Task complexity amplifies spatial asymmetry of muscle synergy plasticity in chronic stroke survivors

Abstract

Background

Motor synergy patterns are recognized as physiological markers of motor cortical damage, providing insights into how motor cortex coordinates spinal motor modules to generate movements. However, how these patterns adapt to tasks of varying complexity following post-stroke cortical damage is not yet fully understood.

Objective

We aimed to understand how motor synergy patterns are distorted across tasks of increasing complexity after stroke induced cortical damage, also to provide a reference for task selection when using muscle synergy patterns as biomarkers for stroke evaluation or intervention.

Methods

This was a pilot, cross sectional study. We investigated muscle synergies during five tasks with varying complexity in 20 healthy individuals (13 females and 7 males, aged 64.33 ± 6.94 years) and in 12 participants with chronic stroke (4 females and 8 males, aged 64.4 ± 6.54 years). Surface electromyographic activities were recorded from 16 upper limb muscles (8 muscles per limb: upper/lower trapezius, anterior/posterior deltoid, triceps brachii lateral head, biceps brachii short head, flexor digitorum superficialis, and extensor digitorum communis). Non-negative matrix factorization was performed to extract the muscle synergies. We categorized the stroke-induced synergy plasticity based on healthy synergy centroids, compared the synergy plasticity between affected and unaffected limbs, and investigated the correlation between synergy plasticity and patient’s motor function,

Results

In healthy individuals, the number of muscle synergies exhibited a U-shaped pattern as task complexity increased, whereas in stroke patients, both the affected and unaffected limbs showed a decreasing trend in muscle synergy number with increasing task complexity. Besides, the unaffected arm exhibited significantly more preservation synergies (synergies resembling healthy patterns) than the affected arm in moderate (Placing 30 cm: Z(11) = -2.144, corrected p = 0.031, Rosenthal’s r = -0.646) and high complexity tasks (Z(11) = -2.558, corrected p = 0.028, Rosenthal’s r = -0.771), and fewer mutation synergies (synergies deviating from healthy patterns) with marginal significance (Placing 30 cm: Z(11) = -1.992, corrected p = 0.058, Rosenthal’s r = -0.600; Drinking: Z(11) = -2.070, corrected p = 0.058, Rosenthal’s r = -0.624). Notably, this asymmetry in preservation synergies was significantly correlated with patients’ motor function (Fugl-Meyer Assessment-Upper Limb: R = -0.711, permutation p = 0.010; Modified Ashworth Scale for elbow flexion: R = 0.603, permutation p = 0.044).

Conclusion

This study is among the first to investigate how task complexity influences muscle synergy plasticity and its asymmetry in participants with chronic stroke. Patients demonstrate spatial asymmetry in muscle synergies between the unaffected and affected sides. This asymmetry is magnified by task complexity and shows a strong correlation with motor performance. Therefore, we recommend that when using muscle synergy patterns as biomarkers for stroke assessment, the influence of task complexity should be explicitly considered, as it plays a critical role in shaping these patterns.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12984-026-01889-9.

Introduction

Impairment of upper limb motor function is one of the most common consequences of stroke, significantly affecting the ability to reach, grasp, and manipulate objects, thereby disrupting patients’ daily lives and social participation [1]. Motor synergy patterns have been evidenced as physiological markers of motor cortical damage, reflecting the way motor cortical areas orchestrate motor modules in the spinal cord to generate movement [2–4]. The current mainstream view is that descending cortical signals serve as neuronal drives that select, activate, and flexibly combine muscle synergies specified by networks in the spinal cord and/or brainstem [5]. This multilayered architecture enables the motor system to manage its large number of degrees of freedom efficiently. However, when cortical outflow is disrupted, as in stroke, the ability to process and execute tasks of varying complexity becomes significantly impaired, producing cascading effects on the control of motor modules in the spinal cord. Stroke survivors often exhibit altered muscle synergy patterns, reflecting neural reorganization and compensatory strategies after cortical damage [6, 7]. These distortions in motor synergies have been identified as potential biomarkers for assessing motor recovery and guiding rehabilitation strategies [8]. To gain deeper insight into these disruptions, we investigated the muscle activation patterns of stroke survivors performing tasks with different levels of complexity. Our aim was to elucidate how motor synergies are distorted across tasks of increasing complexity following cortical impairment and to provide a reference for selecting tasks when using muscle synergy patterns as biomarkers for stroke evaluation and intervention.

Preservation, merging, and fractionation are three distinct patterns of muscle synergy plasticity that reflect the multiple neural responses occuring poststroke [2]. In patients with mild impairment, as indicated by higher FMA scores, muscle synergies in the stroke-affected upper limb closely resemble those in the unaffected upper limb. This preservation suggests that, despite observable variations in motor performance, the ability to recruit and coordinate muscle synergies at the spinal level remain largely intact. However, as impairment severity increases, a noticeable divergence between the muscle synergies of the affected and unaffected upper limbs emerges. In severely affected limbs, two distinct synergy plasticity patterns—merging and fractionation—are commonly observed. Merging involves the combination of multiple muscle synergies into a single, less complex synergy, reflecting a loss of motor flexibility and adaptability. In contrast, fractionation refers to the splitting of a single muscle synergy into multiple fragmented components, indicating inefficient and disorganized motor control. Previous studies have mainly observed preservation, merging, and fractionation as synergy plasticity patterns in the affected upper limb relative to the unaffected limb, concluding that stroke primarily disrupts descending cortical signals while sparing spinal motor module control. However, we contend that the unaffected upper limb in stroke patients is an unreliable reference due to documented alterations in its neuromuscular control and motor behavior [9, 10]. These changes, including compensatory motor strategies and modified muscle synergy patterns, can confound comparisons and obscure the true extent of neural and muscular adaptations post-stroke. To address this issue, we used healthy individuals as the reference standards. By clustering muscle synergies from healthy individuals across tasks of varying complexity, we established robust synergy centroids as references. Using this baseline, we identified four synergy plasticity patterns in stroke patients: preservation, merging, fractionation, and a newly discovered pattern, mutation. This pattern was observed not only in the affected upper limb but also, notably, in their unaffected upper limb.

In this study, we designed five upper-limb tasks with varying levels of complexity to simulate common daily activities, accommodate the motor abilities of stroke patients, and align with established methodologies in the field [11]. These tasks were performed by both stroke patients and healthy participants. Muscle synergies were extracted from EMGs using non-negative matrix factorization (NMF), applying a variance accounted for (VAF) threshold of 0.98 to ensure accurate EMG reconstruction. The muscle synergies of the affected and unaffected limbs of stroke patients were compared to the healthy synergy centroids to calculate the proportions of the four distinct synergy patterns. We then analyzed the differences between the affected and unaffected limbs of stroke patients for each synergy pattern, investigating how muscle synergy distortion and imbalance evolved with increasing task complexity. We hypothesized that task complexity amplifies stroke-induced asymmetry of muscle synergy patterns, and that asymmetry is related to motor function. This study provides insights into how task complexity influences muscle synergy patterns in stroke survivors, contributing to a more nuanced understanding of motor control impairments and their relationship to functional tasks demand.

Methods

This study was a pilot, cross-sectional study. The study followed the guidelines of the Standard Protocol Items: Recommendations for Interventional Trials (SPIRIT) [12]. Ethical approval was obtained from the Human Subjects Ethics Sub-committee of the Hong Kong Polytechnic University (HSEARS 20240125003).

Participants

Stroke participants were recruited from hospitals and the Hong Kong Stroke Association according to the following inclusion criteria: (1) Age > 18 years; (2) First-time ischemic or hemorrhagic stroke; (3) The affected hand can hold a cup and drink water; (4) Modified Ashworth Scale (MAS) < 3; (5) Time from onset of stroke ≥ 6 months; (6) Sufficient cognitive ability to follow experimental procedure (Mini-Mental State Examination Score, MMSE ≥ 20). The exclusion criteria were: (1) Severe verbal comprehension deficit, apraxia, and/or visuospatial neglect; (2) Severe respiratory and circulatory failure; (3) Presence of non-stabilized fractures; (4) Traumatic brain injury; (5) Drug-resistant epilepsy; (6) Swallowing disorders and (7) Any pain during passive and active range of upper limb movement. Age-matched healthy participants with no history of neurological or orthopedic disorders were recruited through community posters, online platforms, and social networks. All participants signed an informed consent form prior to enrollment.

System setup and experimental procedure

System setup

Before the experiment, stroke patients were instructed to complete the cognitive assessment (MMSE), motor function assessments (FMA-UE and ARAT) and muscle spasticity assessments (MAS). These assessments were administered by a therapist who did not participate in EMG data collection. Thereafter, all participants were guided to perform five tasks (Fig. 1A): maximum isometric contraction of elbow extension and flexion, placing an object on a low platform (10 cm), placing an object on a medium platform (20 cm), placing an object on a high platform (30 cm), and drinking from a high platform (30 cm). The platform heights were selected based on established ergonomic principles, corresponding to the Neutral Reach Zone, Secondary Zone, and Tertiary Zone, ensuring realistic daily reaching contexts [13]. Detailed task descriptions and complexity definitions provided in Table 1 and Supplementary Fig. 1. An experimenter provided real-time, step-by-step instructions and monitored the experimental procedures. EMGs were recorded at a sampling rate of 1000 Hz using a wireless multichannel EMG system (The Biometrics Ltd., DataLITE, 16 channels; Fig. 1B). Surface electrodes were placed along the longitudinal midline of the target muscles, following the specifications and anatomical guidelines outlined in the Surface Electromyography for the Non-Invasive Assessment of Muscles [14]. The activities of eight muscles in both limbs were recorded simultaneously (Fig. 1B): upper trapezius (UT), lower trapezius (LT), anterior deltoid (ANDE), posterior deltoid (PODE), triceps brachii lateral head (TBLH), biceps brachii short head (BBSH), flexor digitorum superficialis (FDS), and extensor digitorum communis (EDC). Signal quality was confirmed through muscle test contractions.

Fig. 1.

A schematic illustration of the experimental design. A Tasks with increasing complexity. B Eight muscles for collecting EMG signals in each upper limb. C Raw EMG signals and envelopes extract from the raw EMG signals. D The change of Variance Accounted For (VAF) with increasing number of muscle synergies. The VAF exceeds our predetermined threshold (0.98) at Number of Synergy = 4. E A schematic illustrating how muscle synergies are linearly combined to generate EMGs. F The comparison of EMG waveforms resulting from the activations of individual synergies, and then summed together to reconstruct the EMGs (black lines) and recorded real EMGs (blue lines)

Table 1.

Task characteristics and complexity levels based on Motion, Coordination, and dynamics

Task	Content	Range of motion	Muscle coordination	Temporal dynamics	Task steps	Overall complexity
MVCE/F	Conduct IMVC of elbow extension and flexion for 5s (three repetitions of 2 minutes [20])	Static	Single muscle activation	No dynamics	Single step	Low
Placing low (10 cm)	A 100 g wooden block was placed 30 cm from the table edge (Fig. 1A) [21]. Upon a ‘ready and go’ command, participants used their affected, non-affected limb, or both to place the block on 10, 20, and 30 cm platforms for three repetitions with 10-second intervals, rested for 5 minutes [22], and repeated the task on 20 and 30 cm platforms [23].	Small range	Proximal muscles (elbow-dominant)	Low dynamics	Single step	Low
Placing moderate (20 cm)		Moderate range	Proximal muscles (elbow-dominant)	Low dynamics	Single step	Low to Moderate
Placing high (30 cm)		Moderate range	Proximal muscles (elbow-shoulder coordination)	Low dynamics	Single step	Moderate
Drinking	A half-filled plastic cup was placed 30 cm from the table edge on a 30 cm platform (Fig. 1A). Participants, seated with wrists on the table’s edge, reached for the cup, used wrist pronation to take a sip, and returned it to its original position.	Moderate range	Proximal + distal muscle coordination	High dynamics	Multi-step	High

MVCE/F  maximum isometric contraction of elbow extension and flexion

Task complexity

Participants were instructed to sit beside a table, with height adjusted to create a 30° angle between their shoulders and the tabletop. The sequence of the five tasks were randomized for each participant; details are provided in Table 1. Each task was repeated three times at a natural speed, starting with the dominant or unaffected arm [15, 16]. The choice of three repetitions was based on previous evidence showing that three trials yield good-to-excellent test–retest reliability for most kinematic variables in stroke populations [16]. At the beginning of each task, the experiment operator pressed the recording button and instructed the participant to begin. After the participant completed three continuous repetitions, the operator stopped the recording. The recorded data were used for Non-negative Matrix Factorization to extract muscle synergies. We defined target complexity, temporal dynamics, and overall complexity based on specific task characteristics. Target complexity was determined by the precision and variability of the target: tasks with static and easily reachable targets, such as the elbow MVCE/F, reaching tasks at 10 cm, 20 cm, and 30 cm, were classified as having simple targets, while tasks involving multi-step or dynamic goals, such as the drinking task, were classified as having complex targets. Temporal dynamics referred to the time synchronization demands of the task: MVCE/F was static and thus had no dynamics; reaching tasks allowed for self-paced execution, making them low dynamics; and the drinking task required precise timing and coordination between actions, making it high dynamics. Overall complexity was determined by integrating factors such as range of motion, muscle coordination, target complexity, and temporal dynamics, which is similar to other studies involving task complexity as a factor [17–19]. Based on these criteria, task complexity was categorized progressively from low (e.g., MVCE/F, placing task at 10 cm) to moderate (e.g., placing task at 20 and 30 cm) to high (e.g., drinking task).

Non-negative matrix factorization to extract muscle synergies

Various methods, including Principal Component Analysis [24], Independent Component Analysis [25], and Factor Analysis [26], have been used to extract and calculate muscle synergy patterns [26], each offering performance advantages [27]. However, these methods allow negative values, producing spatial synergies with negative weightings that do not align with the physiological nonnegativity of muscle activation and are difficult to interpret. Evidence supports the increased neurophysiological relevance of factors derived from Non-Negative Matrix Factorization (NMF) [28]. Therefore, NMF [29] has become the preferred method for studying muscle synergies in motor-related diseases, as it ensures non-negative results and provides a more physiologically meaningful representation of muscle activations.

NMF was applied to EMG signals from eight upper limb muscles to extract muscle synergies, following a preprocessing pipeline to ensure biologically meaningful input. Raw EMG signals were first band-pass filtered (20–450 Hz) to remove noise and artifacts, rectified by taking their absolute values, and then smoothed to extract their envelopes. The envelope was computed using MATLAB’s envelope function in RMS mode, which calculates the RMS of the signal within sliding windows of a specified size. The RMS envelope is defined by:

where the window size was N = 500 in this study, and represents the signal samples within the window. This method captures the energy profile of the signal, providing a smooth and biologically relevant representation of muscle activation intensity (Fig. 1C). The processed EMG matrix V (8×T) was decomposed into a synergy matrix W (8×r) and activation matrix H (r×T) by minimizing reconstruction error () under non-negativity constraints (Fig. 1DEF). This revealed a small number of synergies (r), representing coordinated muscle activation patterns and their time-varying activations. In addition, the silhouette score, which is the number of synergies in the NMF, were used to evaluate cluster quality of the muscle synergies [30]. This approach also determined the optimal number of synergies, thereby improving the objectivity of the analysis.

Synergy categorization

To study the distortions in muscle synergy patterns in the affected and unaffected upper limbs of post-stroke patients, we used the muscle synergy centroids from healthy individuals as references. K-means algorithm was used to cluster muscle synergy centroids from the dominant side of healthy individuals, minimizing within-cluster variance via . where the K is the number of clusters, represents the set of data points (synergies) in the k cluster. ​ is the squared Euclidean distance between a data point and its cluster centroid. The silhouette coefficient was employed to evaluate clustering quality. Where is the average intra-cluster distance for point , is the smallest average distance of to all points in other clusters. The resulting cluster centroids, representing dominant-side synergy patterns in healthy individuals, were used as a baseline to analyse synergy deviations in stroke patients. These centroids are shown in Supplementary Fig. 2.

Through observation, we categorized the plasticity patterns of muscle synergies into four types: Preservation, Merging, Fractionation, and Mutation (in Fig. 2). Each synergy from stroke participants was compared to healthy centroids and categorized based on their cosine similarity (). Synergies that exceeded a predefined similarity threshold (0.80) [2] with any healthy centroid were categorized as “Preservation” (see preservation example in Fig. 2). For non-preserved synergies, further classification involved assessing their similarity to linear combinations of healthy centroids (“Merging”, see merging example in Fig. 2). or determining whether the synergy could be represented as a linear combination of other synergies within the same subject to reconstruct any healthy synergy centroid (“Fractionation”, see fractionation example in Fig. 2). Synergies that did not fit into the above categories were classified as “Mutation” (see mutation example in Fig. 2), indicating a pattern distinct from those observed in healthy individuals. The linear combination weights were determined using non-negative least squares [31]. A flowchart illustrating this procedure is provided in Supplementary Fig. 3.

Fig. 2.

Examples of muscle synergy plasticity patterns (Preservation, Merging, Fractionation, and Mutation) in the limbs of stroke participants compared to muscle synergy centroids from healthy participants. In Preservation plasticity, the affected limb synergy (A1, MVCE/F) closely matched a healthy synergy centroid (H centroid 6; cosine similarity = 0.90). In Merging plasticity, the affected limb synergy (A5, placing 20 cm) was reconstructed by merging three healthy synergy centroids (H centroid 2, 3, 5). In Fractionation plasticity, the affected limb synergy (A1, placing 30 cm) was fractionated by a healthy synergy centroid (H centroid 6). In Mutation plasticity, the unaffected limb synergy (U5, placing 10 cm) had limited similarity to any healthy centroid (highest similarity = 0.63) but showed morphological resemblance to H centroid 3, suggesting it is a mutated version of this centroid

Statistical analysis

Muscle synergy numbers extracted using NMF were analyzed to compare differences across tasks (MVCE/F, placing 10 cm, 20 cm, 30 cm, and drinking) and three limb categories (affected limb, unaffected limb, and healthy dominant limb). The statistical analyses were structured as follows: (1) Healthy vs. Affected limb: A two-way repeated measures ANOVA was conducted, with group (Healthy vs. Affected) as the between-subject factor and task as the within-subject factor. (2) Healthy vs. Unaffected limb: A two-way repeated measures ANOVA was conducted, with group (Healthy vs. Unaffected) as the between-subject factor and task as the within-subject factor. (3) Affected vs. Unaffected limb: A two-way repeated measures ANOVA was conducted with task and limb side (Affected vs. Unaffected) as within-subject factors, reflecting the paired design for limbs within the same individuals. Levene’s test (p > 0.05) confirmed the homogeneity of variances for between-group comparisons. Mauchly’s test was used to verify the sphericity assumption for the repeated measures factors, and Greenhouse-Geisser corrections were applied where sphericity was violated. False Discovery Rate (FDR) corrections, specifically the Benjamini-Hochberg method, were used to adjust for multiple comparisons [32].

Once all muscle synergies from stroke limbs were classified into Preservation, Merging, Fractionation, and Mutation plasticity, we analyzed the difference in the number across the five tasks between the affected and unaffected limbs. Since not all of the data meet normality assumptions (tested by Anderson-Darling test, p < 0.05), we employed non-parametric Wilcoxon Signed-Rank Tests to evaluate differences between the affected and unaffected limbs for each task. FDR corrections were applied to control for multiple comparisons across tasks.

We also investigated the relationship between synergy asymmetry (Preservation, Merging, Fractionation, and Mutation) with motor function (FMA-UE, ARST, and MAS) and stroke duration (Time since stroke onset) of patients. Synergy asymmetry is defined as the difference in the number of muscle synergies between the unaffected side and the affected side, calculated as the number of synergies on the unaffected side minus those on the affected side. Correlation analyses were conducted using both parametric and non-parametric methods depending on the distribution of the data. The normality of the variables was assessed using the Anderson-Darling test, with results guiding the selection of appropriate correlation measures. For normally distributed variables, Pearson correlation was applied. For non-normally distributed variables, Spearman correlation was used. Significance of the correlations was determined using permutation tests to account for potential biases in the data, with 10,000 iterations performed for each test to derive robust p-values. All statistical analyses were performed using a significance threshold of p < 0.05.

Results

Subject demographics

From May to July 2024, 12 chronic stroke patients and 20 age-matched healthy individuals were recruited for this pilot study. The demographic characteristics of the participants are presented in Table 2. As a pilot study, the sample size followed methodological recommendations, suggesting 10–30 participants per group [33–35]. In particular, Julious (2005) proposed a “12 per group” rule of thumb for pilot studies [33]. Thus, the final sample of 12 stroke participants and 20 healthy controls is consistent with this recommendation. Chi-square tests revealed no significant difference in sex (χ²(1) = 3.04, p = 0.081) or handedness (χ²(1) = 3.56, p = 0.059) between groups. A sample t-test revealed no significant age differences (t(30) = 0.029, p = 0.977).

Table 2.

Demographic characteristics of participants

Variables	Stroke group (N = 12)	Healthy group (N = 20)	Sig.(p)
Gender (m/f)	8/4	7/13	0.08
Affected side (R/L)	6/6	–	–
Dominant hand (R/L)	10/2	20/0	0.06
Age (yrs)	64.33 ± 6.94	64.4 ± 6.54	0.977
Time since stroke (months)	70.58 ± 47.48	–	–
FMA-UE total	52.5 ± 7.3	–	–
FMA-UE shoulder	27.08 ± 3.94	–	–
FMA-UE wrist	7.67 ± 1.87	–	–
FMA-UE finger	11.75 ± 4.33	–	–
MMSE	28.75 ± 2.59	–	–
ARAT	43.75 ± 14.09	–	–
ARAT grasp	14.42 ± 5.07	–	–
ARAT grip	10 ± 3.1	–	–
ARAT pinch	11.5 ± 5.87	–	–
ARAT gross	7.75 ± 1.35	–	–

The data are expressed as mean ± SD. m male, f   female, R right side, L left side; yrs years; FMA-UE Fugl-Meyer Assessment- Upper Extremity, MMSE Mini-Mental State Examination, ARAT Action Research Arm Test

Parameter testing in the NMF

To determine the threshold of Variance Accounted for (VAF) in the NMF algorithm, we conducted parameter testing for each task. It can be observed in Fig. 3, across all tasks, the number of muscle synergies in the affected limb of stroke patients was generally lower than in the unaffected limb and the dominant limb of healthy participants. This finding aligns with our expectations and previous research. However, the degree of distinction varied significantly depending on the VAF threshold. Regions showing clearer distinctions are highlighted with blue dashed lines. To ensure comparability across tasks, a unified VAF threshold was required. At VAF = 0.98, the muscle synergy numbers across all five tasks effectively distinguished the affected limb, unaffected limb, and healthy dominant limb. Therefore, VAF = 0.98 was selected as the unified threshold for this study.

Fig. 3.

The curves of the muscle synergy number with the threshold of variance accounted for (VAF) in the task of: A elbow MVCE/F; B placing 10 cm; C placing 20 cm; D placing 30 cm; E drinking. Blue dash line marks the VAF = 0.98, where the muscle synergy numbers effectively distinguished the affected (Aff), unaffected (Unaff) and healthy dominant (Hdomi) conditions across the five tasks

Spatial characteristics of muscle synergy by tasks and groups

For stroke patients, the number of muscle synergies in both the affected and unaffected limbs showed a decreasing trend as task difficulty increased (Fig. 4A). In contrast, for the dominant limb of healthy individuals, the number of muscle synergies followed a U-shaped curve, first decreasing and then increasing as task complexity rose (Fig. 4A). The three-dimensional coordinate system was employed to clearly present results at both the individual and group levels: the X-Z plane displays the group mean and standard deviation, while the X-Y plane shows individual data points, with each point representing a participant. Figure 4B provides specific examples of muscle synergy numbers for a healthy dominant limb, a stroke unaffected limb, and a stroke affected limb across five different tasks. The synergy numbers for the Healthy dominant limb are 6, 3, 2, 3, and 6; for the Stroke unaffected limb, 6, 3, 2, 2, and 1; and for the Stroke affected limb, 5, 3, 2, 2, and 1 (MVC, Placing 10, 20, 30, and Drinking, respectively). These results align with the group-level trends shown in Fig. 4A.

Fig. 4.

A Comparison of muscle synergy numbers across tasks and groups. The X-Z plane presents the group mean and standard deviation, while the X-Y plane illustrates individual data points, with each point corresponding to an individual participant; B Examples of muscle synergy numbers for a healthy dominant limb, a stroke unaffected limb, and a stroke affected limb across five different tasks; C The results of statistical analysis of group/limb side and task effects on muscle synergy numbers. Hdomi: Healthy dominant side; Unaff: Stroke unaffected side; Aff: Stroke affected side

The muscle synergy numbers varied significantly across tasks in all comparisons (Healthy vs. Unaffected, Healthy vs. Affected, and Affected vs. Unaffected limbs; Partial η²: 0.215, 0.170, and 0.529, respectively; All p < 0.01; Fig. 4C). In the Healthy vs. Stroke unaffected limb comparison, neither the main effect of group nor the task × group interaction was significant, indicating no overall difference in muscle synergy numbers and similar task effects between the two groups. In the Healthy vs. Stroke affected limb comparison, a significant main effect of group was observed (F(1, 30) = 7.936, corrected p = 0.024, Partial η² = 0.209), with muscle synergy numbers being higher in the Healthy dominant limb, while the task × group interaction was not significant, suggesting similar task effects for the two groups. Finally, in the Stroke affected vs. Stroke unaffected limb comparison, neither the main effect of limb side nor the task × limb side interaction was significant, indicating no difference in muscle synergy numbers and similar task effects between the Affected and Unaffected limbs.

Figure 5A shows that the unaffected limb preserved more healthy-like synergies, whereas the affected limb had higher proportions of merging and mutation. These results were similar to those reported by Cheung et al. [2] Making a further step, we calculated the differences in synergy proportions between the two limbs to examine limb asymmetry. As shown in Fig. 5B, task complexity increased the asymmetry in preservation and merging plasticity, while fractionation showed no clear trend. Mutation imbalance followed a U-shaped trend, initially increasing and then decreasing with rising task complexity.

Fig. 5.

A Comparison of the proportion of muscle synergy preservation, merging, fractionation across tasks and groups based on the healthy synergy centroid. B The trend of imbalance of four types of muscle synergy in stroke participants (unaffected - affected) with increasing task complexity

Wilcoxon Signed-Rank Tests revealed that the number of preservation synergies on the affected side was significantly different from that in the unaffected side for tasks of moderate and high complexity (Placing 30 cm and Drinking tasks). Specifically, the preservation synergy numbers were lower in the affected side compared to the unaffected side for these tasks (Fig. 6A; Placing 30 cm: Z(11) = -2.144, corrected p = 0.031, Rosenthal’s r = -0.646; Drinking tasks: Z(11) = -2.558, corrected p = 0.028, Rosenthal’s r = -0.771). The mutation synergy numbers were higher on the affected side compared to the unaffected side during the two tasks, with a marginally significant p-value after multiple comparison corrections. (Fig. 6D; Placing 30 cm: Z(11) = -1.992, corrected p = 0.058, Rosenthal’s r = -0.600; Drinking tasks: Z(11) = -2.070, corrected p = 0.058, Rosenthal’s r = -0.624).

Fig. 6.

The differences in synergy numbers of: A Preservation, B Merging, C Fractionation, D Mutation between the affected and unaffected limbs across five tasks (MVCE/F, placing 10 cm, placing 20 cm, placing 30 cm, drinking)

Correlations between motor evaluation and muscle synergy pattern

Correlation analysis revealed significant associations between motor performance and asymmetry of preservation synergies during the high-complexity task (drinking task). Specifically, FMA-UE scores showed a significant negative correlation with the asymmetry of preservation synergies (Fig. 7A for the full correlation matrix, Supplementary Fig. 4 for the R values, and Fig. 7B for the masked correlation matrix), with R = -0.711 and p = 0.010 (permutation test). The corresponding scatter plot (Fig. 7E) demonstrates a clear downward trend, with a regression line and narrow confidence intervals further confirming this relationship. In contrast, MAS elbow flexion scores exhibited a significant positive correlation with the asymmetry of preservation synergies, as shown in the scatter plot (Fig. 7G, with the data points jittered to enhance visual clarity), with R = 0.603 and p = 0.044 (permutation test). The histogram of MAS elbow flexion (Fig. 7F) indicates that the data do not follow a normal distribution (Anderson-Darling Test: p = 0.047), justifying the use of Spearman correlation for this analysis. Furthermore, Fig. 7C and D illustrate the histograms of the preservation synergy asymmetry and FMA-UE scores, respectively, both of which follow a normal distribution (Anderson-Darling Test: p = 0.097 for Fig. 7C and p = 0.242 for Fig. 7D). Additional exploratory multivariate regression analyses and results are reported in Supplementary Analysis 1.

Fig. 7.

Correlation between motor performance metrics and significantly asymmetric muscle synergy patterns. A. Full correlation matrix. B Masked correlation matrix showing only significant correlations (p < 0.05, permutation test). C Histogram of preservation synergy asymmetry in the drinking task (Anderson-Darling Test: p = 0.097; data follow a normal distribution). D Histogram of FMA-UE scores (p = 0.242; data follow a normal distribution). E Scatter plot showing a significant negative correlation between FMA-UE scores and preservation synergy asymmetry in the drinking task (R = -0.711, p = 0.010, permutation test). F Histogram of MAS elbow flexion scores (Anderson-Darling Test: p = 0.047; data do not follow a normal distribution). G Scatter plot showing a significant positive correlation between MAS elbow flexion scores and preservation synergy asymmetry in the drinking task (R = 0.603, p = 0.044, permutation test). 1 + and 1- MAS scores were assigned 1.33 and 0.66, respectively, to distinguish them from integer scores for quantification. Only muscle synergies with significant asymmetry, as identified in prior analyses, were included in the correlation analysis

Discussion

This study is the first to investigate how task complexity influences muscle synergy patterns and their distortion in patients with stroke. Our findings demonstrate that complicated tasks requiring greater motor demands, such as Placing 30 cm and Drinking, significantly amplify the distortion and asymmetry of muscle synergies between the affected and unaffected limbs. In these high-complexity tasks, preservation and mutation synergies showed significant synergy asymmetry between limbs. Furthermore, correlation analyses revealed that higher asymmetry in preservation synergies during the Drinking task was strongly associated with reduced motor function and increased spasticity. All significant effect sizes (Partial η², Rosenthal’s r, and R) exceeded medium threshold, indicating substantial and meaningful effects. These results underscore the importance of considering task complexity when using muscle synergy patterns as markers of motor impairment in stroke survivors and suggest that tasks with higher motor demands may provide more sensitive assessments of stroke-induced motor deficits.

Classification of muscle synergy

Preservation, merging, and fractionation represent the three typical types of plasticity in muscle synergies observed in post-stroke patients [2, 36]. These plasticity patterns were found to correlate with the level of motor impairment, as assessed by the Fugl-Meyer Scale, and the time since stroke onset [2]. These observations can be perfectly explained within the framework of the hierarchical structure of motor control, which consists of a high-level (cortical) layer responsible for planning and coordinating motor tasks, and a low-level (spinal cord) layer that executes specific muscle activations through predefined synergistic modules [5, 37]. However, most prior studies have adopted the unaffected limbs as the references for evaluating changes in the paretic limbs [2, 5, 38]. While this approach offers the advantage of providing direct within-subject comparison and highlights the extent of motor impairment upon stroke, one potential limitation is that motor control of the non-paretic side may also be affected following stroke [39, 40]. Our findings, as shown in Fig. 5A, quantitatively demonstrate the deviations in muscle synergies of the non-paretic side relative to those of healthy individuals. As a result, using the non-paretic side as a baseline may not accurately capture the full spectrum of stroke-induced changes in motor control.

Our study addressed this limitation by using synergy centroids from healthy individuals as the reference to examine changes in muscle synergy patterns on both the paretic and non-paretic sides of stroke patients. Beyond the classical plasticity phenomena of preservation, merging, and fractionation [2, 4, 36, 41], we observed a phenomenon that aligns with what has been previously reported as systematic alterations in muscle synergies. Notably, Jinsook Roh et al. also described such alterations in upper limb muscle synergies in stroke survivors with severe motor impairment, distinguishing them from the typical preservation, merging, or fractionation patterns [40]. Building on these earlier observations, our study provides a quantitative characterization of this plasticity pattern, which we term ‘mutation’. Mutation synergy reflects the emergence of substantially altered muscle synergy structural patterns post-stroke, likely representing compensatory response within the central nervous system due to disrupted cortical control and altered descending inputs [42]. These disruptions can force spinal circuits to engage in compensatory reorganization, leading to the formation of atypical synergistic patterns. Our study provides new insights into the broader landscape of synergy plasticity following stroke.

Task complexity reveals divergent muscle synergy patterns in healthy individuals and stroke patients

Previous studies on muscle synergy differences between stroke patients and healthy individuals have been inconsistent, with some reporting no difference [5, 43, 44] and others noting fewer synergies in stroke patients [45, 46]. One potential cause for this inconsistency is the limited consideration of task difficulty. In this study, tasks were designed across multiple complexity levels, and significant effects of task difficulty were observed. Interestingly, contrary to our expectations, the relationship between task complexity and muscle synergy patterns differed significantly between healthy individuals and patients with stroke (Fig. 4A). In healthy individuals, the number of muscle synergies exhibited a U-shaped trend with increasing task complexity. Moderate-complexity tasks required fewer synergies, whereas both simple and highly complex tasks involved more synergies to ensure stability or adaptability. This nonlinear pattern may reflect an adaptive modulation of neural resource allocation within the sensorimotor system. During moderately challenging tasks that are neither trivial nor cognitively demanding, the central nervous system (CNS) likely achieves optimal efficiency through the selective activation of stable, well-trained neural modules, minimizing redundancy and energetic cost [47, 48]. In low complexity tasks, excessive degrees of freedom may exist, prompting the recruitment of additional synergies to stabilize performance. Conversely, at high complexity, increased cortical and subcortical engagement, heightened sensorimotor integration demands, and augmented feedback control likely require the recruitment of additional or overlapping neural synergies to maintain accuracy and postural stability [49–51]. In stroke patients, this nonlinear adaptation was no longer evident. Both the paretic and non-paretic sides showed a continuous decline in the number of synergies as task complexity increased. This finding suggests reduced flexibility in cortico-subcortical coordination and diminished capacity to dynamically reorganize motor modules, likely due to disrupted descending control pathways and compensatory reliance on fewer, less differentiated synergies [5, 40].

Task complexity amplifies synergy asymmetry in stroke patients

The asymmetry of muscle synergies reflects the redistribution of workload and control strategies in post-stroke patients. Previous studies have demonstrated that muscle synergies in the lower limbs are highly sensitive to symmetry in stroke patients, with lateral symmetry improving as a result of rehabilitation interventions [39, 52, 53]. Our results extend these findings to the upper limbs, showing consistent patterns of synergy asymmetry in the upper extremities of stroke patients (Fig. 5A). Moreover, we found that synergy asymmetry becomes more pronounced with greater task complexity (Fig. 5B). At the individual level, the number (modified) of preservation synergies was significantly higher in the unaffected than the affected arm during moderate- and high-complexity tasks (Placing 30 cm and Drinking tasks in Fig. 6). Conversely, mutation synergies–both in number and activation–were lower in the unaffected arm for these tasks (Drinking Task in Fig. 6 and Placing 30 cm in Supplementary Fig. 3). Notably, asymmetry in preservation significantly correlated with motor function in patients with stroke (Fig. 7). These findings highlight the critical importance of task selection in using muscle synergy analysis for stroke evaluation, as tasks of varying complexity elicit different synergy patterns. Higher complexity tasks are more sensitive to revealing asymmetrical and compensatory mechanisms, offering a more comprehensive assessment of motor function and recovery.

The parameter selection for NMF

Finally, we would like to discuss our experience in selecting the muscle synergy number during the extraction process using the NMF algorithm. In studies employing NMF for extracting muscle synergies, the most prevalent approach is to use Variance Accounted For (VAF) to determine the number of synergies required for a given task [41, 54–56]. Although this method is more robust than relying on experience to set the muscle synergy number, the selection of the VAF threshold itself remains somewhat subjective. For instance, some studies use a VAF threshold of 0.80 [56] or 0.90 [57], while most of them between 0.80 and 1.00 [58], leading to significant inconsistencies. In this study, we tested and plotted the relationship between muscle synergy number and VAF for the unaffected and affected sides of stroke patients, as well as the dominant side of healthy individuals. By analyzing these curves across different tasks, we identified the VAF ranges that best distinguished the three groups. Combining the optimal ranges across tasks, we determined that a VAF threshold of 0.98 provides the best balance. This testing methodology can serve as a valuable reference for determining the VAF threshold and muscle synergy number in future related studies.

The disadvantages and future direction

Several limitations should be considered when interpreting these findings. One limitation is the relatively small number of stroke patients, which may influence the generalizability of the results. Additionally, the wide range of stroke chronicity (6–240 months) may introduce variability, but synergy studies [2, 5, 40, 59, 60] indicate that neuromuscular patterns are largely stable in the chronic phase, and our within‑subject design further minimizes between‑subject differences. Another consideration is the sex distribution in our participant groups (stroke: 8 M/4F; controls: 7 M/13F). Although the chi-square test showed no significant difference (χ²(1) = 3.04, p = 0.081), the imbalance in sex ratios may still have influenced the outcomes due to known physiological differences. The study also employed three task repetitions per condition, which, while sufficient for initial analyses [16], could benefit from additional trials to enhance the stability of synergy extraction. Future studies should consider recruiting larger and more balanced samples to reduce variability and enhance generalizability. Increasing the number of task repetitions, refining task complexity levels, and testing a broader range of tasks under diverse conditions could further improve reliability.

In addition, it is worth noting that task complexity in this study was qualitatively classified as low, moderate, or high, following established multidimensional frameworks that encompass range of motion, muscle coordination, temporal dynamics and task steps [17–19]. In future studies, we plan to develop a more refined and systematic quantitative index of task complexity based on these parameters. Establishing such an index will enhance methodological rigor, improve reproducibility, and enable more precise modeling of the relationship between task difficulty and motor control performance.

Conclusion

Post-stroke muscle synergy plasticity on both the unaffected and affected sides can be categorized into preservation, merging, fractionation, and mutation patterns. Unlike healthy individuals whose muscle synergy numbers follow a U-shaped trajectory as task complexity increases, stroke patients exhibit a steady decline in muscle synergy numbers on both sides with increasing task complexity. Furthermore, patients with stroke demonstrate clear spatial asymmetry between limbs, which becomes more pronounced with greater task complexity and is strongly correlated with motor performance. These findings underscore the importance of considering task complexity when using muscle synergy for stroke assessment and rehabilitation.

Author contributions

R. Sun, Y. Wang, P. Cao, R. Song, and R. K. Y. Tong conceptualized and designed the study. Y. Wang, L. Zhong, D. Liao, H. Song, and Q. Meng, C. H. Fong were responsible for data acquisition, analysis, and interpretation. R. Sun and Y. Wang drafted the manuscript. All authors critically reviewed, revised, and approved the final version of the manuscript for submission.

Funding

This study was supported by The Hong Kong Polytechnic University (P0045217).

Data availability

The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request. Please note that the data are provided exclusively for research purposes.

Declarations

Ethics approval and consent to participate

The protocol was approved by the Institutional Review Board of The Hong Kong Polytechnic University (HSEARS20240125003). Participants gave informed consent before taking part in the study.

Consent for publication

Consent obtained from patients/family members.

Competing interests

The authors declare no competing interests.