Factors Influencing Neuromuscular Responses to Gait Training with a Robotic Ankle Exoskeleton in Cerebral Palsy

Abstract

A current limitation in the development of robotic gait training interventions is understanding the factors that predict responses to treatment. The purpose of this study was to explore the application of an interpretable machine learning method, Bayesian Additive Regression Trees (BART), to identify factors influencing neuromuscular responses to a resistive ankle exoskeleton in individuals with cerebral palsy (CP). Eight individuals with CP (GMFCS levels I – III, ages 12 – 18 years) walked with a resistive ankle exoskeleton over seven visits while we measured soleus activation. A BART model was developed using a predictor set of kinematic, device, study, and participant metrics that were hypothesized to influence soleus activation. The model (R² = 0.94) found that kinematics had the largest influence on soleus activation, but the magnitude of exoskeleton resistance, amount of gait training practice with the device, and participant-level parameters also had substantial effects. To optimize neuromuscular engagement during exoskeleton training in individuals with CP, our analysis highlights the importance of monitoring the user’s kinematic response, in particular, peak stance phase hip flexion and ankle dorsiflexion. We demonstrate the utility of machine learning techniques for enhancing our understanding of robotic gait training outcomes, seeking to improve the efficacy of future interventions.

Introduction

The field of neurorehabilitation has seen significant strides in the development of robotic devices for improving mobility and augmenting rehabilitation for individuals with neuromuscular injuries (Beretta et al., 2020; Mehrholz, Thomas, Kugler, Pohl, & Elsner, 2020; Nam et al., 2017). Many of these devices work by guiding or resisting limbs through specific phases of the gait cycle in an attempt to re-train important spinal pathways (Nam et al., 2017) or increase supraspinal drive (Winchester et al., 2005). Specifically, robotic gait training interventions have demonstrated promising outcomes for individuals with cerebral palsy (CP) (Conner, Remec, Orum, Frank, & Lerner, 2020; Kang et al., 2017; Wu, Kim, Gaebler-Spira, Schmit, & Arora, 2017), a pediatric-onset movement disorder characterized by progressive declines in mobility into adulthood (Graham et al., 2016).

While the demonstrated benefits of recent robotic gait training interventions are promising for augmenting rehabilitation effects and improving plantarflexor control (Conner, Schwartz, & Lerner, 2021) in CP, there are current limitations to their implementation. CP is a broad diagnosis, encompassing any injury or insult to the developing brain, which can vary in timing, location, and severity (Graham et al., 2016). Severity of walking disability in CP is typically linked to reduced complexity of neuromuscular control of gait, as measured by the total variance accounted for by one muscle synergy, with greater co-activation across the hamstrings and gastrocnemius in early stance, and rectus femoris and anterior tibialis during late stance (Steele, Rozumalski, & Schwartz, 2015). Additionally, robotic interventions have a range of modifiable parameters (e.g., level of difficulty, walking speed, etc.) that can be tuned based on user ability. These factors can compound one another, making it difficult to determine which device parameters are ideal for each participant. It may be beneficial to have a method for evaluating the factors, and their relative importance, that are influencing the response to robotic gait training, with the goal of optimizing intervention efficacy in clinical settings (Conner, Remec, & Lerner, 2022).

There has been recent and rapid growth in the use of machine learning to enhance diagnosis and treatment of individuals with neuromuscular disorders (Halilaj et al., 2018). Support vector machines and artificial neural networks have seen broad adoption, as they can model non-linear relationships from the diverse predictor sets inherent in gait analysis; such algorithms have been used in CP specifically to predict patient outcomes following surgical intervention and device prescription (Rajagopal et al., 2018; Ries, Novacheck, & Schwartz, 2014). One algorithm that has demonstrated promising performance compared to existing techniques, but has so far seen limited application in biomechanics research is Bayesian Additive Regression Trees (BART) (Dorie, Hill, Shalit, Scott, & Cervone, 2019). BART is a ‘sum of trees’ regression method used for non-parametric function estimation, similar to boosting (Freund & Schapire, 1997), and random forest (Breiman, 2001) algorithms. Although, unlike these other ensemble algorithms, BART uses Bayesian methods to constrain tree parameters and prevent data overfitting, resulting in models that have low levels of bias and variance (Dorie et al., 2019). This may make BART particularly advantageous in understanding robotic gait training, where complex and multifactorial user-device interactions are expected. However, BART has not been previously used to this end.

The purpose of this study was to evaluate the factors influencing the neuromuscular response to gait training with a resistive ankle exoskeleton and demonstrate the general application of machine learning techniques for evaluating which factors are important to consider for improving the response to robotic gait training interventions. We hypothesized that the neuromuscular response would be most influenced by the level of resistance applied by the device, followed by the amount of gait training practice (i.e., visit number), and the functional level of the individual (i.e., gross motor function classification system (GMFCS) level). We also evaluated walking kinematics and measures of motor control to determine if any compensatory gait mechanics were influencing the neuromuscular response.

Methods

The protocol used in this study was approved by the Northern Arizona University Institutional Review Board (#986744) and utilized participants recruited for a clinical trial that was prospectively registered at ClinicalTrials.gov (NCT04119063). After detailing the nature and key elements of involvement in the study, informed written consent was received by all participants 18 years or older, and by parents or legal guardians of participants under 18 years of age (minors also provided verbal assent).

Participants

Individuals with CP between the ages of 10 to 18 years were recruited from the local area for participation in this study. Inclusion criteria was a confirmed diagnosis of CP, GMFCS levels I-III, the ability to walk on a treadmill with or without support for at least 10 minutes, and the ability to understand and follow verbal directions. Exclusion criteria was orthopedic surgery within the past six months unless approved by the participant’s orthopedic surgeon, a botulinum toxin injection to the lower limbs within the past six months, and any other condition that would preclude safe participation in the study. Eight total participants were recruited (see Table 1 for participant characteristics). Participants were evaluated by a licensed physical therapist who measured anthropometrics and spasticity level at the ankle (dorsi/plantar flexion) and knee (flexion/extension) using a Modified Ashworth Scale. The Modified Ashworth Scale is a 5-point scale commonly used in clinical evaluation where (1) indicates no increase in tone and (5) indicates full rigidity.

Table 1.

Participant characteristics

	Gender	Age (yrs)	Height (m)	Weight (kg)	GMFCS level*	CP type*	Walking speed (non- dimensional)*
P1	M	18	1.76	62.1	II	SD	0.31
P2	F	12	1.45	42.4	II	SD	0.28
P3	M	13	1.58	42.2	III	SD	0.22
P4	M	15	1.65	60.8	II	SD	0.25
P5	M	13	1.49	39.0	II	SD	0.30
P6	M	12	1.43	39.5	I	SH	0.31
P7	M	15	1.56	69.0	III	SD	0.23
P8	M	16	1.65	65.8	I	SH	0.27

GMFCS level: Gross Motor Function Classification System level

CP type: spastic hemiplegic (SH), spastic diplegic (SD)

Walking speed: non-dimensionalized according to (Hof, 2018)

Study visit allocation

Participants completed seven visits using the resistive ankle exoskeleton (Fig. 1A), including “assessment” visits to record the neuromuscular responses across different resistive torque magnitudes (Fig. 1B) before and after four “gait training practice” visits (Fig. 1C) designed to explore how repeated, short-term exposure may influence the neuromuscular response. First, participants completed an initial “assessment” session (Visit 1), which was followed by the four gait training practice sessions (Visits 2-5). Following the gait training practice sessions, participants completed a post-practice assessment (Visit 6), and a two-week follow-up assessment (Visit 7). Visits 1 – 6 took place over a two-week period, with 24 – 72 hours between visits. Visit 7 was completed two weeks after visit 6.

Figure 1.

A) Ankle exoskeleton device worn by all participants, consisting of a motor and control assembly worn at the waist and bilateral ankle assemblies; B) Assessment visits completed on visits 1, 6, and 7, whereby lower limb kinematics and muscle activity were collected under four conditions: 1) Baseline (device powered off, foot plates disconnected), 2) Low resistance (0.10 Nm/kg), 3) Moderate resistance (0.15 Nm/kg), and 4) High resistance (0.20 Nm/kg) (order of resisted trials was randomized); C) Gait training practice visits completed on visits 2 – 5, where participants walked for 20 minutes/visit at progressively increasing levels of resistance and with soleus audiovisual (AV) biofeedback.

Resistive ankle exoskeleton device

This study utilized an untethered, battery-powered ankle exoskeleton, the details of which can be found in (Conner, Luque, & Lerner, 2020; Orekhov, Fang, Cuddeback, & Lerner, 2021) (Fig. 1A). Briefly, this device consisted of an assembly worn at the waist and bilateral ankle assemblies. The motor and control assembly actuated the ankle pulleys via Bowden cables, allowing the device to resist ankle plantar flexion. The research team was able to wirelessly control the level of resistance delivered by the device via Bluetooth and a custom MATLAB graphical user interface.

During the gait training practice visits only (see “Gait training pracice” section below), participants walked with the resistive ankle exoskeleton device while also receiving audiovisual biofeedback of their soleus muscle from their more affected limb. This biofeedback system displayed a user’s soleus surface electromyography (sEMG) signal, smoothed via an 80 ms moving average filter, alongside a horizontal target line. A tone was played whenever this target line was surpassed. The sEMG biofeedback target was programmed to automatically adjust higher or lower so that the 10-stride success rate was between 50 – 75%.

Assessments (visits 1, 6, and 7)

For each assessment visit, bilateral sEMG sensors (Noraxon, Scottsdale, AZ, USA; 1000 Hz) were placed on the soleus, tibialis anterior, vastus lateralis, and semitendinosus according to SENIAM recommendations. Participants were also outfitted with reflective markers in accordance with Vicon’s Lower Limb Plugin Gait Model for measurement of three-dimensional lower limb kinematics (Vicon, Denver, CO, USA; 100 Hz). To assess baseline function on the first visit, participants performed barefoot overground walking trials at a self-selected speed until six complete gait cycles on their more affected side were recorded.

For each assessment session, participants walked for two minutes at a self-selected speed on a treadmill during a baseline condition where foot plates were detached from the ankle pulley system to prevent back drive from the motor assembly. Participants then completed two-minute walking trails at three different ankle exoskeleton resistance levels: 1) 0.1 Nm/kg (i.e., “low” resistance), 0.15 Nm/kg (i.e., “moderate” resistance), and 0.2 Nm/kg (i.e., “high” resistance); delivery was prescribed in a way to evenly distribute the condition order across participants. These resistance levels were chosen based on a pilot study that utilized this device; these levels were feasible and effective (Conner, Remec, et al., 2020; Conner et al., 2021). Muscle activity and lower limb kinematics were recorded during the final minute of each two-minute trial to allow time for each participant to acclimate to each resistance level before capturing their biomechanical responses. Treadmill speed was matched across all visits within a participant. All participants were instructed to hold onto the treadmill hand supports while walking for safety, regardless of GMFCS level.

Gait training practice (visits 2 – 5)

During gait training practice visits, participants walked with ankle exoskeleton resistance and audiovisual biofeedback for two 10-minute treadmill walking bouts, separated by a five-minute seated break; walking speed was kept constant across all visits. Participants started at 0.1 Nm/kg of nominal peak ankle resistance, which was increased to 0.125 Nm/kg after the first 10 minute round. On visit 3, participants started at 0.125 Nm/kg, which was again increased by 0.025 Nm/kg to 0.15 Nm/kg after the first 10 minute round. This progression was used across gait training practice visits so that 0.20 Nm/kg was reached by the final visit.

Data analysis

Ten consecutive gait cycles on a participant’s more affected side were evaluated from each assessment visit. Using a custom MATLAB script (MathWorks, Natick, MA, USA; v2019a), EMG data were demeaned, high pass filtered (4^th order Butterworth, 20 Hz), full wave rectified, low pass filtered (4^th order Butterworth, 10 Hz), and down-sampled to align with kinematics. Muscle activation curves were normalized within each session to the peak activation of each respective muscle from the baseline condition. To account for potential slight differences in electrode placement and skin conductivity between visits, the relative within-visit changes in EMG between resisted and baseline walking were used for comparison across visits. Data were segmented into individual gait cycles, defined as heel strike to consecutive heel strike. Kinematic data were low pass filtered (4^th order Butterworth, 6 Hz) and joint angles at the ankle, knee, and hip for individual gait cycles were calculated via inverse kinematics using Vicon’s Plugin Gait Model (Vicon, Denver, CO, USA). We computed total variance accounted for by one synergy (tVAF1), which has been shown to be a robust and repeatable summary measure of motor complexity in cerebral palsy that, importantly, is related to walking function (Shuman, Schwartz, & Steele, 2017) and treatment outcomes (Shuman, Goudriaan, Desloovere, Schwartz, & Steele, 2018).

Bayesian Additive Regression Trees (BART)

A BART model was developed to identify factors which were predictive of soleus activity during walking with an ankle exoskeleton. The predictor set included kinematic, device, study, and participant metrics that could influence soleus modulation. All input variables for the BART model are outlined in Table 2. Details on how the spasticity score, tVAF₁ variables, and kinematic variables were calculated can be found in Supplementary Material.

Table 2.

Model variables

Response	Description
Mean stance-phase soleus activity	Z-score of stance-phase soleus activity, normalized to baseline treadmill walking for each participant
Predictors	Description
Visit number	Numerical value of 1, 6, and 7 (assessment visits only)
GMFCS	Numerical value from 1-3, indicating the participan’s GMFCS level. Level determined by a licensed physical therapist during a physical exam
Speed	Non-dimensional walking speed, normalized to participant leg length
Spasticity score	Summary score of four lower-limb modified Ashworth scores, derived using principle component analysis
Δ tVAF1	Difference in motor control complexity between baseline treadmill walking and walking with the ankle exoskeleton
Baseline tVAF1	Measure of motor control complexity during overground barefoot walking
Exoskeleton resistance level	Numerical value from 1-3, corresponding to low (0.1 Nm/kg), moderate (0.15 Nm/kg), and high (0.2 Nm/kg) resistance levels
Kinematic parameters	Z-score values for sagittal plane kinematic parameters at the hip, knee, and ankle (see Section 2.4), normalized to baseline treadmill walking for each participant

Model validation

BART model development and hyperparameter tuning were performed using 10-fold cross-validation (parameters: k = 2, q = 0.9, nu = 3, num_trees = 200, seed = 50) in RStudio (RStudioTeam, 2020; bartMachine package) (Kapelner & Bleich, 2016). Pseudo R² and out-of-sample root-mean-squared error (RMSE) are reported as measures of model quality.

Model interpretation

Model outputs were interpreted using accumulated local effect plots (ALE) (Apley & Zhu, 2020). ALE plots illustrate the effect that individual predictors have on the response variable, conditioned on all other model covariates, making them useful to understand non-linear relationships (e.g., discontinuities, parabolic trends). ALE plots are advantageous to other visualization techniques, such as partial dependence plots (PD), as they quantify local rather than marginal effects and are, therefore, unbiased to highly correlated predictors (Apley & Zhu, 2020). For each predictor, an ALE plot was generated using a bootstrap procedure. This involved drawing samples of the original data set with replacement (n = 50) to generate a series of ALE plots from which the average trend (± 1SD) could be derived (Apley, 2018). As a secondary smoothing operation, the predictor data were then binned in evenly spaced increments over its entire range. Using the resultant plots, the net effect of each predictor was approximated as the difference between the 90^th and 10^th percentile of the response (Fig. 2). All effects were then ranked by magnitude to identify those predictors which had the largest overall influence on soleus activity. If soleus modulation was most sensitive to device resistance, GMFCS level, and gait training practice, as hypothesized, we would expect that the net effect of these variables in the BART model would be larger than other potential predictors. Predictors were calculated relative to an individual’s baseline values (see Supplementary Material for more detail) so that interpretation of results could be implemented into practice. For example, if a therapist was evaluating how a patient walked with the resistive device, it would be more feasible to assess and coach a patient’s walking mechanics relative to their typical walking pattern (i.e., more or less crouched) than to coach an individual to a specific parameter (i.e., attaining at least 20 degrees of plantar flexion at push-off).

Figure 2. Overview of the processing pipeline to generate the BART model and ALE plots.

BART model development and hyperparameter tuning were performed using 10-fold cross validation with the bartMachineCV package (Kapelner & Bleich, 2016). Following model development, ALE plots were generated from bootstrapped samples (Z¹ to Z⁵⁰) of the original predictor and response data sets (Table 2) and the average ALE plot (± 1SD) was identified (Apley, 2018). The net effect, representing the difference between the 90^th and 10^th percentile of the response over the range of the predictor was then quantified for each predictor.

Prior to running our BART model, an outlier analysis was performed on our primary outcome variable (mean stance phase soleus activation). Outliers were defined as values that were greater than 1.5 times the interquartile range above the third quantile or less than 1.5 times the interquartile range below the first quantile (Tukey, 1977). It was found that P8’s visit 6 and 7 mean stance phase soleus activation z-scores met the definition of outliers. For this reason, P8 was completely removed from the BART model. Additionally, P3’s visit 7 was not included in the analysis because EMG data was not available for this visit due to a dropped signal from the electrodes.

Results

The BART model (R² = 0.94; RMSE = 1.15) identified that kinematic parameters had the largest net effect on soleus modulation, appearing as four of the top ten predictors considered (Fig. 3). Specifically, peak stance phase hip flexion (net effect = 1.25) and ankle dorsiflexion (net effect = 1.01) had the largest overall effect, with a decreased peak for both variables resulting in greater stance phase soleus activation. Knee flexion at heel strike (net effect = 0.75) and peak stance phase hip extension (net effect = 0.57) also influenced soleus activation (Fig. 4).

Figure 3.

Net effects for each parameter when controlling for all other covariates, grouped by kinematic parameters (blue), motor control and functional status parameters (dark blue), and testing and device parameters (grey).

Figure 4.

Kinematic ALE plots, displaying the accumulated local effects of peak stance-phase hip flexion, peak stance-phase ankle dorsiflexion, knee angle at heel strike, and peak stance-phase knee extension (all relative to baseline) on mean stance-phase soleus activation. Vertical bars indicate ± 1 standard deviation and the size of the individual points represent the number of samples in each bin.

Both the change in motor control complexity (Δ tVAF₁; net effect = 0.83) and baseline motor control complexity (tVAF₁; net effect = 0.65) appeared to affect soleus activation level, whereby an increase in tVAF₁ with resistance and higher baseline tVAF₁, representing a decrease in motor control complexity, were associated with greater activation. Additionally, greater spasticity (net effect = 0.83) and higher functional level (i.e., lower GMFCS level; net effect = 0.68) were associated with greater soleus activation (Fig. 5). Gait training practice with the device (i.e., visit number) had a net effect of 0.73, with increased soleus activation after gait training practice to the device. Notably, soleus activation decreased with the two week follow up visit (i.e., visit 7), but stayed above the visit 1 activation level, indicating that participants retained some adaptation to the device after the two-week break. There appeared to be a relatively linear relationship with the magnitude of applied resistance by the device (net effect = 0.70), with higher resistance levels resulting in greater soleus activation (Fig. 6).

Figure 5.

Motor control and functional status ALE plots, displaying the accumulated local effects of change in motor control complexity with resistance (Δ tVAF₁), spasticity score, GMFCS level, and baseline motor control complexity (tVAF₁) on mean stance-phase soleus activation. For both complexity measures, a larger number indicates less complex control. A larger spasticity score indicates a higher amount of spasticity across the ankle and knee. Vertical bars indicate ± 1 standard deviation.

Figure 6.

Testing and device parameter ALE plots, displaying the accumulated local effects of gait training practice (visit number) and magnitude of exoskeleton resistance on mean stance-phase soleus activation. Vertical bars indicate ± 1 standard deviation.

Discussion

The purpose of this study was to evaluate the factors that contributed to the neuromuscular response during gait training with a resistive ankle exoskeleton in a cohort of children and young adults with CP. The findings from our BART model partially supported our hypothesis that the neuromuscular response (i.e., mean stance phase soleus activation) to the resistive robotic exoskeleton would be influenced by the device’s resistance level, amount of gait training practice, and a participant’s functional walking status (i.e., GMFCS level). While these factors fell within the top ten of net effects, we generally found that gait kinematics had a larger effect on soleus activation with the tested device. Specifically, it was found that peak stance phase hip flexion and ankle dorsiflexion had the largest effect on soleus activation, with decreased values relative to baseline resulting in increased activation. Overall, we found that a relative increase in lower-limb extension during the stance phase of gait resulted in greater recruitment of the ankle plantar flexors. This is in line with the finding that a crouched (more flexed) posture can reduce muscle force output of the lower limbs (Hicks, Schwartz, Arnold, & Delp, 2008). We also found that those individuals who decreased their motor control complexity (i.e., increased tVAF₁ from baseline, unresisted walking) when walking with resistance had greater overall soleus activation. This may indicate that some participants adopted an undesirable activation strategy that was greater in magnitude, but less complex and characterized by higher levels of ankle co-activation between the soleus and tibialis anterior through stance phase, which would raise tVAF₁.

The ALE plots allowed us to isolate the effect of resistance level from other variables associated with soleus activation, such as altered kinematics, spasticity level, and GMFCS level. Our findings indicated soleus activation generally increased with higher magnitudes of resistance, but the actual net effect was smaller than anticipated. In a previous study with this resistive ankle exoskeleton, participants were coached on how to properly engage with the device (Conner, Luque, et al., 2020), as had been done in previous robotic gait training studies (Drużbicki et al., 2013; Smania et al., 2011; Wallard, Dietrich, Kerlirzin, & Bredin, 2017). In the present study, however, verbal coaching was not implemented throughout the trial to avoid confounding effects, which may have allowed for compensatory walking strategies. As a result, instead of increasing soleus activation to overcome a larger magnitude of resistance, it appears that some participants may have altered their walking mechanics in a way that allowed them to bypass this demand. These findings highlight the importance of monitoring compensatory strategies when training with robotic gait training devices to ensure proper mechanics and neuromuscular engagement. While the compensatory strategies were reflected in the kinematic data, given that the number of simultaneous performance targets is limited before overwhelming to the user, our results suggest that verbal feedback from a clinician who can quickly and intuitively observe kinematic compensations may be critical for effective robotic training in this patient population. A recommendation for clinicians working with patients during gait training would be to encourage a more upright, less crouched posture when engaging with ankle exoskeleton resistance.

In addition to better understanding an intervention with early evidence of promising benefits for individuals with CP, this study aimed to illustrate how non-linear supervised learning, specifically BART, may increase the efficacy of robotic rehabilitation interventions for heterogeneous patient populations. Various algorithms have been previously used in neuromuscular disorders research to classify gait impairments or predict response to existing intervention (Halilaj et al., 2018). However, as was highlighted in this study, such algorithms may be useful in robotic rehabilitation applications, not only for identifying optimal candidates but also guiding device parameter tuning and training structure. A particular advantage of BART, beyond its predictive capacity, is that it requires very little (if any) hyperparameter tuning and is less computationally expensive than other high-performing paradigms, both of which make it more feasible for a broad user base to implement reliably (Chipman, George, & McCulloch, 2010). Further, BART is capable of handling missing predictor variables and multiple variable types, as are expected in experimental gait analysis (Kapelner & Bleich, 2015).

A limitation of this study was that we evaluated a relatively small sample size of individuals with CP, and given the heterogeneity of this patient population, our findings cannot be applied to all individuals with CP. This limitation further highlights the opportunity for BART to help interpret heterogeneous responses to robotic gait training interventions in largely variable patient populations. Still, the sensitivity of the outcomes reported in this study to different pathological gait patterns (e.g., crouch gait) should be confirmed. Also, factors we did not evaluate may also contribute to soleus activation with exoskeleton resistance, such as baseline plantar flexor strength and gait kinetics.

Conclusion

This study determined the primary factors and their relative influence for more effective neuromuscular engagement during gait training with a resistive ankle exoskeleton in individuals with CP. Our findings underscored the importance of monitoring how users change their gait kinematics when walking with the resistive device, with a specific emphasis on stance-phase lower limb extension. We also highlight the necessity of considering an individual’s functional status and amount of practice with the device, as well as more obvious factors, like device parameters. BART can be used early in the development of robotic gait training interventions to better understand complex and multifactorial user-device interactions. In doing so, device and intervention parameters can then be tuned for future, large scale studies.

Data availability statement for basic data sharing policy

The data collected and analyzed in this study are available from the corresponding author upon reasonable request.