The adaptation of muscle activity during split-belt walking reveals age-dependent decline of motor learning.

We are constantly adjusting our movements to changes in our body or environment through online feedback or 2 planned feedforward processes. It is thought that only feedforward motor commands are adapted in locomotion 3 and underlie the observed aftereffects upon experiencing a novel situation, such as split-belt walking with 4 the legs moving at different speeds. Feedback (a.k.a., reactive) commands are assumed to simply reﬂect the 5 walking condition by suddenly changing when the novel environment is introduced or removed. However, this 6 has been concluded based on kinematics (strongly inﬂuenced by environmental dynamics) and opposes others’ 7 suggestion that feedback processes are adapted according to the environment at hand. Thus, we ask if feedback 8 commands are adapted and if this adaptation is dissociated from the recalibration of feedforward ones through 9 the analysis of muscle activity. To this end, we characterized changes in the activity of 15 muscles during 10 and after split-belt walking with respect to regular walking. Importantly, we exploited the temporal dynamics 11 of feedforward and feedback processes to isolate feedback-generated motor commands and characterized its 12 adaptation after short and long exposures to the split condition requiring distinct recalibration of feedforward 13 commands. We found that feedback-mediated motor patterns were adapted. Speciﬁcally, subjects exhibited on 14 each leg distinct feedback-generated responses upon introduction of the split perturbation that were mirrored 15 between the legs (i.e., muscle activity of the right was observed on the left and vice versa) when the split 16 condition was removed. This mirroring was only observed after long exposure of the split condition and 17 contrasted the purely ipsilateral changes predicted from environment-dependent feedback activity. Both of 18 which indicated that the mirroring of feedback responses reﬂected a learning-dependent process. Further, older 19 individuals exhibited limited adaptation of feedback activity, but had equally adapted patterns at the end of the 20 split-belt walking compared to younger participants, suggesting that healthy aging only affects the adaptation of 21 feedback commands. Taken together our results indicate that corrective actions in walking are adapted and that 22 this adaptation is partially dissociated from planned actions, raising the possibility of distinct neural processes 23 underlying the adaptation of feedback- and feedforward-mediated motor control. Methods) of muscle activity aftereffects onto both kinematic aftereffects and 437 age indicated that both factors were signiﬁcant predictors (Pearson’s R 2 = 0 . 741, p age = 0 . 018 and p SLA = 438 4 . 2 × 10 − 4 , not shown). We interpret these results as supportive of two different learning mechanisms: one 439 that underlies the association between age and EMG aftereffects, and a separate one that is captured by the 440 association between EMG and kinematics aftereffects. In sum, we observed diminished mirroring of feedback 441 activity post-adaptation and smaller EMG aftereffects in older adults, suggesting that healthy aging has a 442 negative impact on learning process updating feedback-generated muscle activity.

Significance statement 25 We characterized the temporal evolution of muscle activity during a motor adaptation and deadaptation task. 26 We found that feedback-generated motor commands, which react to sudden transitions in the environment, are 27 adapted through prolonged exposure to new environmental conditions. Adapted responses are consistent with 28 subjects adopting the new environment as normal, and the extent of this adaptation is modulated by age. This 29 shows that aftereffects are at least partially a response to the sudden transition in the environment, and not 30 solely the continuation of a motor output optimized for different environmental conditions. Characterizing how 31 muscle activity changes throughout this task and how this predicts motor learning (aftereffects) can aid in the 32 design of individualized protocols to maximize the therapeutic effects of split-belt walking. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint interest as they may reflect distinct learning mechanisms (Wagner and Smith, 2008 It is particularly relevant to evaluate the adaptation of feedback and feedforward commands in middle aged 63 and older individuals because of its implications for movement rehabilitation. Namely, locomotor learning 64 induced by split-belt walking can be exploited to improve the gait of stroke survivors (Reisman et al., 2013; 65 Lewek et al., 2017), whose age mostly ranges between 45 and 80+ years old (Mozaffarian et al., 2016). While 66 there are studies reporting age-related decline in subjects' ability to adapt movements and the retention of those 67 changes (Wolpe et al., 2016;Sombric et al., 2017), it is unknown if healthy aging has a differential impact 68 on feedback and feedforward processes contributing to motor adaptation. Thus, we studied how healthy aging 69 affects the plasticity of muscle activity to gain insights on the learning mechanisms available to patients because 70 of their age.

72
In sum, we characterized the adaptation of muscle recordings from a broad set of lower-limb muscles during and 73 after split-belt walking. We used sudden changes in muscle activity upon transitions in the walking environment 74 (e.g., split-belts vs. tied-belts) to dissociate the contributions of feedforward and feedback processes to motor 75 adaptation. We found distinct feedback muscle activity following the same environment transition at different 76 adapted states of feedforward-generated motor commands, suggesting that feedback processes are also influenced 77 by the re-calibration of internal representations of walking. Importantly, we were able to predict the adaptation 78 of feedback-generated activity based on each individual's initial muscle activity during the adaptation period. 79 Interestingly, we observed that older subjects adapted less, revealing age-related limitations in sensorimotor 80 learning.   89 We assessed the adaptation and de-adaptation of muscle activity through the protocol illustrated in Figure   90 1A. The protocol consisted of treadmill walking in six different conditions presented in the following order: 91 Slow Walking (50 strides), Mid Walking (50 strides), Short Exposure (10 strides), Baseline Mid Walking (150 92 strides), Adaptation (900 strides, in three blocks of 300) and Post-Adaptation (600 strides, in two blocks of 93 300 or a single 600 block). Slow and mid walking were used to quantify speed-dependent modulation of 94 muscle activity during regular treadmill walking. Short Exposure and Adaptation were used to differentiate 95 learning-dependent changes in muscle activity that required multiple steps in the split environment from those 96 solely due to unexpected introduction or removal of the split perturbation. Baseline walking was used as a 97 reference for steady state walking on the treadmill prior to motor adaptation. Finally, the Post-Adaptation 98 condition was used to evaluate adaptation aftereffects in muscle activity following split-belt walking. The

99
Adaptation and Post-Adaptation conditions were designed to have enough strides to examine the evolution 100 of muscle activity from a transient to a steady state when the perturbation was either introduced or removed, 101 respectively. These conditions were divided into blocks to minimize fatigue and subjects were instructed not to 102 step between blocks to prevent deadaptation due to unrecorded steps. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint same speed relative to one another (i.e., 'tied' mode) ( Figure 1A). The treadmill was started and stopped at the 108 beginning and at the end of each condition and speeds did not change while each condition was ongoing. Belts 109 moved at one of three possible speeds throughout the experiment: a self-selected ('mid') walking speed, a 'slow' 110 speed of 66.6% of the mid speed, and a 'fast' speed of 133.3% of the mid speed. We used self-selected speeds to 111 ensure subjects from all ages could complete the entire protocol. The self-selected speed was obtained by first 112 averaging each subject's speed when walking over ground in a 50-meter hallway during 6 minutes (i.e., 6-min 113 walking test (Rikli and Jones, 1998)) and then subtracting 0.35m/s, which resulted in a comfortable walking 114 speed on a treadmill based on pilot data in older adults (> 65 yrs). Mid speed was 0.72 ± 0.26m/s (mean ± 115 standard deviation) across the population. Speed values used for all subjects during baseline walking are shown 116 in Table 2.

118
Safety measures were designed such that participants from older ages could complete the study. First, all 119 subjects wore a harness that only provided weight support in the event of falling but not during walking or 120 standing. Also, subjects were told a few seconds in advance that they could hold on to a handrail (directly 121 located in front of them) whenever a condition or block started or finished. Subjects were encouraged to let go 122 of the handrail as soon as they felt comfortable doing so to minimize the effect of external support on muscle 123 recordings. Finally, we monitored subjects' heart-rate continuously and blood-pressure during the rest breaks 124 to prevent over exertion in any of the participants. We collected electromyographic (EMG) signals, kinematics, and kinetic data to characterize subjects behavior.

128
Surface EMG signals from 15 muscles on each leg were recorded for all subjects (see Table 1 for full list and 129 abbreviations) at 2000Hz using a Delsys Trigno System (Delsys Inc., Natick, Massachussets). Signals were 130 highpass-filtered to remove undesired movement artifacts and then rectified. We used a 2nd order Butterworth 131 filter (dual-pass) with a cutoff frequency of 30Hz, which resulted in 80 dB/dec attenuation and zero-lag (Merletti 132 7 . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint and Parker, 2005). Kinematic data was collected at 100 Hz with a passive motion analysis system (Vicon Motion 133 Systems, Oxford, UK). Movements were recorded by placing reflective markers bilaterally on bony landmarks 134 at the ankle (i.e., lateral malleolus) and the hip (i.e., greater trochanter). Ground reaction forces were recorded 135 with an instrumented split-belt treadmill (Bertec Corporation, Columbus, Ohio) and sampled at 1000Hz. Forces 136 along the axis of gravity (Fz) were used to determine when the foot landed (i.e., heel-strike: Fz > 10N) or was 137 lifted off the ground (i.e., toe-off: Fz < 10N). EMG parameters: We characterized the changes in EMG during and after split-belt walking by first generating 141 muscle activity variables that discretized each muscle's activity over the gait cycle ( Figure 1B). Filtered EMG 142 activity was divided in sub-intervals of the gait cycle aligned to gait events to focus on changes in muscle 143 activity within the gait cycle, rather than on activation changes due to differences in timing of the gait cycle 144 across the distinct walking conditions (Dietz et al., 1994;Reisman et al., 2005). For example, in Figure 1B 145 we present sample EMG traces for baseline walking at medium speed (gray trace) and late adaptation (green 146 traces) for the leg walking slow (top panel) and the one walking fast (bottom panel). More specifically, we 147 divided the gait cycle of each leg into 6 intervals according to well defined gait phases (Perry and Burnfield,148 2010): first double-support (from ipsilateral heel-strike to contralateral toe-off), early single-stance (first half 149 of single stance), late single-stance (second half of single stance), second double support (from contralateral 150 heel-strike to ipsilateral toe-off), early swing (first half of swing phase) and late swing (second half of swing 151 phase). In order to achieve better temporal resolution, each of these 6 intervals where further divided into two 152 equal sub-intervals, yielding 12 intervals for each gait cycle. Note that the normalized gait cycles appear to 153 be the same duration because of our normalization procedure. However, each of these sub-intervals' duration 154 ranged between 75 and 120 ms. The precise timing for each sub-interval throughout the different epochs of the 155 study are presented in Table 3. Muscle activity amplitude was averaged in time for each of these sub-intervals, 156 for every stride and muscle resulting in 360 muscle activity variables per subject: 12 intervals x 30 muscles.

157
Sample of these variables for one muscle in one gait cycle for Baseline (B) (gray row) and late Adaptation (lA) 158 8 . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint (green row) are presented in the rows below the EMG trances in Figure 1B. 159 These EMG variables were normalized using the activity during Baseline walking as a reference. To this end, 160 we first computed the mean activity for each sub-interval across the last 40 strides (i.e., steady state) of Baseline 161 walking. Then each muscle's activity was scaled such that the least and most active phases for every muscle 162 during baseline had the value of 0 and 1, respectively. We specifically subtracted the mean value for the least 163 active phase in Baseline from the activity in all sub-intervals for all strides in all conditions and we subsequently 164 divided the shifted values by the mean baseline value for the most active phase. Following this normalization, 165 the units of EMG activity for each muscle of a given subject were represented with respect to the average 166 maximum value recorded in that muscle during Baseline walking of the same subject. This scaling allowed us 167 to aggregate subjects and compare effect sizes across muscles even when recorded EMG amplitudes were very 168 different because of sensor placement or underlying tissue properties ( Figure 1D).

170
Kinematic parameters: The adaptation of movements was characterized with step-length asymmetry, which is 171 a metric known to change during and after split-belt walking (Reisman et al., 2005). We computed step-length 172 asymmetry on each stride cycle by calculating the difference in step lengths (i.e., ankle to ankle distance at 173 foot landing) for two consecutive steps taken with the fast and slow leg. This difference was normalized by the 174 sum of step lengths to obtain a measure that was a proportion of each subjects' step sizes. A zero step-length 175 asymmetry value indicated that steps lengths were even, negative values indicated that the (non-dominant) leg 176 walking on the slow belt was taking longer steps than the (dominant) one on the fast belt and viceversa for 177 positive values. 178 We also computed body displacement with respect to the foot in contact with the ground during the stance phase 179 for each leg. This was done to interpret the changes in muscle activity upon transitions between tied and split 180 conditions. Body displacement was computed as the anterior-posterior distance between the middle of the hip 181 markers (greater trochanter) and the ankle from ipsilateral heel-strike to contra-lateral heel-strike. To focus on 182 displacement changes within this period of the stance phase, we aligned and divided the body displacement in 183 two sub-intervals: double support (from ipsilateral heel-strike to contra-lateral toe off) and single stance (from 184 9 . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint contra-lateral toe-off to contra-lateral heel-strike). This normalization of the time axis facilitated the comparison 185 of body displacement during stance of the fast vs. the slow leg or across walking conditions given the distinct 186 stance durations between legs and treadmill modes (tied vs. split) (Reisman et al., 2005). were considered but the first one was excluded. 195 We compared muscle activity for a given epoch with that of Baseline walking ( Figure 1D). This allowed us to  Figure 1B). We also computed differences in muscle activation 199 variables between consecutive epochs in which subjects transitioned from the split-to-tied (e.g., eP lA = eP − lA)  The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint distinct temporal dynamics of these processes. We specifically assume that feedforward components change slowly upon experiencing a novel situation because they are generated purely through internal models that are updated gradually. On the other hand, feedback components can change rapidly after a transition in the walking environment since sensory information is available immediately following a change in the walking context.
Note that our definition of feedback-generated motor commands includes modulation of muscle activity that occurs at different latencies after a movement disturbance (i.e., short-and long-latency reflex responses and subsequent voluntary responses) (Horak et al., 1990). In sum, we quantified feedback-generated activity as the changes in EMG after a sudden environmental transition (i.e., EMG a f ter ) with respect to EMG activity before the transition (i.e., EMG be f ore ). In equation form: ∆EMG = EMG a f ter − EMG be f ore = FEEDBACK-GENERATED ACTIVITY Thus, even if the change in EMG is strictly equal to the change in both feedback and feedforward-generated 206 activity, we approximated the latter one to be negligible because it occurs at a slower rate compared to changes 207 in feedback-generated activity after an abrupt transition in the walking condition. As an example, in the  The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint (non-adaptive) feedback responses (H2), and 3) adaptive feedback responses (H3). The basis and rationale for 220 each one of our hypothesis is presented in the following sections. One possibility is that muscle activity is mostly controlled in a feedforward manner when the split perturbation is removed. By mostly, we mean that feedback responses are negligible in magnitude when compared to the feedforward components. In other words, muscle activity before (lA B ) and after (eP B ) the split-to-tied transition is roughly the same, as illustrated in the schematic of H1 in Figure 1C. This hypothesis suggests that feedback components are present during the tied-to-split transition (eA B ), but are greatly reduced during the split-to-tied transition because the return to a familiar environment (i.e., tied condition) may not require the same extent of feedback responses as when a novel condition (i.e., split) is experienced. We formulated this hypothesis on the basis that aftereffects result from the continuation of the motor output updated during the Adaptation period Another possibility is that feedback responses are not negligible during post-adaptation, but they are not adaptive in nature. That is, the feedback processes generating them are not updated during split-belt walking and produce corrective muscle patterns solely as a function of the sensed environment. This hypothesis makes two predictions. First, feedback-generated activity would change similarly after removing the split perturbation following a long or short exposure to the split condition because the environment transitions identically in these two situations. Second, we would expect that changes in muscle activity originated by the tied-to-split transition (i.e., eA B ) would be numerically opposite from those from the split-to-tied transition ( Figure 1C, H2 schematic) because the environment changes in exactly the opposite manner in these two transitions. Thus, any muscle whose activity increased when the split perturbation was introduced, should decrease by the same amount when it is removed (and vice versa). We reasoned that this hypothesis would be plausible given that 12 . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint step length asymmetry (conventionally used to characterize locomotor adaptation and after-effects) is perturbed during early adaptation and early post-adaptation by approximately the same amount but in opposite directions (Reisman et al., 2005). Formally expressed: Where eA B represents muscle activity during the Early Adaptation with respect to Baseline. These aftereffects 224 (eP B ) are exactly what would be predicted from a linear model whose input is environmental condition and its 225 output is muscle activity. Finally, we considered the possibility that substantial feedback responses are not only present following transitions in walking speeds, but they are adaptive. According to this hypothesis, changes in activity from the split-to-tied conditions would be different after a short vs. a long exposure to the split environment. We specifically considered that following the long exposure, the motor system would learn that the split (perturbed) condition is the "new normal" and hence departures from it would be processed as perturbations, just as when the novel environment was first introduced. We based this hypothesis on previous work reporting that the removal of the novel condition is in-and-of itself a perturbation (Herzfeld et al., 2014b), and that feedback responses are adapted through experience to an altered environment even if there is no feedback-specific learning opportunities (Wagner and Smith, 2008;Yousif and Diedrichsen, 2012). During the split-to-tied transition in our experiment the leg that was 'fast' moves down in speed, while the one that was 'slow' moves up. This is the mirrored version of the initial tied-to-split transition where the 'fast' leg went up and the 'slow' leg went down. Accordingly, we expect the feedback-generated muscle activity to reflect this mirror-symmetry between belt-speed transitions when the split condition is introduced or removed. Formally expressed: Where eA * B is eA B mirrored (i.e., transposing the values for the left and right legs).

228
13 . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint 2.5 A general regression model to characterize the contribution of feedback-generated activity to aftereffects 229 in muscle activity 230 In summary, we tested our three hypotheses on the structure of feedback-generated components of aftereffects 231 with one single regression model: Where the parameters β S and β M are respectively interpreted as the extent of feedback-generated activity that were determined by using the Wilcoxon signed-rank test (i.e., non-parametric analogue of a paired t-test) on 244 each of the 360 EMG parameters. Effect sizes were computed using median values across subjects since this is a 245 measure less susceptible to outliers. All tests were two-tailed and the null hypothesis was that the (normalized) 246 absolute effect size was smaller than 0.1 (i.e., 10% of the maximum baseline activity for that muscle). This was 247 done to avoid finding significant but small, and presumably meaningless, differences. We corrected for multiple Contribution of feedback activity to aftereffects. The linear regressions for testing our hypotheses on the structure 253 of feedback-generated activity were performed using Matlab's fitlm() function and computing (Pearson's) R 2 254 values that were uncentered, given that our regression model did not include intercept terms.

255
Analysis of inter-subject variability. Lastly, we used Spearman's correlations to determine if either age or walking 256 speed could explain the inter-subject variability in our model's coefficients. We focused on these subject-specific 257 features because they exhibited large ranges in our cohort that could have impacted our results. Spearman's 258 correlations were also used to study the association between aftereffects sizes in muscle activity and step-length 259 asymmetry. For all these analyses we applied Spearman's correlations as a non-parametric alternative to 260 Pearson's correlation because it is more robust to outliers. The correlation value (lowercase r, to distinguish 261 from previous use of R) and the corresponding p-value were presented. To parse the relation between age and 262 kinematic aftereffects onto muscle activity aftereffects, we used a two-factor linear regression. In this case there 263 was no non-parametric method available, so we used a robust linear fit (bisquare weighting) to protect against 264 potential outliers, and we report Pearson's R 2 . 265 2.7 State-space modeling 266 We implemented a linear time-invariant state space model as an analytical tool, rather than a mechanistic model 267 (e.g Thoroughman and Shadmehr, 1999), to characterize the evolution of EMG activity during the Adaptation 268 and Post-Adaptation conditions. In particular, we used this framework to determine the extent to which the 269 state of the system underlying the muscle activity during Adaptation predicted that of Post-Adaptation. To Whereŷ k is the estimated EMG data and x k is the state vector of the system at stride k. C and D are matrices than an input signal v k as we do here. As long as the error signal is presumed to be a linear combination of 297 motor output (y k ) and some external input (v k ), as is the case in all those examples, the models can be expressed 298 in the form given by the equations above. However, a transformation of the parameters needs to be considered 299 when comparing parameters across both formulations (Cheng and Sabes, 2007). 300 We also used the state-space approach described above to fit the EMG difference between legs at each stride k 301 during Adaptation (i.e., y k = EMG sym = EMG dominant − EMG nondominant = 180 x 1 vector) and predict EMG sym The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint during Post-Adaptation. This was done given that EMG s ym would yield the same Post-Adaptation values, 303 regardless of whether the structure of the feedback-generated activity was solely environment-dependent (H2) 304 or adaptive (H3).

305
The goodness of fits and prediction were quantified through the RMS residuals: represents the data for the j-th EMG variable (e.g., j ∈ [1, 360] for bilateral EMG fitting) during the i-th stride,ŷ i j 307 is the corresponding model output, and N is the number of strides considered. k is a model-specific constant that   Early Adaptation immediately after hitting the ground (heel-strike) with the fast or slow legs, respectively.

327
This was likely due to the discrepancy between subjects' expected and actual walking speed (Figure2B).

328
Consistently, the corresponding muscle activity for both legs during 100-400ms (colored rectangles) after   Figure 3B). Namely, we found 352 that if a group of muscles increased activity on one side, the same group decreased activity on the other one.

353
Interestingly, this opposing modulation across legs was not merely determined by ipsilateral walking speed. 354 Should this have been the case, one would expect reduced muscle activity of the slow leg with respect to  The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint 3.3 Feedback responses are transposed across legs when returning to tied-belts after prolonged exposure to 368 split-belts 369 We proposed that aftereffects (eP B ) would be generated by a combination of feedforward and feedback processes.

370
In particular, one of our hypotheses (H1) predicted that feedforward-generated motor commands would dominate 371 EMG aftereffects, such that muscle activity during Early Post-Adaptation would be similar to that of Late 372 Adaptation (lA B ). Qualitatively, we found little resemblance in the muscle activity during these two epochs, 373 as shown by the few significant changes that were similar between them (black outlines in Figure 3C). This half to 4D top half) at these transitions, except for the reduced activity in some muscles during early stance (RF, 390 VL, VM, GLU, TFL). We also found that the body motion exhibited mirroring between the legs' stance phases.

391
For example, the change in body position with respect to the fast stance leg in the tied-to-split transition (Fig.   392 4F, green trace) closely resembled the one during slow stance at the split-to-tied transition (Fig. 4F, magenta 393 20 . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint trace). Thus, both EMG and kinematics exhibited mirroring when we consider changes with respect to the 394 previous steady-state condition. This is consistent with the nervous system (specifically, the feedback response 395 mechanisms) adopting the steady-state behavior as a new reference gait pattern. Perturbation of kinematics 396 away from this new reference, and not perturbations away from the subjects' baseline behavior, may drive 397 feedback-generated muscle activity.

398
To confirm that the mirroring between legs is a learning-dependent process, and not simply due to removal of 399 the split perturbation, we ran the same regressions on the split-to-tied transition following the Short Exposure 400 condition. In this condition subjects did not have time to adapt, so we expected to observe solely environment-dependent 401 changes in EMG activity (H2), rather than mirroring (H3). Indeed, that is what we found with our regression muscle activity can be modeled as environment-dependent, whereas following Adaptation (long exposure) the 408 feedback is well explained by mirroring. Taken together, we found that feedback-generated activity was adapted 409 and strongly contributed to the structure of aftereffects in muscle space. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint activity because of the reduced signal to noise ratio, leading to biased or noisy estimates of β M and β S . To discard 420 this possibility, we correlated the R 2 of the fitted models with age. We found no effects of age (r = −0.36,  Figure 5C). Importantly, walking speed, 428 which naturally alters muscle activity, was not associated to either the magnitude of EMG aftereffects (p = 0.36) 429 nor to the adaptation (speed vs. betaM p = 0.64) or magnitude of split-to-tied feedback responses (p = 0.59).

430
Interestingly, no age mediation was observed in the magnitude of step-length asymmetry aftereffects (r = −0.16, 431 p = 0.55, Figure 5D), which are conventionally used to characterize locomotor learning. This suggests that 432 the adaptation of muscle activity may be more sensitive to individual differences in learning than kinematics.

433
Alternatively, there might be partially different mechanisms involved in the adaptation of muscle activity and 434 step length asymmetry. In agreement, muscle activity aftereffects and kinematic aftereffects were not correlated 435 (r = 0.38, p = 0.15, not shown) unless the effect of age on EMG aftereffects is accounted for. Specifically, 436 the robust linear regression (see Methods) of muscle activity aftereffects onto both kinematic aftereffects and 437 age indicated that both factors were significant predictors (Pearson's R 2 = 0.741, p age = 0.018 and p SLA = 438 4.2 × 10 −4 , not shown). We interpret these results as supportive of two different learning mechanisms: one 439 that underlies the association between age and EMG aftereffects, and a separate one that is captured by the 440 association between EMG and kinematics aftereffects. In sum, we observed diminished mirroring of feedback 441 activity post-adaptation and smaller EMG aftereffects in older adults, suggesting that healthy aging has a 442 negative impact on learning process updating feedback-generated muscle activity. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint

The dynamics of Adaptation predicts Post-Adaptation of EMG symmetry, but not of individual legs 444
The adaptive nature of feedback-generated activity (Hypothesis H3 in Figure 1C) was further supported by 445 the poor prediction of EMG aftereffects of individual legs compared to symmetry measures ( Figure 6). Recall 446 that only purely environment-driven feedback responses (Hypothesis H2 in Figure 1C) can be reproduced by a 447 linear model framework. In other words, a necessary condition for a system to be linear time-invariant (LTI) 448 is that its response (output) upon introduction of a given perturbation (input) would be numerically opposite to 449 the one observed upon removal of said perturbation after reaching steady state. Consistently, our state-space 450 model predicted that any muscle whose activity increased when the split perturbation was introduced during The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint 4 Discussion 470 We have shown that aftereffects following split-belt walking were not merely a continuation of adapted muscle 471 activity patterns, but they exhibited a large contribution of feedback-generated motor commands. Further, 472 we showed that the elicited feedback responses were adapted, such that removing the split condition induced 473 muscle activity similar to the one observed when unexpectedly increasing or decreasing walking speed upon 474 introduction of the perturbed environment. This indicates that subjects adopted the split pattern as their reference 475 gait, which cannot be modeled with a linear system predicting that feedback motor commands would be reversed 476 when the split environment was removed. Interestingly, this phenomenon was age-dependent, with older  Aftereffects in muscle activity are not a continuation of late adaptation activity. 482 We have shown that aftereffects following split-belt walking were not merely a continuation of adapted muscle 483 activity patterns. Notably, only a few muscles exhibited similar changes during late Adaptation and early 484 Post-Adaptation. We expected to find more similarities between these two epochs given that kinematic aftereffects The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint with a distinct approach, such as simulations with models of the musculo-skeletal system (Song and Geyer, We consider three possibilities on how the the feedback loop may be adapted. First, the feedback policy 514 transforming the sensed state into corrective actions might rely on on internal models that are adapted through 515 exposure to an alternative environment (Bhushan and Shadmehr, 1999). Second, the state estimation may be 516 inaccurate due to sensory shifts following split-belt walking (Vazquez et al., 2015). As a result, feedback 517 commands might be different post-adaptation even if the feedback policy remains unchanged. Third, feedback 518 adaptation might be merely a byproduct of feedforward adaptation. That is, feedback activity post-adaptation 519 might be generated because of a misapplied motor pattern, arguably generated through a feedforward mechanism. 520 While possible, we believe this last option cannot fully explain our observations given that age-related decay is 521 25 . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint observed in the adaptation (i.e., mirroring) of feedback activity but not in kinematic aftereffects. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint correlate of neural activity than kinematics, are more sensitive to differences in internal model recalibration 547 (Thoroughman and Shadmehr, 1999). Alternatively, this EMG/kinematic discrepancy on age-related modulation 548 of aftereffects may imply a partial dissociation of the underlying neural circuitry of each type of aftereffect.

549
Further, the magnitude of muscle patterns during Late Adaptation, which is mostly feedforward-generated 550 activity, was not affected with age. The dichotomy in age effects on feedback vs. feedforward activity suggests a 551 partial distinction in the neural mechanism underlying the adaptation or execution of feedback and feedforward  The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint environment (late Adaptation) is not consistent with simple ipsilateral speed-based modulation of activity. Our 573 observation is that distinct muscle groups up-and down-regulate their activity on both legs, whereas ipsilateral 574 speed-based modulation would predict down-regulation only in the slow leg, and up-regulation on the fast one.

575
This highlights the bilateral coordination requirements of walking, and suggests the adapted state of the two 576 legs are not decoupled from one another. In contradiction to our finding, it has been suggested that adaptation 577 to split-belt walking may happen through leg-specific neural substrates that can be adapted independently (Choi 578 and Bastian, 2007). We hypothesize that coupling is likely dependent on the constraints between the legs   Within our findings, three elements appear of importance towards the development of personalized rehabilitation 606 treatments. First, the age dependence of some responses suggests that patient age needs to be considered 607 a potential factor when assessing the effectiveness of treatment. Second, assessment of learning through 608 kinematic-only measurements may be missing interesting insights into neural activity changes which may be 609 better captured through EMG-based measurements. Third, aftereffects following a sudden change in walking 610 dynamics may be more strongly influenced by the feedback reaction to that change than by changes in feedforward 611 circuits, which are arguably the true target of rehabilitation therapies. Thus, long-term rehabilitative potential 612 may not be well captured by measuring behavior immediately after split-belts treadmill walking.   The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint  Figure 2C. (D) Split-to-tied change in EMG data (early Post-Adaptation with respect to late Adaptation). We observe that changes in activity in the fast (slow) leg during this transition resemble those of the slow (fast) leg during the tied-to-split transition (mirroring). For example, TA is not active during stance in steady-state walking (see Figures 1D and 4A,B), but it is active following the tied-to-split transition in the slow leg, and the split-to-tied transition in the fast leg. (E) Quantification of the validity of H2 and H3 in split-to-tied transitions, after the short exposure (SE, gray) and long exposure (LE, magenta) to the split-belts environment. Large dots represent regressor values for mean subject data, small dots represent regressors for individual subjects. Bars indicate median and interquartile range for individual subject regressions. (F) Changes in hip position (with respect to previous steady-state) after the two transitions (tied-to-split in green, split-to-tied in magenta). Line represents mean across subjects, shaded area is standard error. We observe these the kinematics of these two transitions are approximately mirror images of one another, possibly causing the observed mirroring of EMG responses. For comparison, we also present early Post-Adaptation with respect to Baseline (gray).

41
. CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint late Adaptation (steady-state) and early Post-adaptation (aftereffects). Aftereffects, but not steady-state, is correlated with age. This confirms that older subjects are able to modulate muscle activity as much as healthy subjects, but display smaller aftereffects. (D) Step-length asymmetry aftereffects are not correlated with age.
This shows EMG-based measures of learning are more sensitive than kinematic-based ones.

42
. CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint  The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/372359 doi: bioRxiv preprint