group). The analyses of the temporal emergence of aftereffects during learning and the decay rates during washout also showed no differences between groups, further indicating that the two groups performed remarkably similarly. We argue that the most likely explanation for this is that both groups learned and stored the asymmetry through the same mechanism, use-dependent plasticity. This is an important departure from the assumption that sensory prediction error is the primary driver of aftereffects when learning new patterns of walking on the treadmill through distorted visual feedback (Cherry-Allen et al. 2018; Chunduru et al. 2019; French et al. 2018; Hussain et al. 2013; Kim et al. 2015; Kim and Krebs 2012; Kim and Mugisha 2014; Statton et al. 2016). In experiment 2, the REP-DR group demonstrated significantly greater aftereffects compared with the REP group, with the only difference between paradigms being the size of the learned asymmetry. Therefore, the use-dependent bias in step length asymmetry is directly affected by the size of the learned asymmetry.

To understand these results, it is important to consider some of the nuances of the present locomotor learning paradigm. First, although our primary outcome was step length asymmetry, the cues used to drive learning were applied to each leg separately. Separate ANOVAs of the right and left step lengths during Late Baseline, Late Learning, and Early Washout periods for SPE+REP and REP groups showed no main effects of group (both P > 0.61) or interaction effects (both P > 0.29) but a significant effect of epoch (both P < 0.001), just as was shown for step length asymmetry (Fig. 2B). This demonstrates that each individual leg’s behavior was also not different between groups.

Another important aspect of this paradigm is that subjects were provided considerable explicit instructions. We have previously shown that, with this approach, subjects use an explicit strategy to achieve most of the learning (French et al. 2018). This was confirmed here when both SPE+REP and REP groups experienced a large decrease in step asymmetry from the end of the Learning phases to the beginning of the Catch and Washout phases when they were instructed to walk normally (i.e., when instructed to remove any strategy they were using). This ability to voluntarily and immediately implement or inhibit a strategy is a hallmark of explicit learning processes, and the persistent aftereffect during these periods is evidence that an implicit process was also active throughout learning (Bond and Taylor 2015; Taylor and Ivry 2011).

We hypothesized that this implicit process was a combination of repetition and sensory prediction error. However, the present data suggest that sensory prediction errors in the SPE+REP group did not result in an updating of an internal model for step length asymmetry. Although several prior studies with similar paradigms have assumed that sensory prediction errors are an important component of visually driven locomotor learning (Cherry-Allen et al. 2018; Chunduru et al. 2019; French et al. 2018; Hussain et al. 2013; Kim et al. 2015; Kim and Krebs 2012; Kim and Mugisha 2014; Statton et al. 2016), the presence of a sensory prediction error was not definitively demonstrated (i.e., was not separated out from effects of repetition) in any of these works. Thus, our work raises a fundamental question of whether or not erroneous visual feedback of specific gait parameters, such as step lengths, during treadmill walking does in fact use sensory prediction errors to change internal models of feedforward locomotor control. A different study by Maeda et al. (2017) used a modeling approach to suggest that when subjects wear laterally deviating prism goggles, sensory prediction errors are responsible for the adaptation of foot placement during overground walking. If this is correct, then why, in the present paradigm, did sensory prediction errors not contribute to aftereffects? We know that sensorimotor adaptation requires a salient and robust sensory prediction error signal (Brudner et al. 2016; Wei and Körding 2009). We can only speculate, but one possibility is that providing visual feedback of step lengths as bar graphs on a monitor was too abstract or too unfamiliar to sufficiently generate a robust sensory prediction error. This is supported by the fact that humans almost never utilize spatial representations of the lower extremities to guide movements, which contrasts with the frequent use of spatial representations for hand movements such as using smart phones and computers. Similarly, the prism glasses used in Maeda et al. (2017) have the effect of laterally displacing the entire visual field, thus affecting all visual feedback, whereas in our paradigm only one specific aspect of visual feedback was altered (the representation of step lengths), which was perhaps insufficient either to drive a sensory prediction error or to drive a motor response to the sensory prediction error. A related explanation for our findings comes from the theory of credit assignment (Berniker and Kording 2008), suggesting that sensory prediction errors might be less likely to be utilized to update an internal model if they are assigned to a specific tool or to the external environment, rather than to the individual. Perhaps the error signal in this paradigm was assigned to the unique visual display and therefore not incorporated into the motor plan for walking on the treadmill without the visual display. Regardless of the reason, the present work shows that erroneous visual feedback of step lengths during treadmill walking likely does not induce aftereffects, or the storage of learning, from sensory prediction errors above and beyond that which is induced due solely by UDP. This possibility should be further explored in other similar locomotor learning paradigms in which learning is induced with visual representations of specific gait parameters.

Interestingly, in a clinical study comparing error augmentation training (analogous to sensory prediction error-based learning) to error minimization and conventional training with verbal feedback, a similar finding was reported in which individuals with chronic stroke were shown to improve step symmetry and gait speed just as much with verbal feedback alone as they were with approaches based on using sensory prediction error modulation (Lewek et al. 2018). Contrary to our findings, a study by Long et al. (2016) found no use-dependent bias after a visually guided, explicit change to step asymmetry. Compared with the present study, there were differences in walking speed, the number of practice repetitions, size of the asymmetry, and the dependent variables used. Although some factors such as number of repetitions and asymmetry size appear to be essential to the use-dependent process in the present study, it is unclear why no use-dependent aftereffect was observed in the Long et al. study. Other components such as walking speed should be investigated in future studies as potentially necessary criteria to activate the use-dependent process.

We have argued that the aftereffects present in the present study are due to use-dependent plasticity, but it is possible that another process such as reinforcement learning was responsible for these effects. We think this is not the case, largely because we did not provide any explicit rewarding feedback when the target was successfully hit. However, we cannot rule out that an intrinsic target reward was active when subjects perceived they successfully hit the target (Kim et al. 2019; Leow et al. 2020). We speculate that intrinsic reward was not a key factor here because although subjects gradually approached the target zone over the course of learning, they actually stepped within the (presumably rewarding) target zone very infrequently (the mean percentage of strides landing within the target zone during Late Learning 3 was ~17%). Furthermore, the extended amount of time for the aftereffects to decay back to baseline is consistent with a use-dependent process (Diedrichsen et al. 2010).

Because use-dependent plasticity has never before been exclusively investigated in locomotion from a mechanistic perspective, we aimed to determine additional aspects of this form of learning. First, the step asymmetry aftereffects are small. However, they represent ~24% of the total amount of asymmetry learned for both groups in experiment 1 (Initial Washout divided by Late Learning), which leaves ~76% of the asymmetry learned occurring through explicit, strategic mechanisms. Furthermore, the aftereffects are, on average, outside of the range of subjects’ baseline step symmetry variability (1 standard deviation), indicating a true change in behavior. We also found that the use-dependent aftereffect increases with greater practice time with differences between Baseline and Early Catch 1 and between Early Catch 1 and Early Washout. However, it is unclear whether the increase in aftereffect size from Early Catch 1 to Early Washout was due to the increased number of repetitions or the amount of variability in the practice. Recall that practice during the learning phase before Early Catch 1 was gradual and thus more variable than the steady-state learning phases. Therefore, we cannot separate two potentially crucial components of use-dependent plasticity: number of repetitions and variability. However, we did find that one critical component of UDP is the magnitude of the learned asymmetry. In experiment 2, we used a REP-DR group to assess whether the use-dependent aftereffect scaled to the learning size or not. It is possible that the aftereffects between the REP and REP-DR groups would not have been different, suggesting that the aftereffect is not scalable to the size of the asymmetry. On the contrary, by doubling the amount of asymmetry learned, the size of the aftereffect also increased.

Although it is well established that SMA is heavily reliant on the cerebellum, consistent with the theory of cerebellar function as a site of storage and updating of forward internal models for the motor system (Morton and Bastian 2006; Popa et al. 2013; Tseng et al. 2007; Wolpert et al. 1998), the mechanism behind and brain regions involved in UDP are less certain. In the upper extremity, UDP is suggested to be a form of Hebbian learning occurring in the motor cortex (Classen et al. 1998; Orban de Xivry et al. 2011). Although traditionally viewed as a subcortical or spinally driven behavior, it is now well recognized that, for locomotion, the motor cortex and posterior parietal cortex are also important for precise planning and execution of limb placement during visually guided walking (Lajoie et al. 2010; Krouchev and Drew 2013; Marple-Horvat and Criado 1999; Nordin et al. 2019). Therefore, a cortical site for aftereffects induced by UDP of human locomotion is plausible.

In conclusion, we have demonstrated for the first time that the primary driver of implicit storage of a newly learned step asymmetry induced by visually guided feedback during treadmill walking is a use-dependent bias, in which repeated prior movements influence future movements, and not sensory prediction errors associated with sensorimotor adaptation. This surprising result changes our understanding of how sensory prediction errors in visual representations influence locomotor learning paradigms of this sort and sheds new light on the importance of repetition and use-dependent plasticity during visually guided treadmill walking.

GRANTS

The work was supported by NIH Grants P20 GM-103446, S10 RR-028114, F31 NS-00806, and K12 HD-055931.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

J.M.W., H.E.K., M.A.F., D.S.R., and S.M.M. conceived and designed research; J.M.W. performed experiments; J.M.W. analyzed data; J.M.W., H.E.K., D.S.R., and S.M.M. interpreted results of experiments; J.M.W. prepared figures; J.M.W. and S.M.M. drafted manuscript; J.M.W., H.E.K., M.A.F., D.S.R., and S.M.M. edited and revised manuscript; J.M.W., H.E.K., M.A.F., D.S.R., and S.M.M. approved final version of manuscript.