Split‐Belt Treadmill Training to Improve Gait Adaptation in Parkinson's Disease

Gait deficits in people with Parkinson's disease (PD) are triggered by circumstances requiring gait adaptation. The effects of gait adaptation training on a split‐belt treadmill (SBT) are unknown in PD.

display a wide heterogeneity in cognitive profiles, 9,14,15 and therefore, such multimodal training may not be suitable for all individuals. Hence, the current study investigated a paradigm for tackling gait adaptation without adding cognitive load, thereby widening possible gait training options for PD.
Split-belt treadmill (SBT) training offers safe and controlled practice of asymmetric and changing gait patterns in a progressive and individualized manner. A SBT has two belts and allows control of the speed of each belt separately. A switch of speed in one belt prompts a bout of sudden asymmetrical walking, producing a sensory prediction error and a subsequent slow adaptive response to regain symmetric motor control and gait stability. 16 Similar to turning, the body center of mass needs to be controlled and adjusted in multiple directions when the stance phases are manipulated by belt speed switches. Therefore, SBT-perturbations elicit largely automatic responses without demanding explicit attention, 16 which parallel the daily-life adaptations made after a sharp turn or a sudden switch in walking direction. Previous studies found that particularly patients with FOG showed impaired adaptation to SBT-conditions. 2,17 However, when comparing mild PD to healthy controls a very similar adaptive ability was found. 18 This largely preserved response may be explained by the fact that the so-called "fine-tuning" locomotor network, including the frontoparietal cortex and the cerebellum, is relatively spared in PD taking on a compensatory role. 16,19,20 Yet, whether repeated exposure to SBT will enhance long-term acquisition of gait adaptation skills and whether this will improve turning in a wider PD population is currently unknown.
Therefore, we first conducted a pilot trial in a cohort of 45 freezers and 36 healthy controls (HC) to examine if a single session of SBT led to 24-hour retention. 21,22 We also compared various doses of speed contrasts with regular (tied-belt) treadmill (TBT) to be able to underpin the design of a prolonged SBT-program. We found that gait adaptation gains were retained, albeit less so than in HC and that this positively modified turning and percentage time frozen. 21,22 When freezers and HCs were analyzed together, the largest gains pertained to turning speed on the spot (360 ). Given its association with FOG and falls, 23,24 turning speed was chosen as the primary outcome allowing to include both freezers and non-freezers in the trial. This not only provided a pragmatic advantage for recruitment, but also permitted exploring the differential impact of training on the cohorts. Consequently, we set out to investigate whether 4 weeks of progressive SBT versus TBT training would lead to robust gait adaptation as manifest during (1) untrained turning over-ground, (2) after a 4-week follow-up and (3) during dual tasking.

Methods
This multi-center randomized single-blind controlled study was conducted at the Christian-Albrechts-University (CAU), Kiel, Germany and the KU Leuven, Belgium. The trial was preregistered at ClinicalTrials. gov (ID: NCT04176263).

Participants
Participants were included between August 2019 and May 2021, including the period of the coronavirus disease 19 (COVID-19) pandemic. Inclusion criteria were: (1) diagnosis of idiopathic PD as confirmed by a neurologist according to the Brain Bank criteria 25 ; (2) Hoehn and Yahr (H&Y) stage 2 or 3; (3) freezers and nonfreezers who could walk without an aid for at least 5 minutes; and (4) stable medication. Exclusion criteria were: (1) presence of other neurological diseases; (2) cognitive impairment (Mini Mental State Examination ≤24); (3) orthopedic or other health conditions that might influence gait or balance; (4) enrollment in another clinical study; (5) any other gait training 1 month before participation; and (6) cardiovascular exercise risk factors as diagnosed by a doctor (established via questionnaire). Participants were classified as freezers based on subjective reporting of FOG (New Freezing of Gait Questionnaire [NFOG-Q] item 1 = 1) or after visual observation of FOG during the baseline assessments. Both ethical committees approved the study protocol (CAU Kiel: D 454/13; KU Leuven: S62825) and participants gave written informed consent before participation.

Study Design and Procedures
Participants were randomly allocated to either SBT or TBT in blocks of four and stratified per center for H&Y stage and FOG status using the Research Randomizer software (https://www.randomizer.org). Concealment of randomization was performed by a dedicated person in each center, not involved in data acquisition or training. Allocation was communicated via a secured e-mail system. Assessments and the statistical analysis were conducted while blinded to group allocation. Because of the nature of the trial, participants and training staff were unblinded, but participants were kept blind to whether their treadmill program represented the experimental or the control intervention.
Data collection was conducted at three time points: 1 week before the start of training (pre), 1 week after completion (post) and after the 4-week follow-up ( Supplementary Fig. 1). Training and testing were conducted in the on-state, the latter with strict adherence to the time of medication intake and with close monitoring (approximately every hour) of the consistency of patients' on-state. Change in the levodopa equivalent daily dosage (LEDD) over time was documented. Participants wore a safety-harness when walking on the treadmill and were encouraged to walk without the handrails. Only eight participants needed the handrails and usage was kept constant throughout the study. To ensure comparability between centers, in-person meetings and regular online fidelity checks of the study procedures were held between assessors and therapists.

Primary Outcome and Other Over-Ground Mobility Measures
Based on our pilot work and our interest in investigating transfer of SBT-training to a complex gait task offtreadmill, 21 we chose turning speed as our primary outcome. Turning speed was assessed by asking participants to turn as quickly as possible on the spot (360 ) for 1 minute in alternating directions in single task (ST) offering a degree of resemblance to the asymmetrical perturbations experienced during SBT. Turning was also performed under dual-task (DT) conditions because this was found to be particularly sensitive for assessing FOG. 26 The secondary task involved an auditory Stroop task (the words "high" and "low" [in German or Dutch] were presented congruently or incongruently in a high or low pitch). Participants had to respond to the pitch of the word as quickly as possible. 27 Turning kinematics were quantified with APDM Mobility Lab (Portland, USA) accelerometers placed on the lower back and on the shins. A custom MATLAB script was used to compute turning speed, the turning speed coefficient of variation (CV) and the number of turns. 24 Trials with FOG were disregarded (pre: 24%, post: 19%, followup 20%).
For over-ground gait analysis, participants were asked to walk 10 times along a 10-meter distance (4 meter capturing area) at a comfortable pace with and without the auditory Stroop task. Gait kinematics were calculated using 3D motion capture technology (CAU: Qualisys Motion Capture, Systems, Gothenburg, Sweden; KUL: Vicon Motion Systems, Oxford, UK) with six passive reflective markers placed on the feet (lateral malleolus, heel and tip of the shoe on each foot).

Gait Adaptation Outcomes on the Split-Belt Treadmill
Gait adaptation was assessed on the treadmill according to a previously developed protocol. 22 This required three consecutive walking cycles of 1 minute, during which subjects were walking in (1) TBT (baseline); (2) 50% reduction on one side (to assess early and late SBT-adaptation); and (3) TBT (to assess early and late de-adaptation) ( Supplementary Fig. 2). Testing was conducted with and without the auditory Stroop task to probe whether training gains were held when a secondary task was added. The ST-condition was performed before the DT to avoid order effects across repeated measures. Treadmill velocity was personalized to the individuals' comfortable over-ground walking speed. The adaptation protocol was repeated for the left and right body side in a random order. Gait characteristics were calculated using custom MATLAB scripts. The outcome of interest was step length asymmetry averaged over the 3-minute protocol (total adaptation) and over the different adaptation phases separately (baseline, early-split, late-split, early-tied, and late-tied using 10 second segments). 22 Step length asymmetry was calculated as a ratio value: (fast leg parameterslow leg parameter)/(fast leg parameter + slow leg parameter). Values closer to zero represented better gait adaptation. 22 The statistical analysis was conducted on the averaged and normalized value of the total adaptation and the late-split belt phase (see Supporting Data).

Freezing of Gait
FOG-severity was evaluated as a secondary outcome in the group of freezers only by calculating: (1) the FOG-score during walking in a FOG-provoking protocol, 28 (2) the percentage of time frozen, 26 and (3) the FOG-ratio (mediolateral) 29 during the 360 turning test.

Clinical Measures
We also collected the Movement Disorder Society Unified Parkinson's Disease Rating Scale Part III (MDS-UPDRS III), the Mini-Balance Evaluation Systems Test (Mini-BESTest), the Fullerton Advanced Balance Scale (FAB-S) and the Falls Efficacy Scale-International (FES-I). The New Freezing of Gait Questionnaire (NFOGQ), the Montreal cognitive assessment (MoCA), the Frontal assessment battery (FAB), and the trail making test (TMT) were calculated at baseline only.

Training
Both interventions consisted of 4 weeks of supervised treadmill training delivered three times per week with a rest day in between sessions. Training progression was standardized and increased from 30 to 45 minutes per session with 5-minute increments per week. In addition, 16 progression levels were developed to achieve equal training motivation in each arm (Supplementary Tables 1-4). In the SBT-group, the 16 progression levels were determined by varying the magnitude and frequency of the speed contrasts and by including one or two body sides. Belt-ratio reductions between 25% and 50% of comfortable over-ground gait speed were implemented, which were shown in previous study to elicit the best results. 21,22 The frequency of speed changes increased from 4 to 16 switches per 5 minutes. The 16 progression levels in the TBT group were characterized by increasing the walking speed from 80% to 110% of comfortable walking speed and by increasing the duration of bouts from 5 to 20 minutes of continuous walking. Therapists only personalized the training progression if needed by omitting or adding a challenge according to the participants' level of exertion as assessed by heart rate monitors and the Borg scale. However, training duration was always kept in accordance with the required duration for each training week. Details of the training protocols are summarized in the Supporting Data and explained through the TIDieR checklist. 28

Sample Size Calculation
Sample size was determined a priori based on our pilot data, demonstrating an average turning speed of 74.32 (AE27.22) at baseline. We assumed that the SBTgroup would improve 14% compared to the TBTgroup. 21 We calculated a sample size of n = 50, including 20% dropout rate to detect a significant time*training-group interaction with α = 0.05 and a power of 80%.

Statistical Analysis
Mann-Whitney U tests were used to compare between-group differences at baseline. Linear mixed models were calculated for the primary and secondary outcomes. The model included time (3 levels) and training-group (2 levels) as independent variables to investigate the time by training-group interaction as well as center (2 levels) as a covariate. The model included a random effects term with a correlated random intercept and slope for time and subject. A Satterthwaite approximation of the denominator degrees of freedom was used. Histograms and Q-Q plots were used to visually check for normality of the residuals. Post hoc tests were conducted with Tukey adjustment for multiple comparisons. In line with the intention-to-treat approach, all randomized participants were included in the mixed linear models to accommodate missing data. We explored the impact of SBT versus TBT on FOG-related outcomes in freezers as a sub-analysis. To investigate whether freezers and nonfreezers improved similarly, an additional linear mixed model was run with time and FOG-status as independent variables to investigate the time by FOG-status interaction (for the total sample and for the two training-groups separately). Additionally, in an extra analysis including the factor "task" (single vs. dual task) to the model we investigated whether training effects were similar in these conditions. Effect sizes (Cohen's d) 30 were calculated between groups for the differences from pre to post or pre to follow-up, respectively and reported in Table 2. Effect sizes 30 and their confidence intervals 31 are also visually depicted in a metaview of the pre-to follow-up differences only. We interpreted effect size as small (≥0.2), medium (≥0.5), or large (≥0.8). 32 The statistical analysis was conducted with SAS software (version 9.4, SAS Institute, Cary, NC) with a preset level of significance of P < 0.05.

Participants' Characteristics
Fifty-two individuals with PD were included ( Fig. 1) of whom n = 22 were freezers (SBT: n = 12; TBT: n = 10). Two (4%) additional participants were recruited to compensate for the dropout risk because of the COVID-19 pandemic. Overall, compliance and protocol deviations were similar in each group (Supplementary Tables 5 and 6). Table 1 shows that participant characteristics were similar for both groups. There was no significant difference in LEDD between groups (P = 0.408) and between time points (P = 0.840).

Effects on Turning
No significant time by training-group interaction was found for turning speed (ST: P = 0.55; DT: P = 0.95) and for any of the other turning outcomes (Table 2). However, there was a significant time effect for the number of turns. (ST: P = 0.020; DT: P = 0.020). Exploratory within-group analysis showed significant increases in the number of turns performed in ST and DT in the SBT-group from pre to follow-up (ST: P = 0.030; DT: P = 0.004) not present in the TBTgroup. Figure 3 displays the pre to follow-up effect sizes in a metaview, showing effects in favor of SBT for the number of turns, whereas turning speed variability showed larger changes in the TBT-group. Effect sizes were largely similar in ST and DT, which was confirmed by the fact the factor task was not significant in the model (P = 0.42).

Effects on Gait Adaptation
A significant time by training-group interaction was found for the gait adaptation outcomes with significantly greater improvements of step length asymmetry in the SBT-compared to the TBT-group (total:   Step length asymmetry at late-split    Table 2). Figure 2 illustrates that compared to pre, the SBT-group was able to achieve greater reduction in their step length asymmetry at late-split adaptation (ST: P < 0.001, DT: P < 0.001) and for total adaptation (ST: P = 0.020) at post. Furthermore, these improvements were also retained at the 4-week follow-up for late-split (ST: P < 0.001, DT: P = 0.020) and total adaptation (ST: P < 0.001) in both ST and DT. When adding the factor task (ST vs DT) to the model, no significant interactions were found (P = 0.19), indicating that improvements were similar in both conditions.

Effects on over-Ground Gait
For over-ground gait no significant time by traininggroup interactions were found for gait speed (ST:  Table 2). Significant time effects showed that both training groups increased their gait speed and step length under ST and DT conditions (ST: P = 0.026; DT: P < 0.001 and ST: P = 0.028; DT: P = 0.006, respectively). Figure 3 shows a metaview of the effect sizes of the change scores from pre to 4-weeks follow-up between the SBT-and TBT-group, whereby values greater than zero indicate larger effects for the SBT-group. Between-group analyses shows a tendency for larger effects in the SBT-compared to the TBTgroup, suggesting that these time effects were mostly driven by the improvements in the SBT-group. Furthermore, there was a significant time effect for gait asymmetry (ST: P = 0.011), showing decreased asymmetry from pre to follow-up in both training groups.

Effects on FOG
We found no significant time by training-group interactions for the freezing-related outcomes (ie, the mediolateral FOG ratio, the Ziegler score and the percentage time frozen) (Supplementary Table 8). When adding FOG status as a factor in the statistical model, it had no significant impact on the results, indicating that freezers and non-freezers benefitted similarly from the training. The sub-analysis for freezers showed no significant time by training-group interactions for the FOGrelated outcomes. When running models for each group separately, including FOG status and time as fixed factors, it was shown that non-freezers benefited less from TBT than freezers for DT gait speed (P = 0.040) and DT step length (P = 0.013) (Supplementary Table 9).

Clinical Measures
A significant time effect was found for MDS-UPDRS III (P = 0.002) showing a clinically meaningful improvement of À5.1 points from pre to post and À 5.8 points from pre to follow-up for the SBT-group, in contrast to the TBT-group (À2.9 points from pre to post; À3.5 points from pre to follow-up). No significant interaction or time effects were apparent on the clinical balance and fear of falling scales (Supplementary Table 7). Selfperceived fatigue (BORG-scale) during training was similar across both groups, 11.76 (AE2.4) in the SBT-group and 11.73 (AE2.2) in the TBT-group (P = 0.963).

Discussion
This is the first study investigating the effects of prolonged SBT-training in a mid-stage population of people with PD. We found that SBT was superior to TBT in improving gait adaptation on the treadmill with effect sizes ranging from medium to large. Furthermore, these effects were sustained at 4-weeks follow-up and still present during dual task conditions. Gaining consolidated improvements of gait adaptation through an implicit learning paradigm is remarkable in a condition such as PD, which is notorious for its automaticity and motor learning deficits. 33 However, and crucially, the training effects did not transfer to over-ground turning speed, the primary outcome of this trial. Both treadmill groups achieved significant gains for over-ground step length and gait speed, exceeding the minimal clinically important difference. 34 Additionally, the MDS-UPDRS III scores were improved and retained for 4 weeks in both groups, but this only reached clinically meaningful effects in the SBT-group. These results indicate that SBT improved over-ground gait and symptom severity similarly or better than the well-established traditional treadmill training. 10,35 Additionally, it had surplus and specific value for gaining gait adaptation capacity to asymmetric treadmill perturbations. Balance outcomes and falls efficacy were not improved in either group.
The lack of transfer effects to over-ground turning is an important finding, although at the within-group level the number of turns increased significantly in the SBT-and not in the TBT-group. Of note, this was not supported by significant interaction effects. This unexpected result was in contrast to our pilot data indicating that one session of SBT-training not only improved gait adaptation, but also DT turning speed in a population of freezers. 21 We propose four explanations for this lack of transfer. First, because our pilot trial included freezers only, the baseline values of turning speed were worse than in the present mixed cohort. In previous rehabilitation studies, 36,37 worse baseline scores predicted better training effects probably due the larger scope for improvement. The sample size calculation was also based on these pilot data. Hence, our study was probably underpowered. Indeed, although the effects on the turning outcomes were consistently better in SBT than in TBT (recall Fig. 3), they never reached statistical significance.
Second, because SBT-perturbations require multidimensional stepping and postural responses, the choice of the transfer task may have been suboptimal. Transfer of motor learning implies that after the training of one specific task, benefits can be expected in task variants or in novel conditions. 38 Depending on the level of resemblance with the trained task, "near" or "far" transfer is inferred. We used 1 minute of over-ground turning (360 ), while switching directions, as the transfer task. However, turning is deemed a complex task requiring visual guidance, extensive sensory integration, and head-pelvis dissociation. 29 Because turning on the spot does not capture a shift from straight-line walking to asymmetrical gait, it may have been too dissimilar as a near or far marker of transfer.
Third, the lack of transfer could be specific to PD. 33,39 Compared to HCs, people with PD showed overall intact motor learning capacity of goal-directed upper limb tasks, but a clear lack of transfer to different task conditions and different contexts. 40 Additionally, during an adaptive visuomotor reaching task, people with PD had a lower capacity for inter-limb transfer of the acquired adaptation skills. Interestingly, this deficit proved correlated to reduced dopamine transporterbinding in the right striatum. 41 Fourth, the training conditions could have been responsible for the lack of transfer. 39 Motor learning theory stipulates that offering a wide range of task variations will induce better transfer as it builds a global representation of a motor skill, thereby facilitating independence from the learning context. 42 This principle has been confirmed by a recent study showing better transfer to over-ground gait after treadmill training when practicing under varying conditions. 43 Work on SBT-training in healthy people confirms this notion, because it was found that introducing switches every 20 seconds rather than exposure to constant asymmetry, induced better transfer to a novel SBT-task. 44 Similarly, we found in our pilot trial that changing the ratios of the speed perturbations led to better transfer than keeping the contrasts constant. Consequently, this principle was adhered to during this study, yet without the desired transfer. It is possible that for some subjects too much variety was offered over a 4-week training, so that because of their compromised brain capacity predictions could not be made to off-treadmill performance. Certainly, transfer of SBT-training in stroke patients was enhanced by introducing perturbations gradually over a longer period rather than abruptly. 45 Recently, it was proposed that learning a repertoire of motor memories via contextual inference is key to growing transferable sensorimotor repertoires. 46 This would point to the need for adding off-treadmill learning conditions to SBT-training to develop the flexibility to cope with the transitions inherent to daily life. 46 We included both freezers and non-freezers in this study and found that both groups benefitted equally from SBT-training in line with our earlier work. However, the finding that SBT had no effect on FOG-outcomes was not in agreement with our previous data 21 and may partly be attributable to the inability to capture FOG-phenomena during laboratory testing in on. Motor-cognitive treadmill training also showed no effect on self-reported FOG. 47 Interestingly, within-group analyses showed that freezers benefited more from TBT-training than non-freezers, especially for DT gait outcomes. This may be explained by the fact that TBT, especially when provided at the high and progressive dose that was administered in our study, enhanced gait capacity in freezers, which is commonly more depleted than in non-freezers. 37 Overall, both treadmill programs showed significant time effects on overground gait and motor symptom severity, which were consistently larger in the SBT-group yet training effort was the same. Effects sizes for the MDS-UPDRS-III in the SBT-group resembled those of other walking endurance studies, 11,35 supporting the notion that challenging treadmill training constitutes another viable rehabilitation option to alleviate motor symptoms in PD.
The current findings suggest that SBT-training is a promising therapeutic avenue to modify gait adaptation in PD in addition to the similar and sometimes even larger improvements in over-ground gait and disease severity when compared to traditional treadmill training. Because perturbation training relies on wearing a harness during practice, it has potential for a wide segment of the PD population, 48 including those with cognitive decline and compromised balance. However, before wide clinical implementation is recommended, further study is needed to address whether an add-on off-treadmill training can facilitate the transfer of adaptive skills, and whether SBT is useful for more cognitively impaired patients. Furthermore, transfer to daily-life ambulation and unsupervised contexts are yet to be demonstrated. 49 The present results need to be interpreted with some caution as participants and trainers were not blinded and the same trainers were involved in delivering SBT and TBT. However, we minimized performance bias by using structured progression levels in each training arm to ensure equal compliance. Participants were unaware of whether they were receiving the experimental treadmill training or not. Finally, the sub-analysis in freezers was not preregistered and because of the small sample (n = 22) inadequately powered for the primary outcome.

Conclusion
This study has shown that people with PD learn and retain the gait adaptation skills to overcome asymmetric gait-speed perturbations on a treadmill remarkably well, but seem unable to generalize these skills to turning offtreadmill. The current findings warrant further investigation into whether an enriched SBT-training targeted to enhance transfer to over-ground turning, will eventually translate into better mobility in daily life.

Supporting Data
Additional Supporting Information may be found in the online version of this article at the publisher's web-site.