Abstract
Ankle exoskeletons alter wholebody walking mechanics, energetics, and stability by altering centerofmass (CoM) motion. Controlling the dynamics governing CoM motion is, therefore, critical for maintaining efficient and stable gait. However, how CoM dynamics change with ankle exoskeletons is unknown, and how to optimally model individualspecific CoM dynamics, especially in individuals with neurological injuries, remains a challenge. Here, we evaluated individualspecific changes in CoM dynamics in unimpaired adults and one individual with poststroke hemiparesis while walking in shoesonly and with zerostiffness and highstiffness passive ankle exoskeletons. To identify optimal sets of physically interpretable mechanisms describing CoM dynamics, termed template signatures, we leveraged hybrid sparse identification of nonlinear dynamics (HybridSINDy), an equationfree datadriven method for inferring sparse hybrid dynamics from a library of candidate functional forms. In unimpaired adults, HybridSINDy automatically identified springloaded inverted pendulumlike template signatures, which did not change with exoskeletons (p > 0.16), except for small changes in leg resting length (p < 0.001). Conversely, poststroke pareticleg rotary stiffness mechanisms increased by 37–50% with zerostiffness exoskeletons. While unimpaired CoM dynamics appear robust to passive ankle exoskeletons, how neurological injuries alter exoskeleton impacts on CoM dynamics merits further investigation. Our findings support HybridSINDy’s potential to discover mechanisms describing individualspecific CoM dynamics with assistive devices.
Similar content being viewed by others
Introduction
Ankle exoskeletons are used to improve walking function and gait mechanics^{1,2,3}. Personalized passive and powered ankle exoskeletons have shown promise to improve walking function and muscle coordination in some unimpaired adults and individuals with neurological injuries^{2,4,5}. However, changes in gait in response to ankle exoskeletons are highly individualized, especially following neurological injury, making device personalization critical to improving mobility. For example, in stroke survivors and children with cerebral palsy, ankle exoskeletons elicit diverse—and sometimes detrimental—impacts on gait mechanics, walking speed, step length, and the energetic cost of walking^{1,2,3,4}. Quantifying and characterizing these responses remain challenging and hinders clinicians’ and designers’ abilities to customize exoskeletons to support walking function.
Characterizing the changes in the neural and biomechanical processes governing centerofmass (CoM) motion (i.e., CoM dynamics) with ankle exoskeletons may help explain heterogeneous exoskeleton impacts on tasklevel goals during walking. Despite observed changes in centerofmass (CoM) mechanics and energetics with ankle exoskeletons, little is known about how CoM dynamics change with ankle exoskeletons to achieve tasklevel goals, like walking stably and efficiently^{2,5,6,7,8,9,10}. CoM energetics are altered in unimpaired adults walking with ankle exoskeletons and following neurological injuries (e.g., poststroke, the paretic leg exhibits reduced power generation and changes in leg power with exoskeletons differ between individuals)^{2,5,11}. However, whether these changes in leg power are accompanied by changes in CoM dynamics or simply altered CoM kinematics is unclear.
Reducedorder representations of CoM dynamics, often termed template models, provide a foundation to quantify complex exoskeleton responses using interpretable mechanical elements^{12,13,14,15,16,17}. For example, Full & Kodistchek (1999) used template models of CoM dynamics, such as the springloaded inverted pendulum (SLIP), to quantify strategies to stabilize the CoM in response to perturbations^{15}. Such reducedorder representations of gait encode neural and biomechanical dynamics using a minimalist set of physicsbased mechanisms. Specifically, common template models of CoM dynamics use a variety of mechanisms, such as SLIP leg springs or the rigid legs of an inverted pendulum walker, to describe relationships between leg kinematics and CoM accelerations during gait^{10,13,15,16,18,19,20,21,22}. Each mechanism within a model, therefore, encodes a hypothesis about how neural and biomechanical subsystems interact to achieve tasklevel walking goals.
However, determining which template mechanisms are needed to optimally describe an individual’s CoM dynamics remains challenging. Inverted pendulum templates have been useful in modeling CoM energetics, the transition from walking to running^{16,18}, and strategies for energetically efficient CoM acceleration^{8,22,23} and lateral stabilization^{10,24}. Higherdimensional template structures, such as the bipedal SLIP, were needed to more accurately predict sagittalplane ground reaction forces (GRFs)^{20,25}. Additional mechanisms applied to the bipedal SLIP template, such as leg dampers, rotary springs, or curved feet, have been used to further improve the accuracy of CoM dynamics models during walking^{19,26}. This breadth of templates proposed for human walking suggests that the mechanisms that best describe gait are individualspecific^{20,26}. Selecting individualspecific template structures (i.e., the mechanisms included in the template) is, therefore, critical to quantifying exoskeleton impacts on CoM dynamics using template models.
To emphasize the individualspecific nature of templates, we denote the combination of mechanisms best describing individualspecific CoM dynamics as a template signature. Inter or intraindividual differences in the template signature mechanisms that best describe CoM dynamics for an individual and walking condition, or the coefficients estimated for each mechanism, may provide insight into how exoskeletons impact CoM dynamics. For example, template signatures in children with hemiparetic cerebral palsy differed from typically developing children and were asymmetric, with increased stiffness—defined by the coefficient of the stiffness mechanism—in the paretic leg^{27,28}. Atypical and asymmetric CoM dynamics suggested that, following neurological injury, people may adopt individual and legspecific strategies to accelerate the CoM.
However, characterizing changes in template signatures with exoskeletons or following neurological injury requires addressing a major methodological challenge: Manually identifying optimal template signature structure is a slow, ad hoc process that relies on firstprinciples knowledge of the system^{19}. Using this manual approach, comprehensively comparing candidate template signatures for even a moderate number of template mechanisms is challenging: the number of comparisons increases combinatorially with the number of candidate mechanisms. New approaches are needed to select mechanisms rapidly and systematically from a literaturebased library of candidate mechanisms.
Recent advances in datadriven modeling and machine learning provide powerful tools to identify template signatures from walking data^{29,30,31,32}. One such algorithm, HybridSINDy (SINDy: Sparse identification of nonlinear dynamics^{29}), identifies sparse nonlinear dynamics in hybrid systems from timeseries data, making it particularly appropriate for identifying template models of walking, which have distinct dynamics based on foot contact configuration^{28,31}. HybridSINDy automatically identifies and compares a large number of candidate dynamical models (e.g., template signatures) from an arbitrary library of possible functional forms (i.e., mechanisms). The algorithm uses information criteria to determine the relative plausibility of each candidate model and selects only those that are highly plausible (i.e., that are both parsimonious and highly representative of the system)^{33}. When applied to human walking data this approach will, therefore, enable rapid, systematic identification of individualspecific template signatures.
The purpose of this study was to identify changes in templatesignaturebased representations of CoM dynamics in response to ankle exoskeletons. We used the HybridSINDy algorithm to identify physically interpretable, lowdimensional template signatures describing CoM dynamics during walking in unimpaired adults and evaluated how template signature coefficients changed with hinged and stiff ankle exoskeletons. We hypothesized that the addition of ankle exoskeleton frame and stiffness would alter template signature coefficients. Additionally, to examine the potential of HybridSINDybased template signatures to reveal changes in CoM dynamics with ankle exoskeletons for individuals with neurological injuries, we present a case study evaluating altered template signatures in one individual with poststroke hemiparesis.
Methods
Data collection
We collected 3D marker trajectories using a tencamera motion capture system (Qualisys AB, Gothenburg, SE) and GRFs using an instrumented treadmill (Bertec Inc., Columbus, USA) in twelve unimpaired adults (6 M/6F; age = 23.9 ± 1.8 years; height = 1.69 ± 0.10 m; mass = 66.5 ± 11.7 kg) and one stroke survivor with lowerlimb hemiparesis (sex not disclosed; age = 24 years; height = 1.70 m; mass = 68.0 kg). Participants walked at their selfselected steadystate speed on a fixedspeed treadmill in shoesonly and with bilateral passive ankle exoskeletons under two conditions: zero resistance to ankle flexion (i.e., zero stiffness; K_{0}) and high dorsiflexion resistance (i.e., high stiffness; K_{H} = 5.1 Nm/deg; Fig. 1). The order of walking conditions was randomized. A detailed description of the acclimatization protocol and data preprocessing can be found in Ref.^{34}. Briefly, in a second session following a practice session, data were collected while participants walked at their selfselected speed—determined in the practice session—for six minutes per condition, including two minutes to acclimate to the treadmill before data were recorded. Only the third and fourth minutes of data were used in this study. To mitigate fatigue, the poststroke participant walked under the same protocol, but for only three minutes per condition^{35}. This study was approved by the University of Washington Institutional Review Board (#47744). The study was performed in accordance with the approved protocol and University of Washington Institutional Review Board guidelines and regulations. All participants provided written informed consent prior to participating in the study.
Estimating template signatures with HybridSINDy
We used the HybridSINDy algorithm to identify template signatures during walking with and without ankle exoskeletons, separately for each walking condition.
Kinematic variable extraction
For each exoskeleton condition, we used OpenSim’s Body Kinematics algorithm to estimate the CoM and foot positions^{36,37}. In this section, we describe the SINDy and HybridSINDy algorithms in the context of identifying template signatures, while more detailed explanations can be found in Refs.^{29}^{,}^{31}. The 3D CoM accelerations, \(\ddot{{\varvec{q}}}\left(t\right)\in {\mathbb{R}}^{m\times n}\) (Fig. 1B), were described by continuoustime nonlinear dynamics, \(f\left(q\left(t\right), \dot{q}\left(t\right)\right)\), where m denotes the number of samples and n = 3 denotes the output variables:
where time is denoted by \(t\in {\mathbb{R}}^{m\times 1}\), and \(q(t)\) and \(\dot{q}(t)\) represent CoM positions and velocities relative to the feet, respectively, in \({\mathbb{R}}^{m\times n}\), in the anterior–posterior, vertical, and mediolateral directions. We assume that only a small number of functional forms (i.e., mechanisms) in \(f(q\left(t\right), \dot{q}(t))\) describe most of the system’s behavior. We omit the time notation in the remaining sections.
Sparse identification of nonlinear dynamics (SINDy)
The SINDy algorithm^{29} recovers sparse nonlinear dynamics from a library of candidate functional forms, which may consist of arbitrary nonlinear functions of system measurements. Adopting the notation from Ref.^{31}, we can rewrite the dynamics in Eq. (1) as:
where \({\varvec{\Xi}}\in {\mathbb{R}}^{p\times n}\), is a linear map from nonlinear function library encoding mechanisms that may be useful in describing CoM dynamics, \({\varvec{\Theta}}({\varvec{q}},\dot{{\varvec{q}}})\in {\mathbb{R}}^{m\times p}\), to CoM accelerations, \(\ddot{{\varvec{q}}}\). The coefficients in \({\varvec{\Xi}}\), therefore, describe how each template signature mechanism accelerated the CoM, with nonzero coefficients denoting the active terms that define the template signature structure. In the context of the template model investigated here, nonzero coefficients denote the mechanisms included in the model. We included p = 14 functional forms (mechanisms) in the function library (7 per leg), described below. The SINDy algorithm promotes sparsity in the model using sequential leastsquares regression with hard thresholding, with the threshold defined by the sparsity parameter, \(\lambda\) (Eq. (3))^{31}. This thresholding approach penalizes the zeronorm of \({\varvec{\Xi}}\) and solves:
HybridSINDy
HybridSINDy extends SINDy in two important ways. First, HybridSINDy uses clustering to generalize SINDy to hybrid systems. For human walking, clustering enables unique dynamics to be identified in each gait phase, defined by foot contact configuration (i.e., single and doublelimb support)^{9,14,18}. We replace the term hybrid regime, used in the original HybridSINDy manuscript, with gait phase for clarity in the context of human walking^{31}. Second, HybridSINDy uses information criteria to automatically select the system dynamics that best describe the data. This approach enables competing hypotheses about the mechanisms describing CoM acceleration to be rapidly and systematically compared, thereby highlighting mechanisms that are critical to describing CoM dynamics across individuals and mechanisms unique to a subset of individuals. Note that HybridSINDy does not compare entirely distinct sets of governing equations (e.g., SLIPlike template vs. a passive dynamic walker with knees^{12}). Rather, the algorithm selects which mechanisms should be included in a SLIPlike template model to best describe CoM dynamics.
Applying hybridSINDy to walking
We applied the HybridSINDy algorithm to human gait using the following steps for each participant and walking condition (outlined in Fig. 2; example results shown in Fig. 3). Note that within each gait phase, we expanded upon the original HybridSINDy algorithm by using multimodel inference to define a single template signature when multiple signatures were plausible (Step 5)^{31}.

1.
Clustering (Figs. 2, 3B): We used a clustering approach to increase robustness to measurement noise and identify frequently occurring template model structures^{31,38}. For the first 3600 samples (30 s; ~ 25–30 strides) in the training set (10,800 samples were available for clustering in each trial), we generated clusters of each sample’s 800 nearest neighbors and identified the centroid of each cluster (3600 clusters, total). These clusters were used to estimate template model coefficients (Step 2). During clustering, nearest neighbors were selected based on their continuous kinematic phase: the phase angle of the right/nonparetic leg angle and angular velocity relative to vertical (Fig. 1A; right)^{34,39}. We also used this phase variable to normalize stride progression. The last 3600 samples of each dataset were withheld from training and used during model evaluation and selection (Step 3).
Some clusters contained data in both single and doublelimb support phases. However, these clusters tended to have relatively large error during model evaluation, such that they would not be selected as plausible in Step 3, below^{31}. Further, we selected our cluster size to ensure that clusters were small enough to contain data from only one gait phase: the average cluster width (800 samples) spanned only 7.4% of the training data, smaller than the duration of doublelimb support (10–12%)^{40}.

2.
Model estimation (Figs. 2, 3B,C): For each training cluster, we used SINDy to estimate the coefficients of multiple template signatures by sweeping 40 sparsity threshold values, ranging logarithmically from 1–100% of the largest magnitude coefficient in the fulldimensional model in each cluster. This approach typically produced 5–15 unique signatures per cluster.

3.
Model evaluation and selection: Using the 3600 samples of heldout data, we evaluated the ability of each template signature to reconstruct CoM accelerations in the anterior–posterior, vertical, and mediolateral directions. We computed the average absolute reconstruction error of the heldout data over the gait cycle (Eq. (4)).
$${{\text{error}}}_{\uppsi }={\left{\varvec{\Theta}}\left({\varvec{q}},\dot{{\varvec{q}}}\right){\varvec{\Xi}}\ddot{\mathbf{q}}\right}_{\uppsi },$$(4)where \(\psi\) represents the continuous phase of the gait cycle, from 0 to 100% of a stride.
We selected template signatures based on two criteria: First, we discarded signatures that were identified in less than 1% of training clusters. Frequently occurring template signatures are more likely to be robust to measurement noise or stridetostride variability, making them better representations of an individual’s gait dynamics^{31}.
Second, for each gait phase—single and doublelimb support—we selected the frequently occurring template signatures that had the highest likelihood according to the Akaike Information Criterion (AIC)^{33,41}. The AIC is widely used to compare candidate representations of a system (e.g., template signatures) according to their number of free parameters and loglikelihood^{33}. According to the AIC, a candidate representation that has a lower AIC score than competing representations is considered the most plausible (i.e., best) candidate representation of the system. Adopting the formulation in Ref.^{31}, assuming that model errors are independently, identically, and normally distributed, the AIC can be written in terms of the number of free parameters, k, number of samples, \(\rho\), and the sum of squared residuals:
$$AIC=2k+\rho {\text{ln}}\left(\frac{{\sum }_{j=1}^{\uprho }{\sum }_{i=1}^{I}{{\left(\widehat{f}\left({\text{q}}\right)\dot{q}\right)}_{i,j})}^{2}}{\rho }\right),$$(5)where in the outer summation is over \(\rho\) = 3600 samples and the inner summation, is over the I = 3 output states.
The AIC favors parsimonious, highly representative models, which is ideal for identifying minimalist representations of gait dynamics. Like Mangan and colleagues^{31}, we used the AIC corrected for finite sample sizes (AICc):
$$AICc=AIC+\frac{2\left(k+1\right)\left(k+2\right)}{\rho k2}.$$(6)The correction term approaches zero as the number of samples, \(\rho\), increases. We then determined the relative plausibility of competing template signatures using their relative AICc score, \({\Delta AICc}_{j}\)^{31,33}:
$${\Delta AICc}_{j}=AIC{c}_{j}AIC{c}_{min},$$(7)where \({\mathrm{\Delta AICc}}_{{\text{j}}}\) represents the relative AICc score for the j^{th} model. AICc_{min} represents the AICc of the model with the lowest AICc among the models compared within a gait phase. The best model according to the relative AICc has a score of \({\Delta AICc}_{j}=0\) and all other models had higher scores. Burnham and Anderson noted that models with \({\Delta AICc}_{j}\le 2\) have substantial support, while \({\Delta AICc}_{j}>7\) have low support^{33}. We adopted the threshold of^{31}, deeming template signatures with \({\Delta AICc}_{j}\le 3\) to be plausible.

4.
Multimodel inference (Fig. 2): Since human gait dynamics are not strictly hybrid and template models are approximations of CoM dynamics, multiple template signature structures may be plausible in each gait phase. To construct a single template signature for each gait phase, we computed a weightedaverage signature using Akaike weights, \({\omega }_{j}\), where j is the j^{th} plausible model in the gait phase^{33}. Note that we performed multimodel inference separately for each gait phase. Akaike weights are defined as:
$${\omega }_{j}=\frac{{\text{exp}}\left(\frac{{\mathrm{\Delta AICc}}_{j}}{2}\right)}{{\sum }_{r=1}^{R}{\text{exp}}\left(\frac{{\Delta {\text{AICc}}}_{r}}{2}\right)},$$(8)where \({\text{exp}}\left(\frac{{\Delta AICc}_{j}}{2}\right)\) defines the likelihood of the j^{th} template signature given the observations^{33}. The denominator denotes the summation of exponentially scaled relative AIC scores over all R candidate models. This approach weighs each signature based on its likelihood relative to the other plausible signatures.

5.
Uncertainty estimation and model accuracy (Figs. 2, 3D): To evaluate the robustness of the template signatures to noise and stridetostride variations in the data, we performed 200 bootstrapped estimates of each template signature coefficient in each gait phase separately for each trial. Each bootstrapping iteration randomly selected 3600 samples to estimate template signature coefficients, with replacement. We quantified the robustness of each template signature coefficient to variability in the data using the coefficient of variation (CV) of each participant and condition^{27}. Template signature coefficients for each participant, condition, and gait phase, were defined by the mean of the bootstrapped estimates. Figure 3D shows the estimated coefficients of the bootstrapping procedure applied to a synthetic SLIP model (see Supplemental S2). Synthetic SLIP model parameters and simulation results (Fig. 3D; gray bars and trajectories, respectively) were used as ground truth to validate the algorithm’s ability to identify template models of walking (Supplemental S2).
Template signatures mechanisms and dynamics
To model threedimensional CoM dynamics during walking, we created a function library of candidate mechanisms based on prior literature (Fig. 3A; Table 1). We included leg springs^{13,19,20,21,25} and dampers^{26}, which produce force along the leg. Leg springs are common energetically conservative mechanisms used to describe walking and running dynamics and enable a doublelimb support phase. Leg dampers are less common, but have been used to capture nonconservative gait dynamics^{26}. We also included rotary springs^{13,19} and dampers in the sagittal and frontal planes, which enable forcing transverse to the leg axis. The addition of rotary springs has been shown to improve reconstructions of anterior–posterior GRFs in a bipedal SLIP^{19}. We did not identify rotary damping elements in prior literature but included them as candidate mechanisms describing nonconservative forcing transverse to the leg.
We included only passive mechanical elements in the mechanism library because these elements can be used in a hybrid modeling framework to approximate active control of walking, such as for lateral stabilization or to inject and dissipate energy^{19,42,43,44}. Active mechanisms or statebased controllers could produce more parsimonious models but would be more challenging to interpret.
The mechanisms selected by the HybridSINDy algorithm define the template signature structure, which describes characteristic strategies to accelerate the CoM. The identified template signature coefficients describe each mechanism’s contribution to CoM accelerations.
The dynamics of a threedimensional bipedal SLIP augmented with damping and rotary mechanisms may be written as
where M is body mass, \({\varvec{g}}\) is the gravity vector, \(\phi\) describes the traverseplane leg angle, and \(\theta\) describes the leg angle from vertical in the direction defined by \(\phi\) (Fig. 1C)^{19,25}. The summation represents the total force generated by the legs on the CoM. The leftmost brackets contain mechanisms that impart forces radially along the leg: \({k}_{L}\) is the leg stiffness, \(L\) is the instantaneous leg length (Fig. 1C), \({L}_{0}\) is the leg resting length, \({c}_{L}\) is the leg damping, \(\dot{L}\) is the instantaneous leg velocity. We henceforth denote \({L}_{0}\) as leg length for clarity. The middle bracket contains mechanisms that impart forces transverse to the leg axis in the sagittal plane: \({k}_{s}\) and \({c}_{s}\) are the sagittalplane rotary stiffness and damping, respectively. \({L}_{s}\) denotes the sagittalplane leg projection. Analogously in the rightmost brackets, \({k}_{f}\) and \({c}_{f}\) represent the frontalplane rotary stiffness and damping, respectively, and \({L}_{f}\) denotes the frontalplane leg projection. The derivation of system dynamics can be found in Supplemental S1.
Normalized template mechanisms
To account for interindividual differences in walking speed and body size during analysis, we normalized the template signatures (Table 1). Leg stiffness was normalized as in Refs.^{17,19,31}. Leg resting length was normalized to the measured leg length^{19,26}. Rotary stiffness was normalized according to^{19}. All damping terms were converted to damping ratios^{26}. The normalized leg, sagittalplane, and frontalplane stiffness mechanisms are denoted by \({\kappa }_{L}, {\kappa }_{s},\) and \({\kappa }_{f}\), respectively. The normalized leg, sagittalplane, and frontalplane damping mechanisms are denoted by \({\zeta }_{L}, {\zeta }_{s},\) and \({\zeta }_{f}\), respectively. Normalized leg length is denoted \({\widetilde{L}}_{0}\). We can rewrite Eq. (9) as a linear combination of our normalized coefficients and nonlinear transformations of our states:
The CoM position and velocity relative to the feet were used to compute candidate template signature states: leg lengths and lengthening velocities, sagittalplane leg angles and angular velocities relative to vertical, and frontalplane leg angles and angular velocities relative to vertical. The complete function library can be found in Supplemental S1.
Evaluating HybridSINDy’s ability to select template signatures
We evaluated the HybridSINDy algorithm’s ability to accurately identify walking dynamics in the presence of noise and an incomplete mechanism library (i.e., missing functional forms relative to the true system dynamics) using forward simulations of a bipedal SLIP walking model (example shown in Fig. 3)^{25}. The synthetic SLIP model analysis and results are described in Supplemental S2.
Identifying template signatures in human gait
To quantify how well template signatures captured COM dynamics, we computed coefficients of determination (r^{2}) between the measured CoM accelerations and those predicted by each participant’s template signatures, averaged over the anterior–posterior, vertical, and mediolateral directions.
To evaluate the extent to which each mechanism described CoM accelerations in unimpaired adults, we determined the proportion of participants for whom each template signature coefficient was selected and the average number of nonzero mechanisms in each gait phase. Template signature terms that are identified across individuals may represent mechanisms fundamental to CoM dynamics, while infrequently identified mechanisms may describe individualspecific features of CoM dynamics.
To determine if unimpaired CoM dynamics during shoesonly walking generalized to walking with ankle exoskeletons, we evaluated the ability of shoewalking template signature structures to reconstruct CoM accelerations in the K_{0} and K_{H} conditions. We used leastsquares regression to estimate template signature coefficients for the K_{0} and K_{H} conditions using the shoesonly template signature structure. We compared the AICc scores between these signature structures and those of the signature structures specific to the K_{0} and K_{H} trials. To determine if shoesonly template signatures were less plausible than signature structures selected for the K_{0} and K_{H} conditions, we used onesample righttailed ttests (α = 0.05) to test if differences in the average relative AICc scores were greater than three (e.g., \(AIC{c}_{Shoe}AIC{c}_{K0}> 3\) for the K_{0} condition)^{33}.
To determine if the ankle exoskeleton frame or mass impacted CoM dynamics, we compared template signature coefficients between the K_{0} and Shoe conditions. Similarly, to evaluate the impacts of exoskeleton stiffness on CoM dynamics, we compared template signature coefficients in the K_{H} and K_{0} conditions. For both comparisons, we used paired independentsamples ttests with HolmSidak stepdown corrections for multiple comparisons (α = 0.05)^{45}. Because shoesonly template signature structures were plausible for most unimpaired participants, we reestimated template signatures in the exoskeleton conditions using the shoesonly template signature structure before comparing coefficients across walking conditions.
To determine if CoM dynamics may be altered poststroke, we computed the percent difference in the nonparetic and paretic leg template signature coefficients during shoesonly walking in one individual with poststroke hemiparesis. We also evaluated changes in poststroke CoM dynamics with ankle exoskeletons by computing percent changes in template signature coefficients for the K_{0} condition compared to the shoesonly and K_{H} conditions.
Results
When walking in shoesonly, template signatures reveal common and more individualspecific representations of CoM dynamics across unimpaired participants. Unimpaired template signatures were not significantly different between legs (paired 2sample ttest; p > 0.080). In all gait phases, SLIP mechanisms—leg stiffness and leg length—were selected (i.e., had nonzero coefficients) in 100% of legs (\({\kappa }_{L}\) and \({L}_{0}\) in Fig. 4). Rotary stiffness and damping mechanisms were selected in less than 30% of legs in singlelimb support and swing. On average across participants and legs, 2.3 ± 0.8 and 2.9 ± 1.3 terms had nonzero coefficients in the stance and swing legs, respectively. More mechanisms were selected during doublelimb support phases: Rotary stiffness terms were selected in 79–83% of legs in the leading leg and 67–79% of legs in the trailing leg. Damping mechanisms were selected most frequently (33–67%) in the doublelimb support phases (\(\zeta\) terms in Fig. 4). On average, 5.0 ± 1.8 and 4.6 ± 1.6 terms had nonzero coefficients in first and second doublelimb support, respectively.
For each unimpaired participant, shoesonly template signature coefficients were reliable, having low bootstrapped coefficients of variation (CVs), during singlelimb support and swing: stiffness and leg length CVs were less than 0.02 and 0.06, respectively (Fig. 5A). In both doublelimb support phases, coefficient estimates were less reliable: CVs ranged from 0.06–4.34 across coefficients. During swing, coefficient estimates were generally reliable (CV = 0.03–0.10).
Across unimpaired participants, singlelimb support leg length was the least variable coefficient (0.97 ± 0.03), while leg stiffness values were more variable between participants (16.7 ± 5.8; Fig. 5A). Interindividual variability in doublelimb support leg and rotary stiffness was larger than singlelimb support coefficients. For example, leg stiffness (1.7 ± 6.0 in first doublelimb support) was lower and more variable than in singlelimb support. During swing, leg stiffness was relatively small (− 1.5 ± 0.8) compared to singlelimb support. Template signatures explained 83 ± 7% (range: 67–94%) of the variance in participants’ CoM (Example reconstruction of experimental data in Fig. 5B).
Passive ankle exoskeletons elicited only small changes in unimpaired template signatures: in singlelimb support, the shoesonly template signature structures reconstructed CoM dynamics with similar accuracy to template signature structures selected specifically for the K_{0} condition (mean difference in \(AICc=1.5\pm 8.4;\) p = 0.99) and the K_{H} condition (mean difference in \(AICc=6.8\pm 14.6;\) p = 0.10; Fig. 6A). Negative AICc scores indicate that the shoesonly template signature structure was more plausible than the exoskeletonspecific signature structure, which can occur if the shoesonly template signature structure is not identified in individual clusters for the exoskeleton conditions. In doublelimb support, shoesonly signatures were not statistically less plausible than K_{0} or K_{H} CoM template signatures (p > 0.657). However, the relative AICc scores were highly variable, ranging from − 258 \(\le\Delta AICc<\) 282, such that doublelimb support did not reliably indicate the plausibility of shoesonly template signature structures for the exoskeleton conditions. Because shoesonly template signatures were reliably plausible for gait with exoskeletons in singlelimb support, we constrained each participant’s template signature structures to their shoesonly signature structure. Therefore, we compared template signature coefficients between exoskeleton conditions using coefficients fit to each participant’s shoesonly signature structures.
Using this approach, we found that the only significant difference in unimpaired template signature coefficients was in the leg length coefficient during singlelimb support, which differed between walking in shoesonly and the zerostiffness exoskeletons (K_{0}; p = 7.9e − 4; α_{Sidak} = 9.2e − 4). However, this change was small (\(\Delta {\widetilde{L}}_{0}=0.01\pm 0.01)\), with leg length being slightly longer in the shoesonly condition. Neither ankle exoskeleton mass and frame (K_{0}) nor stiffness (K_{H}) altered other template signature coefficients in any gait phase (p > 0.16; α_{Sidak} = 9.3e − 4) (Fig. 6B).
One stroke survivor’s shoesonly template signature was symmetric in singlelimb support, swing, and first doublelimb support, but was asymmetric in second doublelimb support (Fig. 7). In singlelimb support and swing, leg stiffness and leg length mechanisms were selected and reliably estimated for both legs, with sagittalplane rotary stiffness also selected in swing (all CV < 0.02). In singlelimb support the paretic leg (\({\kappa }_{L}\) = 11.0; white bars in Fig. 7) was slightly (6%) stiffer than the nonparetic leg (\({\kappa }_{L}\) = 10.4; colored bars in Fig. 7). During first doublelimb support, template signature coefficients differed between legs, but only the paretic leg coefficients were reliably estimated (CV = 0.00–0.06). Conversely, in second doublelimb support, the sagittal and frontalplane rotary stiffness mechanisms were selected for the paretic, but not the nonparetic leg. Unlike first doublelimb support, nonparetic leg stiffness and resting length coefficients were more reliable (CV = 0.00–0.26) than those in the paretic leg (CV = 0.34–8.01).
The exoskeleton mass and frame (Shoe vs. K_{0}) primarily impacted the stroke survivor’s nonparetic and paretic leg stiffness in singlelimb support, and paretic leg rotary stiffness in second doublelimb support (Fig. 8). Note that these coefficients were reliably estimated by HybridSINDy. In singlelimb support, the zerostiffness (K_{0}) exoskeleton template signatures had 33% greater leg stiffness in the paretic leg and 19% lower leg stiffness in the nonparetic leg compared to shoesonly signatures (Fig. 8; dashed bars vs. solid color bars). In second doublelimb support, sagittal and frontal plane rotary stiffness were 37 and 50% greater, respectively, in the paretic limb in the K_{0} condition compared to shoesonly. Exoskeleton stiffness (K_{0} vs. K_{H}) had smaller impacts on paretic leg template signatures: in singlelimb support, paretic leg stiffness was 22% less than in the K_{0} condition (Fig. 8; solid black vs, dashed bars). Conversely, nonparetic leg stiffness was 85% lower in doublelimb support. Paretic leg sagittal and frontalplane rotary stiffness were both 33% lower in second doublelimb support than in the K_{0} condition.
Discussion
We evaluated the impacts of passive ankle exoskeletons on individualspecific templatebased representations of CoM dynamics—described by template signatures—using a recently developed datadriven modeling framework, HybridSINDy. Despite balancing model accuracy with parsimony, template signatures captured CoM dynamics with similar accuracy to prior work using predefined 2D template structures^{20,26}. The symmetric and SLIPlike unimpaired template signatures automatically selected by HybridSINDy during walking in shoesonly were consistent with prior template models of CoM dynamics during walking, but suggest that dynamics described by rotary mechanisms represent more individualspecific structures describing CoM accelerations^{19,25}. Contrary to our hypothesis, templatebased representations of unimpaired CoM dynamics were robust to the mechanical constraints of passive ankle exoskeletons: frame and mass (K_{0}), and dorsiflexion stiffness (K_{H}). Conversely, in our poststroke case study, asymmetric shoesonly signatures and changes in template signatures with exoskeletons support HybridSINDy’s potential to identify interpretable representations of pathological CoM dynamics, motivating future investigation into how neurological injuries impact templatebased representations of CoM dynamics with ankle exoskeletons^{28}.
Unimpaired and poststroke template signatures highlight potential interindividual differences in CoM dynamics. The selection of leg stiffness and resting length as active mechanisms (i.e., terms with nonzero coefficients) in 100% of legs and their selection as the only mechanisms in singlelimb support for most participants is consistent with common template walking models and supports the perspective that elastic legs are foundational mechanisms for describing CoM accelerations during unimpaired walking^{12,20,21,25,46}. Conversely, our finding that rotary mechanisms were selected to describe CoM dynamics in only 33–83% of legs is consistent with their less frequent application in template walking models^{13,19}. One interpretation of individual differences in selected mechanisms is that leg stiffness and resting length describe coordination patterns necessary for stable or efficient walking, while rotary mechanisms describe coordination patterns that have more individualspecific impacts on gait^{47}. Alternatively, differences in the selected mechanisms may be due to covariation among template signature state variables. We observed moderate covariation between some variables, particularly in doublelimb support (Supplemental S2: Covariation of template signature state variables). However, HybridSINDy penalizes the selection of strongly covarying states, such that both variables would not likely be selected unless they independently increased model likelihood^{31,41}.
Contrary to our primary hypothesis, unimpaired CoM dynamics are robust to altered ankle constraints due to passive ankle exoskeletons, despite observed changes in kinematics and muscle activity^{34}. Similarly, Collins and colleagues (2015) observed small changes in total CoM power with passive ankle exoskeletons compared to walking in shoesonly^{5}. Our findings suggest that, if changes in CoM dynamics in singlelimb support and swing are captured by the templatebased mechanisms in our function library, these small changes in CoM power are driven by changes in CoM kinematics, rather than changes in the underlying CoM dynamics. Note that these findings may not generalize to powered ankle exoskeletons, which elicit larger changes in gait kinematics and kinetics than did our passive exoskeletons and may yield larger changes in templatebased representations of CoM dynamics^{2,6,34}. Note that we limit our interpretation of template signature coefficients in doublelimb support, as they were not reliably estimated. More data may be needed to robustly estimate model coefficients in shorter gait phases.
However, changes—or a lack thereof—in template signature coefficients may be biased by an incomplete function library. Because template models are incomplete representations of gait, they may be sensitive to both measurement noise and unmodeled dynamics. Our analysis of a synthetic SLIP (Supplemental S2: Effects of measurement noise on algorithm performance) suggests that HybridSINDy can accurately identify template signature coefficients at measurement noise levels comparable to markerbased motion capture. However, changes in true CoM dynamics with exoskeletons are not completely represented by our mechanism library: template signatures accounted for less than 94% of the variance in human CoM accelerations. For example, our mechanism library did not include torso dynamics, which are known to contribute to angular momentum regulation during poststroke gait and may be altered with exoskeletons^{48,49}. Omitting functional forms from the mechanism library induces at least 5–10% differences in template signature coefficient estimates (Supplemental S2: Effects of missing physics on algorithm performance). While we encoded common functional forms from literature, larger function libraries or novel functional forms may increase the robustness of inter or intraindividual differences in template signatures to variations in kinematics between trials^{13,19,20,21,25,26,50}. However, even if a complete set of mechanisms is included in the library, HybridSINDy may not select small but important mechanisms needed to reconstruct CoM dynamics across tasks^{31,41}. Future studies should consider the tradeoff between the improved interpretability of moreparsimonious models with decreased model accuracy across the tasks of interest.
In our case study of one stroke survivor, asymmetric template signatures suggest that interleg differences in the mechanisms describing CoM dynamics can be automatically identified from data. Consistent with templatebased studies in children with cerebral palsy, the poststroke participant’s paretic leg was slightly (6%) stiffer than the nonparetic leg in singlelimb support, and rotary stiffness mechanisms were selected only in the nonparetic leg in doublelimb support^{27,28}. These mechanisms may reflect a more rigid paretic leg or reliance on proximal muscles for propulsion^{51,52}. However, as discussed above, the small difference in stiffness may be driven by unmodeled torso dynamics^{48}. While only a case study, these results highlight the potential interpretability of subjectspecific template signatures to understand how the legs contribute differentially to CoM accelerations following neurological injury.
Changes in our poststroke case study’s template signatures with exoskeletons support HybridSINDy’s ability to identify interpretable impacts of ankle exoskeletons on templatebased representations of CoM dynamics. For example, increases in paretic leg stiffness (K_{0} vs. shoesonly) may stem from the exoskeleton frame restricting inversion of the paretic ankle, which the participant noted during data collection. Conversely, reduced nonparetic leg stiffness and paretic leg rotary stiffness with stiff exoskeletons (K_{H} vs. K_{0}) may reflect a compensatory strategy to avoid reliance on the paretic leg, if it did not adapt effectively to the stiff exoskeleton^{2}. Note that these findings represent a proofofconcept that will not generalize across stroke survivors. Larger studies are needed to understand how individualspecific neural or biomechanical constraints may alter exoskeleton impacts on CoM dynamics^{2,5,7,28,34,53}.
The HybridSINDy algorithm^{31,32} was essential to discovering individualspecific and legspecific changes in CoM dynamics with ankle exoskeletons in the present study. Manually testing all possible template signatures for our fourteendimensional mechanism library would require a combinatorially large number of models to be fit and compared. Conversely, HybridSINDy automatically selected mechanistic representations of CoM dynamics that are consistent with the literature in a fraction of the time required to manually compare each candidate model^{13,19,20,21,25,26}. While HybridSINDy has largely been applied to synthetic systems, this work supports its ability to capture dynamics from human gait data^{31,32,54}.
While alternative modeling approaches, such as principle components analysis (PCA) or nonnegative matrix factorization (NMF), could identify lowdimensional representations of CoM dynamics, HybridSINDy provides immediately interpretable insight into the structure of these dynamics^{55,56}. HybridSINDy facilitates the interpretation of CoM dynamics by selecting a sparse set of readily interpretable template variables derived from expert knowledge of human gait^{12,13,19,20,21,25,26}. PCA or NMFbased template signatures would be harder to interpret, as they would not be sparse in the space of the template variables: each mode of PCA or NMF would likely contain nonzero coefficients for all state variables^{13,19,20,21,25,26}. Moresimilar approaches to ours used stepwise regression to model ankle quasistiffness using mechanicsbased functional forms^{57} or multilayer optimization to identify humanlike template dynamics for robot controllers^{58}. HybridSINDy is distinct from these approaches in its use of clustering and the AIC to evaluate the relative plausibility of competing representations of CoM dynamics, a limitation of stepwise regression^{41,59,60}. For all participants, HybridSINDy identified a single plausible shoesonly template signature structure in each gait phase, providing confidence that the identified signatures have strong statistical support compared to alternative signatures (see Supplemental S1: Comparison to alternative modeling frameworks).
Additional limitations should constrain the interpretation of template signatures. First, human gait dynamics are continuously phasevarying rather than hybrid, such that template signatures may vary continuously over the gait cycle. Our preliminary analyses found that predicted CoM dynamics became less accurate near the transitions between single and doublelimb support. Further, our prior work shows that phasevarying models of gait have higher predictive accuracy, but quantifying changes in the coefficients of continuously phasevarying models is challenging^{34,50}. Because our goal was to quantify changes in dynamics rather than predict CoM motion with maximal accuracy, defining gait phases based on contact configuration was reasonable and consistent with existing hybrid template models of walking^{19,20,25,31}. Second, we assumed that CoM dynamics were timeinvariant, though they may change with adaptation to exoskeletons^{61}. We included a twominute adaptation period for each walking condition, but additional adaptation may elicit larger changes in participants’ template signatures. Identifying the plausibility of template signatures across or within trials could improve our understanding of how templatebased representations of CoM dynamics change during adaptation^{26}. Third, our test set—90120 s of each trial recording—contained 25–30 strides of data per trial and was not a rigorous evaluation of model generalizability. Our approach identifies template signatures specific to a task and walking condition, and we do not expect these signatures to generalize across tasks or to overground walking. However, because template signature coefficients are known to vary with speed^{26} and kinematics (Supplemental S2: Effects of missing physics on algorithm performance), testing on withheld speeds or walking conditions was not practical. Finally, our limited sample size may have masked exoskeleton impacts on unimpaired gait.
Conclusions
We quantified changes in individualspecific template modelbased representations of CoM dynamics in response to passive ankle exoskeletons using an interpretable physicsinformed datadriven modeling framework, HybridSINDy. The template mechanisms describing salient features of unimpaired CoM dynamics were insensitive to ankle exoskeleton frame or stiffness. Interpretable ankle exoskeleton impacts on template representations of CoM dynamics in a case study of one individual poststroke support the utility of template signatures in quantifying CoM dynamics with exoskeletons in people with neurological injuries. These findings also support the potential of datadriven frameworks like HybridSINDy to accelerate the investigation of individualspecific representations of CoM dynamics during walking.
Data availability
All experimental data and modeling code used in this study are freely available at https://simtk.org/projects/ankleexopred.
References
Kerkum, Y. L. et al. The effects of varying ankle foot orthosis stiffness on gait in children with spastic cerebral palsy who walk with excessive knee flexion. PLoS ONE 10, e0142878 (2015).
McCain, E. M. et al. Mechanics and energetics of poststroke walking aided by a powered ankle exoskeleton with speedadaptive myoelectric control. J. Neuroeng. Rehabil. 16, 1–12 (2019).
Ries, A. J., Novacheck, T. F. & Schwartz, M. H. The efficacy of anklefoot orthoses on improving the gait of children with diplegic cerebral palsy: A multiple outcome analysis. PM&R 7, 922–929 (2015).
Lerner, Z. F., Harvey, T. A. & Lawson, J. L. A batterypowered ankle exoskeleton improves gait mechanics in a feasibility study of individuals with cerebral palsy. Ann. Biomed. Eng. 47, 1345–1356 (2019).
Collins, S. H., Wiggin, M. B. & Sawicki, G. S. Reducing the energy cost of human walking using an unpowered exoskeleton. Nature 522, 212–215 (2015).
Jackson, R. W., Dembia, C. L., Delp, S. L. & Collins, S. H. Muscle–tendon mechanics explain unexpected effects of exoskeleton assistance on metabolic rate during walking. J. Exp. Biol. 220, 2082–2095 (2017).
Sawicki, G. S. & Khan, N. S. A simple model to estimate plantarflexor muscletendon mechanics and energetics during walking with elastic ankle exoskeletons. IEEE Trans. Biomed. Eng. 63, 914–923 (2016).
Donelan, J. M., Kram, R. & Kuo, A. D. Mechanical work for steptostep transitions is a major determinant of the metabolic cost of human walking. J. Exp. Biol. 205, 3717–3727 (2002).
Kuo, A. D. The six determinants of gait and the inverted pendulum analogy: A dynamic walking perspective. Hum. Mov. Sci. 26, 617–656 (2007).
Kuo, A. D. Stabilization of lateral motion in passive dynamic walking. Int. J. Robot. Res. 18, 917–930. https://doi.org/10.1177/02783649922066655 (1999).
Farris, D. J., Hampton, A., Lewek, M. D. & Sawicki, G. S. Revisiting the mechanics and energetics of walking in individuals with chronic hemiparesis following stroke: From individual limbs to lower limb joints. J. Neuroeng. Rehabil. 12, 1–12 (2015).
Alexander, R. M. Simple models of human movement. (1995).
Davoodi, A., Mohseni, O., Seyfarth, A. & Sharbafi, M. A. From template to anchors: Transfer of virtual pendulum posture control balance template to adaptive neuromuscular gait model increases walking stability. R. Soc. Open Sci. 6, 181911 (2019).
McGeer, T. Passive dynamic walking. I. J. Robotic Res. 9, 62–82 (1990).
Full, R. J. & Koditschek, D. E. Templates and anchors: Neuromechanical hypotheses of legged locomotion on land. J. Exp. Biol. 202, 3325–3332 (1999).
Cavagna, G. A., Heglund, N. C. & Taylor, C. R. Mechanical work in terrestrial locomotion: Two basic mechanisms for minimizing energy expenditure. Am. J. Physiol.Regul. Integr. Compar. Physiol. 233, R243–R261 (1977).
Holmes, P., Full, R. J., Koditschek, D. & Guckenheimer, J. The dynamics of legged locomotion: Models, analyses, and challenges. SIAM Rev. 48, 207–304 (2006).
Srinivasan, M. & Ruina, A. Computer optimization of a minimal biped model discovers walking and running. Nature 439, 72 (2006).
Antoniak, G. et al. Springloaded inverted pendulum goes through two contractionextension cycles during the singlesupport phase of walking. Biol. Open https://doi.org/10.1242/bio.043695 (2019).
Lipfert, S. W., Günther, M., Renjewski, D., Grimmer, S. & Seyfarth, A. A modelexperiment comparison of system dynamics for human walking and running. J. Theor. Biol. 292, 11–17 (2012).
Maus, H.M., Lipfert, S., Gross, M., Rummel, J. & Seyfarth, A. Upright human gait did not provide a major mechanical challenge for our ancestors. Nat. Commun. 1, 1–6 (2010).
Garcia, M., Chatterjee, A., Ruina, A. & Coleman, M. The simplest walking model: stability, complexity, and scaling. (1998).
Buczek, F. L. et al. Performance of an inverted pendulum model directly applied to normal human gait. Clin. Biomech. 21, 288–296 (2006).
Donelan, J. M., Shipman, D. W., Kram, R. & Kuo, A. D. Mechanical and metabolic requirements for active lateral stabilization in human walking. J. Biomech. 37, 827–835 (2004).
Geyer, H., Seyfarth, A. & Blickhan, R. Compliant leg behaviour explains basic dynamics of walking and running. Proc. R. Soc. B Biol. Sci. 273, 2861–2867 (2006).
Kim, S. & Park, S. Leg stiffness increases with speed to modulate gait frequency and propulsion energy. J. Biomech. 44, 1253–1258. https://doi.org/10.1016/j.jbiomech.2011.02.072 (2011).
Fonseca, S. R. T., Holt, K. G., Fetters, L. & Saltzman, E. Dynamic resources used in ambulation by children with spastic hemiplegic cerebral palsy: Relationship to kinematics, energetics, and asymmetries. Phys. Ther. 84, 344–354 (2004).
Holt, K. G., Fonseca, S. T. & LaFiandra, M. E. The dynamics of gait in children with spastic hemiplegic cerebral palsy: Theoretical and clinical implications. Hum. Mov. Sci. 19, 375–405 (2000).
Brunton, S. L., Proctor, J. L. & Kutz, J. N. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proc. Natl. Acad. Sci. 113, 3932–3937 (2016).
Brunton, S. L. & Kutz, J. N. DataDriven Science and Engineering: Machine Learning, Dynamical Systems, and Control (Cambridge University Press, 2019).
Mangan, N. M., Askham, T., Brunton, S. L., Kutz, J. N. & Proctor, J. L. Model selection for hybrid dynamical systems via sparse regression. Proc. R. Soc. A 475, 20180534 (2019).
Mangan, N. M., Kutz, J. N., Brunton, S. L. & Proctor, J. L. Model selection for dynamical systems via sparse regression and information criteria. Proc. R. Soc. A 473, 20170009 (2017).
Anderson, D. & Burnham, K. Model Selection and MultiModel Inference Vol. 63 (SpringerVerlag, 2004).
Rosenberg, M. C., Banjanin, B. S., Burden, S. A. & Steele, K. M. Predicting walking response to ankle exoskeletons using datadriven models. J. R. Soc. Interface 17, 20200487 (2020).
Genthe, K. et al. Effects of realtime gait biofeedback on paretic propulsion and gait biomechanics in individuals poststroke. Top. Stroke Rehabil. 25, 186–193 (2018).
Delp, S. L. et al. OpenSim: Opensource software to create and analyze dynamic simulations of movement. IEEE Trans. Biomed. Eng. 54, 1940–1950 (2007).
Rajagopal, A. et al. Fullbody musculoskeletal model for muscledriven simulation of human gait. IEEE Trans. Biomed. Eng. 63, 2068–2079 (2016).
Abdullah, F., Wu, Z. & Christofides, P. D. Handling noisy data in sparse model identification using subsampling and coteaching. Comput. Chem. Eng. 157, 107628 (2022).
Villarreal, D. J., Poonawala, H. A. & Gregg, R. D. A robust parameterization of human gait patterns across phaseshifting perturbations. IEEE Trans. Neural Syst. Rehabil. Eng. 25, 265–278 (2016).
Whittle, M. W. Clinical gait analysis: A review. Human movement science 15, 369–387 (1996).
Akaike, H. Selected Papers of Hirotugu Akaike 215–222 (Springer, 1974).
Dingwell, J. B. & Cusumano, J. P. Identifying stridetostride control strategies in human treadmill walking. PloS One 10, e0124879 (2015).
Dingwell, J. B. & Cusumano, J. P. Humans use multiobjective control to regulate lateral foot placement when walking. PLoS Computat. Biol. 15, e1006850 (2019).
Maus, H.M. et al. Constructing predictive models of human running. J. R. Soc. Interface 12, 20140899 (2015).
Glantz, S. (New York: McGrawHill, 2012).
Mochon, S. & McMahon, T. A. Ballistic walking. J. Biomech. 13, 49–57 (1980).
Lee, C. R. & Farley, C. T. Determinants of the center of mass trajectory in human walking and running. J. Exp. Bio. 201, 2935–2944 (1998).
Nott, C., Neptune, R. & Kautz, S. Relationships between frontalplane angular momentum and clinical balance measures during poststroke hemiparetic walking. Gait Posture 39, 129–134 (2014).
Vistamehr, A., Kautz, S. A. & Neptune, R. R. The influence of solid anklefootorthoses on forward propulsion and dynamic balance in healthy adults during walking. Clin. Biomech. 29, 583–589 (2014).
Ebers, M. R., Rosenberg, M. C., Kutz, J. N. & Steele, K. M. A machine learning approach to quantify individual gait responses to ankle exoskeletons. J. Biomech. 157, 111695 (2023).
Clark, D. J., Ting, L. H., Zajac, F. E., Neptune, R. R. & Kautz, S. A. Merging of Healthy Motor Modules Predicts Reduced Locomotor Performance and Muscle Coordination Complexity PostStroke. J. Neurophysiol. 103, 844–857. https://doi.org/10.1152/jn.00825.2009 (2010).
Allen, J. L., Kautz, S. A. & Neptune, R. R. The influence of merged muscle excitation modules on poststroke hemiparetic walking performance. Clin. Biomech. 28, 697–704 (2013).
Steele, K. M., Jackson, R. W., Shuman, B. R. & Collins, S. H. Muscle recruitment and coordination with an ankle exoskeleton. Journal of biomechanics 59, 50–58 (2017).
Novelli, N., Lenci, S. & Belardinelli, P. Boosting the model discovery of hybrid dynamical systems in an informed sparse regression approach. J. Computat. Nonlinear Dyn. https://doi.org/10.1115/1.4053324 (2022).
Lee, D. D. & Seung, H. S. Learning the parts of objects by nonnegative matrix factorization. Nature 401, 788 (1999).
Ivanenko, Y. P., Cappellini, G., Dominici, N., Poppele, R. E. & Lacquaniti, F. Coordination of locomotion with voluntary movements in humans. J. Neurosci. 25, 7238–7253 (2005).
Shamaei, K., Sawicki, G. S. & Dollar, A. M. Estimation of quasistiffness and propulsive work of the human ankle in the stance phase of walking. PloS One 8, e59935 (2013).
Chen, Y.M. & Posa, M. in 2020 IEEE International Conference on Robotics and Automation (ICRA). (IEEE). 8753–8760.
Whittingham, M. J., Stephens, P. A., Bradbury, R. B. & Freckleton, R. P. Why do we still use stepwise modelling in ecology and behaviour?. J. Anim. Ecol. 75, 1182–1189 (2006).
Anderson, D. R. & Burnham, K. P. Avoiding pitfalls when using informationtheoretic methods. J. Wildlife Manag. 66, 912–918 (2002).
Seethapathi, N., Clark, B. & Srinivasan, M. Explorationbased learning of a step to step controller predicts locomotor adaptation. bioRxiv https://doi.org/10.1101/2021.03.18.435986 (2021).
Funding
This material is based upon work supported by the National Science Foundation (https://nsf.gov/) under grant no. CBET1452646 to Katherine. M. Steele, the National Science Foundation Graduate Research Fellowship Program (https://www.nsfgrfp.org/) under grant no. DGE1762114 to Michael. C. Rosenberg, and a University of Washington Gatzert Child Welfare Fellowship (https://grad.uw.edu/graduatestudentfunding/fundinginformationforstudents/fellowships/listoffellowships/gatzertchildwelfarefellowship/) to Michael. C. Rosenberg. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Author information
Authors and Affiliations
Contributions
M.C.R., K.M.S. and J.L.P. contributed equally to the conception and design of this work and to the interpretation of results. M.C.R. collected data, conducted all analyses, created figures, and wrote the initial manuscript draft. J.L.P. provided theoretical guidance on the implementation of the HybridSINDy algorithm. K.M.S. provided experimental resources and guidance on the analysis of gait with exoskeletons. All authors contributed to substantive revisions of the manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Rosenberg, M.C., Proctor, J.L. & Steele, K.M. Quantifying changes in individualspecific templatebased representations of centerofmass dynamics during walking with ankle exoskeletons using HybridSINDy. Sci Rep 14, 1031 (2024). https://doi.org/10.1038/s41598023509990
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41598023509990
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.