Synergies are minimally affected during emulation of cerebral palsy gait patterns

Abstract

Muscle synergy analysis is commonly used to characterize motor control during dynamic tasks like walking. For clinical populations, such as children with cerebral palsy (CP), synergies are altered compared to nondisabled (ND) peers and have been associated with both function and treatment outcomes. However, the factors that contribute to altered synergies remain unclear. In particular, the extent to which synergies reflect altered biomechanics (e.g., changes in gait) or underlying neurologic injury is debated. To evaluate the effect that altered biomechanics have on synergies, we compared synergy complexity and structure while ND individuals (n = 14) emulated four common CP gait patterns (equinus, equinus-crouch, mild-crouch, and moderate crouch). Secondarily, we compared the similarity of ND synergies during emulation to synergies from a retrospective cohort of individuals with CP walking in similar gait patterns (n = 28 per pattern). During emulation, ND individuals recruited similar synergies as baseline walking. However, pattern-specific deviations in synergy activations and complexity emerged. In particular, equinus gait altered plantarflexor activation timing and reduced synergy complexity. Importantly, ND synergies during emulation were distinct from those observed in CP for all gait patterns. These results suggest that altered gait patterns are not primarily driving the changes in synergies observed in CP, highlighting the value of using synergies as a tool to capture patient-specific differences in motor control. However, they also highlight the sensitivity of both synergy activations and complexity to altered biomechanics, which should be considered when using these measures in clinical care.

1. INTRODUCTION

For individuals with cerebral palsy (CP), neurologic injury near birth alters motor control, often making walking challenging and inefficient (Gage et al., 2009; Waters and Mulroy, 1999). As such, many individuals will undergo surgical interventions to improve mobility. Yet treatment outcomes are variable and often unsatisfactory (Hicks et al. 2011; Steele et al 2013; Ubhi et al, 2000).

To improve the consistency of outcomes, recent research has investigated new methods, such as synergy analysis, to directly characterize impaired motor control in CP (Bekius et al., 2020). Synergy analysis identifies weighted groups of co-activating muscles (i.e., synergies) from electromyography (EMG) data collected during movement (Lee, 1999; Ting and McKay, 2007). Prior work has demonstrated that nondisabled (ND) adults recruit a small number of synergies during walking (Ivanenko et al., 2004), running (Cappellini et al. 2006), and balance tasks (Chvatal et al., 2011; Torres-Oviedo and Ting, 2010). In contrast, individuals with CP use fewer synergies during gait, similar to stroke survivors (Clark et al., 2010) or individuals with spinal cord injury (Fox et al. 2013) or Parkinson’s disease (Rodriguez et al., 2013). Further, more impaired synergies are associated with worse function and treatment outcomes (Schwartz et al., 2016; Shuman et al., 2019; Tang et al., 2015). While these results highlight the efficacy of using synergies as a measure of motor control, it is unclear what factors contribute to altered synergies in CP,

Given the complex presentation of CP, many factors could impact synergies. Because CP affects the motor cortex, altered synergies may reflect atypical neurophysiology. This stems from work that suggests that stepping in early development is driven by spinal circuits (e.g., central pattern generators) but becomes increasingly shaped by supraspinal pathways during central nervous system maturation (Leonard et al., 1991). Under this hypothesis, aberrant supraspinal input due to neurologic injury may increase reliance on spinal circuits, resulting in simplified control (Tresch and Jarc, 2009). This is supported by work which has demonstrated that individuals with CP and spinal cord injury, as well as stroke survivors, use similar synergies to those observed during rhythmic-stepping in infants (Cappellini et al., 2016; Clark et al., 2010; Dominici et al., 2011; Fox et al., 2013). However, biomechanical constraints may also affect synergies. Individuals with CP adopt altered gait patterns, often characterized by toe-walking (equinus) and excessive knee and hip flexion (crouch; Wren et al., 2005). These patterns may restrict joint ranges of motion, effectively reducing the feasible activation space for muscles and contributing to altered synergies (Kutch and Valero-Cuevas, 2012; Steele et al., 2015b).

Analyzing how biomechanical and neurological constraints impact synergies in CP will inform the clinical use of synergy analysis. However, because these constraints are intricately coupled following neurologic injury, evaluating their contributions in CP directly is challenging. Alternatively, ND emulation of CP gait patterns presents a useful experimental paradigm to understand the effect of biomechanical constraints on synergies. Prior studies have used emulation to probe how CP gait impacts kinetics (Davids et al., 1999; Thomas et al., 1996), metabolics (Thomas et al., 1996), muscle activity (Davids et al., 1999; Romkes and Brunner, 2007; Thomas et al., 1996), and muscle length (van der Krogt et al., 2007). While many of these studies noted altered muscle activations during emulation, whether synergies are affected remains unknown.

The purpose of this study was to evaluate whether synergies during ND emulation of CP gait (1) differed from those used during baseline walking and (2) aligned with synergies recruited by individuals with CP walking in similar gait patterns. We hypothesized that ND individuals would use similar synergies during baseline and emulated gait and that these synergies would be more complex than synergies used by individuals with CP. If supported, this hypothesis would suggest that altered synergies in CP reflect neurological rather than biomechanical constraints. This investigation will help identify the factors captured by synergy-based measures to inform the efficacy of using synergy analysis to guide treatment in CP.

2. METHODS

2.1. Experimental Protocol

We recruited fourteen ND adults (7M/7F; median [IQR]: 23 years [21,25]) to analyze synergies during CP gait pattern emulation. All participants provided written consent and procedures were approved by the Institutional Review Board at the University of Washington.

Participants walked on a treadmill at self-selected speed (1.05 m/s [1.0,1.1]; Bertec Corp., Columbus, USA) while performing one three-minute baseline trial followed by three, three-minute trials in the following patterns: equinus (Eq), equinus-crouch (Eq-Cr), mild crouch (Mi-Cr), and moderate crouch (Mod-Cr). These patterns were selected as they represent approximately 69% (crouch) and 61% (equinus) of ambulatory children with CP (Wren et al., 2005).

Verbal instructions and visual cues were used to guide emulation. To emulate equinus, participants were instructed to maintain ankle plantarflexion through stance while minimizing knee flexion. For mild and moderate crouch, 20° and 30° of knee flexion were measured, respectively, using a goniometer during stance and reinforced in-trial with a hanging marker (Rozumalski and Schwartz, 2009). Participants were instructed to keep the marker at eye-level by maintaining knee flexion and minimizing trunk flexion (van der Krogt et al., 2007). To emulate equinus-crouch, participants adopted the same equinus posture along with 20° of knee flexion, set and reinforced using the same protocol as crouch.

Full-body motion data were collected using a modified Helen-Hayes marker set and a 10-camera motion capture system at 120 Hz (Qualisys AB, Gothenburg, SE ). Joint angles were computed from marker trajectories in OpenSim v3.3 (Stanford, USA) using a 33 degree-of-freedom model scaled to each subject (Delp et al., 2007; Rajagopal et al., 2016). Across trials, the root-mean-square (RMS) and maximum model error for all markers were 1.4 cm and 2.5 cm, respectively, which align with best practices for model quality (Hicks et al., 2015).

Surface EMG data (Delsys Inc, Natick, MA) were collected bilaterally for the gluteus maximus (GMAX), lateral hamstrings (LH), medial hamstrings (MH), vastus medialis (VM), soleus (SOL), tibialis anterior (TA), and medial gastrocnemius (GAS). Raw EMG signals were high pass filtered (4th order Butterworth; 20 Hz), rectified, low pass-filtered (4th order Butterworth; 10 Hz), and normalized to the 95^th percentile of maximum activation across trials (De Luca et al., 2010; Shuman et al., 2017). Any data spikes due to sensor movement were removed using a robust-PCA algorithm (Candès et al., 2011; Lin et al., 2013).

2.2. Synergy Analysis

Synergies were calculated using non-negative matrix factorization (NMF; Lee, 1999). NMF models EMG data as a linear combination of non-negative, time-invariant synergies (W) and their activation patterns (C) such that:

$$EMG=W_{mxn}*C_{nxt}+error$$ (1)

where m is the number of muscles, n is the number of synergies, and t is the number of time points. For each gait pattern, n = 1 to 7 synergies were calculated using dominant-leg EMG data. Because synergy analysis is sensitive to the amount of EMG data used, we analyzed the same number of strides (n = 43) across gait patterns and participants (Oliveira et al., 2014). For the emulated gait patterns, these data were taken from the middle of the final trial to minimize learning effects due to pattern novelty. The structure and complexity of the resultant synergies were then compared across gait patterns.

2.2.1. Synergy Complexity

We quantified synergy complexity as the total variance accounted for (tVAF) in the EMG data by a set of synergies (n), defined as:

$$tVA{F}_n=\left(1-\frac{{\sum }_j^t{{\sum }_i^m\left({(W_n*C_n)}_{i,j}-EM{G}_{i,j}\right)}^2}{{\sum }_j^t{{\sum }_i^m\left(EM{G}_{i,j}\right)}^2}\right)X100\%$$ (2)

Here, a lower tVAF indicates that a set of synergies captures less variance in the EMG data, suggesting more complex motor control. Importantly, tVAF is associated with impairment level, as children with CP have higher tVAF than ND children (Schwartz et al., 2016; Steele et al., 2015a; Tang et al., 2015). If altered gait patterns influence tVAF in CP, we would expect higher tVAF during ND emulation than baseline gait.

2.2.2. Synergy Structure

To evaluate whether synergy structure changed during emulation, we first identified synergy weights (W) and activations (C) that were shared across patterns using k-means cluster analysis (MacQueen, 1967). For each synergy solution, n, weights from all participants and gait patterns were sorted into k clusters in Matlab (MathWorks, Natick, USA). To allow for the possibility that pattern-specific synergies may emerge, clustering was performed with k ranging from k = n (i.e., baseline and emulated gait patterns shared a set of synergies) to k = 5*n (i.e., all gait patterns had unique synergies). We selected k as the number of clusters with the maximum silhouette coefficient (Rousseeuw, 1987). Median weights and activations for each gait pattern were then calculated from the resultant clusters.

We quantified the similarity of the synergy structures recruited during baseline and emulated gait by calculating how much of the variance in the emulated gait EMG data could be explained by baseline synergy weights (W_baseline). We used the multiplicative update rule from NMF to iteratively solve for C for each emulated pattern (C_emulation) while keeping W_baseline constant. For each synergy solution, n, the reconstructed total variance accounted for (tVAF_{n_recon}) was calculated using W_baseline and C_emulation (Eq. 2) and compared to the original tVAF; if synergy structure was similar during baseline and emulated gait, we would expect both tVAF values to also be similar.

2.3. Comparison to CP

Synergy structure and complexity during emulated gait were also compared directly to CP synergies, using retrospective, overground walking data collected at Gillette Children’s Specialty Healthcare. Two individuals with CP were matched to every ND individual for each emulated gait pattern. Individuals with CP were included in the matching pool if they had diplegia and walked without an assistive device, regardless of intervention history or injury severity. Matches were identified as those individuals with the smallest total RMS difference in sagittal-plane hip, knee, and ankle kinematics,. To increase diversity in our CP cohort, matches were selected without replacement, resulting in 28 unique individuals for each pattern (median age [IQR]; Eq: 11.1 [7.5,15.5], Eq-Cr: 10.6 [7.5,13.3], Mi-Cr: 12.5 [9.4,15.9], Mod-Cr: 12.1 [8.6,13.7]). Across all matches, the median cosine similarity was 0.94 [0.90,0.97], 0.98 [0.97,0.99], and 0.89 [0.77,0.95] and the median RMS difference was 13.7° [9.6,20.1], 9.0° [6.6,12.4], and 7.1° [5.2,10.0] at the hip, knee, and ankle, respectively. Dimensionless walking speed was also similar between cohorts (Mann-Whitney U test; p > 0.42) (Hof, 1996).

For the CP cohort, we processed kinematic and EMG data (vastus lateralis (VL), MH, LH, TA, GAS) using the procedures outlined above and performed synergy analysis for n = 1-4 synergies using three concatenated strides. For individuals with more than three recorded strides, we performed a bootstrapping procedure by calculating synergies from three random strides in the available set and replicating the process until a normal distribution was achieved. Average synergy weights and activations from the resultant distribution were then used to represent the individual’s recorded data. We also recalculated ND synergies for the reduced muscle set and three concatenated strides using the same bootstrapping procedure; however, VM was used in place of VL, as the latter was not collected experimentally. In both cohorts, k-means clustering was used to calculate median weights and activations for each gait pattern.

2.4. Statistical Analysis

To evaluate if synergy complexity and structure changed between baseline and emulated gait patterns, Wilcoxon signed-rank tests were used with a Holm-Šídák correction for multiple comparisons (n = 4); non-parametric statistics were selected due to the small sample size (Ghasemi and Zahediasl, 2012). Significant deviations in kinematics and integrated EMG between emulated and baseline walking were also identified using Wilcoxon signed-rank tests with correction for multiple comparisons. Further, Mann-Whitney U tests were used to compare synergy complexity between ND and CP cohorts for all synergy solutions (n). For all tests, we define significance as p < α for α = 0.05 and report medians [IQR] unless otherwise noted. All statistical analyses were performed using the Matlab Statistical Toolbox.

3. RESULTS

All participants emulated the major kinematic trends seen in equinus, equinus-crouch, mild crouch, and moderate crouch gait (Figure 1). In equinus, participants increased plantarflexion during weight acceptance (−6.85° [−9.8°, −1.4°]) and terminal swing (−2.2° [−9.4°, −0.1°]) and decreased dorsiflexion through stance compared to baseline (p < 0.001; Table 1). This corresponded to increased co-contraction of antagonists at the ankle at midstance (p < 0.02). Mild and moderate crouch were characterized by increased hip flexion, knee flexion, and ankle dorsiflexion in stance, which corresponded to increased quadricep activity and reduced gastrocnemius activity (p < 0.002). In equinus-crouch, knee and hip flexion increased through stance and ankle dorsiflexion decreased at weight acceptance (0.02° [−3.0°, 3.5°]; p < 0.01). These changes were associated with increased quadricep and soleus activity through stance and increased gastrocnemius activity at weight acceptance (p < 0.001).

Figure 1:

Median sagittal-plane kinematics (Top) and EMG (Bottom) during baseline, equinus, equinus-crouch, mild crouch, and moderate crouch gait by ND adults. Shading indicates the IQR for baseline gait. Muscles: gluteus maximus (GMAX), lateral hamstrings (LH), medial hamstrings (MH), vastus medialis (VM), medial gastrocnemius (GAS), soleus (SOL), and tibialis anterior (TA).

Table 1:

Sagittal plane kinematics of ND adults during baseline and emulated gait patterns.

	Baseline	Equinus		Equinus Crouch		Mild Crouch		Moderate Crouch
Gait Variable (deg)	Median [IQR]	Median [IQR]	p	Median [IQR]	p	Median [IQR]	p	Median [IQR]	p
Hip
Max extension	14.7 [11.6,17.9]	15.1 [8.3,18.4]	0.63	10.5 [8.0,12.9]	0.051	9.3 [6.8,10.4]	0.002*	5.7 [1.4,10.4]	0.002*
Max flexion	25.7 [20.9,30.0]	26.4 [24.1,32.9]	0.30	34.9 [32.4,41.6]	<0.001*	33.3 [32.0,38.9]	<0.001*	37.9 [33.2,40.9]	<0.001*
Knee
Mean flexion in stance	19.4 [16.4,20.8]	21.3 [20.0,24.0]	0.02*	37.6 [35.7,45.8]	<0.001*	37.8 [35.0,40.7]	<0.001*	42.2 [39.0,47.7]	<0.001*
Ankle
Max dorsiflexion	21.4 [19.7,25.5]	10.5 [8.6,15.5]	<0.001*	22.2 [17.6,26.8]	0.98	28.2 [25.4,29.7]	<0.001*	30.5 [29.1,30.9]	<0.001*
Mean dorsiflexion in stance	10.4 [9.7,12.9]	4.3 [2.2,8.2]	<0.001*	11.6 [10.7,18.5]	0.051	17.0 [14.8,19.8]	<0.001*	19.0 [17.7,23.2]	<0.001*

^*Significant difference (α = 0.05) from baseline using Wilcoxon signed-rank test with Holm-Šídák correction

3.1. Synergy Complexity

The tVAF significantly increased during equinus gait for all synergy solutions except n = 2 (p = 0.07; Figure 2). However, in all other gait patterns tVAF was similar to baseline for all synergy solutions (p > 0.09). At baseline, one synergy accounted for 68.2% [63.0,73.4] of the variance in the EMG data compared to 76.1% [70.9,81.4], 73.8% [67.3,77.4], 66.0% [63.3,71.7], and 70.0% [65.5,74.8] for Eq, Eq-Cr,Mi-Cr, and Mod-Cr walking, respectively.

Figure 2:

The total variance accounted for (tVAF) by n= 1 to 5 synergies for baseline gait and each emulated gait pattern. An increase in tVAF corresponds to a decrease in control complexity. * denotes significant difference from baseline gait (α = 0.05) using Wilcoxon signed-rank tests with a Holm-Šídák correction for multiple comparisons and dots indicate outliers. Gait patterns: baseline, equinus (Eq), equinus-crouch (Eq-Cr), mild crouch (Mi-Cr), and moderate crouch (Mod-Cr).

3.2. Synergy Structure

Synergy weights were also minimally affected during emulation. The three-synergy solution was compared across patterns, which accounted for over 90% of the variance in muscle activity for most participants (Figure 3). Although k could vary, three clusters emerged across all patterns that were dominated by the gluteus maximus and quadriceps (W₁), the hamstrings (W₂), and the plantarflexors (W₃), similar to prior findings in ND adults (Allen and Neptune, 2012; Clark et al., 2010).

Figure 3:

Synergy weights (W) and activations (C) for the three-synergy solution for baseline and emulated gait patterns. Median synergy weights across participants as well as synergy weights for each participant, sorted in descending value order, are displayed. Muscles: gluteus maximus (GMAX), lateral hamstrings (LH), medial hamstrings (MH), vastus medialis (VM), medial gastrocnemius (GAS), soleus (SOL), and tibialis anterior (TA). Gait patterns: baseline, equinus (Eq), equinus-crouch (Eq-Cr), mild crouch (Mi-Cr), and moderate crouch (Mod-Cr).

Although emulated and baseline synergy weights were similar, differences in activations emerged. Equinus resulted in increased activation of W₃ through stance and terminal swing (C₃); the latter trend has been reported in equinus as evidence of initial contact planning (Romkes and Brunner, 2007). Additionally, mild and moderate crouch had increased activation of W1 through stance (C₁), due to a larger relative contribution of the quadriceps, and increased coactivation of the tibialis anterior and hamstrings through swing (C₂). In equinus-crouch walking, both W₁ and W₃ had increased activation in early stance following trends in equinus and crouch patterns, respectively. These differences corresponded to a significant difference between tVAF₃ and tVAF_{3_recon} (i.e., tVAF by baseline weights) for all emulated patterns. Baseline weights accounted for 6.5% [5.2,11.7], 10.2% [7.1,14.5], 12.3% [10.4,18.4], and 19.1% [12.0,22.7] less of the variance in the EMG data for Eq, Eq-Cr, Mi-Cr, and Mod-Cr gait, respectively, than the pattern-specific weights (p < 0.001 for all patterns).

3.3. Comparison to CP

Comparing emulated gait synergies to the CP cohort, notable differences in both structure and complexity emerged (Figure 4). Despite having similar kinematics, tVAF₁ was significantly larger in the CP cohort than the TD cohort for all gait patterns except equinus (Table 2). This indicates that the CP cohort generally relied on less complex motor control during gait than ND individuals emulating similar patterns. This difference in control is further supported when looking at synergy structure. When k-means clustering was performed on the two-synergy solution, four unique synergies emerged for the ND cohort compared to two synergies for the CP cohort (Figure 4). Synergies, dominated by the tibialis anterior (W₁) and the hamstrings (W₂), emerged in both cohorts. However, in W₂, the CP cohort also had increased quadricep and gastrocnemius activity that was not observed in the ND participants. The ND cohort also had unique synergies dominated by (1) the tibialis anterior and gastrocnemius (W_ND1), which emerged for the majority of individuals during equinus gait and (2) the quadriceps (W_ND2), which emerged for the majority of individuals during equinus-crouch, mild crouch, and moderate crouch.

Figure 4:

Synergy weights (W), activations (C), and complexity (tVAF) for the ND and CP cohorts during equinus (Eq), equinus-crouch (Eq-Cr), mild crouch (Mi-Cr), and moderate crouch (Mod-Cr) gait. Synergy weights for both CP and ND groups represent the two-synergy solution. K-means clustering was performed on each group and resulted in two clusters for the CP cohort (W₁ and W₂) and four clusters for the ND cohort (W₁, W₂, W_ND1 and W_ND2). Percentages reflect the number of ND individuals that recruited each synergy. * denotes significant difference between ND and CP for each n using Mann-Whitney U tests (α = 0.05) and dots indicate outliers. Muscles: medial hamstring (MH), vastus lateralis (VL), medial gastrocnemius (GAS), and tibialis anterior (TA). Note that in the ND cohort, the vastus medialis was used in place of the VL, as the VL was not collected experimentally.

Table 2:

Motor control complexity (tVAF₁) during ND emulated gait patterns compared to a CP cohort.

	tVAF1 (%)
Gait Pattern	CP Cohort (N = 28/pattern)	TD Cohort (N = 14)	p
Equinus	77.0 [73.2,81.3]	79.1 [76.3,82.4]	0.4
Equinus Crouch	79.5 [76.2,83.4]	73.4 [68.8,76.0]	0.001*
Mild Crouch	78.0 [73.0,82.0]	70.8 [67.4,72.1]	<0.001*
Moderate Crouch	80.0 [76.4,84.3]	73.7 [69.3,76.7]	0.002*

Note: Values are listed as median [IQR]. tVAF₁ (%) is the total variance in EMG data accounted for by a 1-synergy solution

^*Significant difference (α = 0.05) between CP and ND cohorts using Mann-Whitney U tests

4. DISCUSSION

Although ND individuals were able to emulate CP gait, synergies did not deviate largely from baseline measures and, importantly, were distinct from synergies recruited by individuals with CP walking in similar patterns. However, emulation did alter the synergy activations and inconsistently affected synergy complexity (i.e., tVAF). These results collectively suggest that synergies in CP are not solely a reflection of altered gait patterns but rather capture aberrant neurophysiology.

4.1. Synergies are invariant to imposed biomechanical constraints

Three synergies emerged across all gait patterns, indicating that synergy weights were largely invariant to imposed biomechanical constraints. It is important to note that while >74% of the variance was captured when using baseline weights to reconstruct muscle activity during emulated gait patterns, there was a significant difference between tVAF₃ and tVAF_{3_recon} for all gait patterns. However, this difference was largely due to changes in the relative magnitude of muscle activations; importantly, similar muscles were recruited in each synergy across gait patterns.

The observed similarity of synergy weights during baseline and emulated gait patterns aligns with previous work. In ND adults, prior work has found that similar synergies are recruited with varying gait speed (Ivanenko et al., 2004), incline (Rozumalski et al., 2017), bodyweight loading (McGowan et al., 2010), and stepping condition (Rouston et al., 2014). This observation also extends to clinical populations where synergy structure was similar following surgical intervention and gait biofeedback training in CP (Booth et al., 2019; Shuman et al., 2019) and across stepping patterns in stroke (Rouston et al., 2014), despite significant changes in gait. Further, children with Duchenne Muscular Dystrophy who had altered gait due to non-neurological muscle weakness (i.e. another form of biomechanical constraint) were shown to have similar synergy structure as ND children (Goudriaan et al., 2018; Vandekerckhove et al., 2020).

However, our study and prior work also suggest a relationship between changing biomechanical constraints and synergy activations. Shuman et al. (2019) observed that although synergy weights were unchanged following surgical interventions in CP, individuals whose synergy activations became more similar to ND trends after surgery had greater improvements in gait. Similar outcomes were reported by Rouston et al. (2013) who found that stroke survivors who improved synergy activation timing following training had greater improvements in walking performance. These findings suggest that for both ND and clinical populations, adaptability to imposed biomechanical constraints is maintained by tuning the activation of a consistent set of synergies.

4.2. Synergy complexity may be sensitive to gait pattern

Our results also highlight that changes in synergy activations can impact measures of synergy complexity . We observed that equinus-crouch, mild, and moderate crouch patterns did not affect complexity; however, equinus increased tVAF. This suggests that tVAF may be elevated for individuals who walk in an equinus pattern due to biomechanical constraints. Similar observations have been made regarding spasticity, as selective dorsal rhizotomy has been shown to increase tVAF₁ due to the pre-operative spasticity masquerading as more complex control (Shuman et al., 2019). These results have important clinical implications, as tVAF₁ has been used to predict treatment outcomes in CP. Prior research has suggested that children with CP who have higher tVAF₁ have greater improvements in gait after common interventions (Schwartz et al., 2016). As such, the sensitivity of synergy complexity to biomechanical constraints needs to be considered when using this measure in clinical decision making.

4.3. Emulated gait synergies do not align with synergies in CP

For all gait patterns, synergies were different between ND and CP cohorts, despite the groups being kinematically-matched. The CP cohort had increased coactivation of the hamstrings and quadriceps in all gait patterns, resulting in two synergies that generally aligned with swing and stance phase, similar to rhythmic-stepping patterns in infants (Dominici et al., 2011). In contrast, four unique synergies emerged in the ND cohort.. The additional synergies identified in the ND cohort may suggests that the ND nervous system is more flexible to changing biomechanical constraints and highlight a shift in reliance from supraspinal mechanisms to spinal circuitry to control gait following neurologic injury (Leonard et al., 1991). This is supported by Rouston et al (2014) who reported that post-stroke patients with more severe impairment (i.e., less complex control) were unable to selectively tune synergy activation to accommodate task demand.

4.4. Methodological considerations

When evaluating the results of this study, certain limitations need to be considered. Gait patterns were performed in a non-randomized order (i.e., Eq, Eq-Cr, Mi-Cr, Mod-Cr) introducing fatigue as a potential confounder. To mitigate fatigue, we required one-minute seated breaks between trials and closely monitored fatigue levels. Further, as prior work has reported that synergies are unchanged during cyclic tasks performed to failure, we do not expect fatigue to significantly influence our reported outcomes (Turpin et al., 2011). Participants also had minimal experience walking in the emulated patterns, which may have introduced learning effects. While we attempted to attenuate these effects by only analyzing the final trial for each pattern, we recognize that motor control strategies may change as a function of experience on a time scale longer than we could capture (Sawers et al., 2015). However, prior work has demonstrated that repeated exposure to novel tasks decreases co-contraction and tVAF (Sawers et al., 2015). As such, if participants were given a longer acclimation time for each pattern, we would expect tVAF values to decrease, and, therefore, diverge further from CP values.

To compare ND emulation results to CP, we had to rely on retrospective data. These data were collected as part of routine care and therefore, only included overground trials with a limited muscle set and number of strides. Because smaller data sets generally overestimate tVAF, we recalculated ND synergies using the same number of strides and muscles as the CP cohort (Oliveira et al., 2014; Steele et al., 2013). While we were not able to control for differences in walking condition, we would expect the ND cohort to walk with greater variability overground (Mileti et al., 2020) which would decrease reported tVAF values and increase the difference between cohorts. During matching, we did not exclude individuals based on injury severity or intervention history. This was a consequence of our hypothesis; if a direct tie existed between gait and synergies, kinematically-matched cohorts should recruit the same synergies regardless of other inter-cohort differences. Similarly, because of the typical age intervention in CP, there was a significant difference in age between ND and CP cohorts (p < 0.001). However, prior work by Dominici et al (2011) demonstrated that synergies stabilize early in ND development and are not expected to change significantly between the ages tested. Finally, we did not match frontal plane kinematics between cohorts. While hip adduction kinematics generally aligned, others, particularly at the pelvis, deviated between cohorts. This highlights an inherent limitation of emulation studies in CP; many gait patterns are multiplanar making them difficult to comprehensively emulate. However, because the muscle set analyzed primarily drives sagittal plane motion, we do not expect that matching frontal plane kinematics would significantly impact our conclusions.

4.5. Conclusion

This study suggests that altered synergies are not primarily a reflection of gait patterns in CP, but likely capture underlying neurologic pathology. While these results lend credence to the clinical use of synergy analysis, they also suggest that targeting gait improvements during intervention may not produce meaningful changes in synergies. Whether synergies can be changed in CP remains largely unknown. There is evidence that synergies can be modified with long-term training (Sawers et al., 2015) or exoskeleton assistance (Ranganathan et al., 2016) in ND adults. Further, recent research has suggested that resistance training with an exoskeleton in CP may improve synergy complexity (Conner et al., 2021). However, more work is needed to understand the plasticity of synergies in CP. Identifying factors that influence synergies will further inform the use of synergy analysis in clinical care and guide the development of interventions that can more consistently improve function for individuals with CP.