Exoskeleton-Assisted Sit-to-Stand Training Improves Lower-Limb Function Through Modifications of Muscle Synergies in Subacute Stroke Survivors

Abnormal muscle synergies during sit-to-stand (STS) transitions have been observed post-stroke, which are associated with deteriorated lower-limb function and mobility. Although exoskeletons have been used in restoring lower-limb function, their effects on muscle synergies and lower-limb motor recovery remain unclear. Here, we characterized normal muscle synergy patterns during STS activity in ten healthy adults as a reference, comparing with pathological muscle synergy patterns in ten participants with subacute stroke. Moreover, we assessed the effects of a 3-week exoskeleton-assisted STS training intervention on muscle synergies and clinical scores in seven stroke survivors. We also investigated correlations between neuromuscular complexity of muscle synergies and clinical scores. Our results showed that the STS task involved three motor modules representing distinct biomechanical functions among healthy subjects. In contrast, stroke participants showed 3 abnormal modules for the paretic leg and 2 modules for the non-paretic leg. After the intervention, muscle synergies partially shifted towards the normal pattern observed in healthy subjects on the paretic side. On the non-paretic side, the synergy modules increased to three and neuromuscular coordination improved. Furthermore, the significant intervention-induced increases in Fugl-Meyer Assessment of Lower Extremity and Berg Balance Scale scores were associated with improved muscle synergies on the non-paretic side. These results indicate that the paretic side demonstrates abnormal changes in muscle synergies post-stroke, while the non-paretic side can synergistically adapt to post-stroke biomechanical deviations. Our data show that exoskeleton-based training improved lower-limb function post-stroke by inducing modifications in muscle synergies.


I. INTRODUCTION
S TROKE is a leading cause of long-term disability worldwide [1] and presents a strong need for lowerlimb rehabilitation. Post-stroke individuals spend more time completing sit-to-stand (STS) movements, during which they spontaneously move the trunk towards the unaffected side, resulting in asymmetrical weight bearing and higher falling risk [2], [3], [4]. Influenced by many physiological and psychological processes [5], STS is an important prerequisite to achieving many functional goals such as walking and daily activity [6], [7], and is commonly impaired and not easily recovered after stroke [8]. Regaining functional independence in STS movements as soon as possible after stroke is one of the challenges of stroke rehabilitation targeting lower-limb motor recovery. Currently, stroke rehabilitation guidelines recommend STS should be included in task-specific training [9]. However, during training, STS movements are typically assisted manually by physiotherapists, making high dosage of STS repetitive practice difficult to be achieved. One study reported the mean daily STS activity poststroke was below the level of frail older people receiving rehabilitation [10]. Although most stroke survivors can gradually develop the primary STS movement and transition to walking, a subset of stroke survivors who do not get sufficient practice with STS will have persistent mobility deficits that limit community participation. Furthermore, because most This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ motor recovery occurs in the early subacute stage (in the first three months) after stroke [11], it is essential for stroke survivors to gain independence in the STS and its component movements earlier, enabling recovery of lowerlimb function, dynamic balance, walking ability, and daily activities performance. Therefore, it is imperative to gain a better understanding of the complex and demanding processes underlying STS as well as STS training during subacute stroke.
Stroke damages the central nervous system and may cause abnormal motor control and disordered muscle activation resulting in impaired STS performance such as being timeconsuming or leading to falls [2]. Previous studies showed that muscle activation during STS was impaired on both sides with more severe impairments on the paretic side, such as delayed or reduced activation of the tibialis anterior, and earlier or excessive activation of the soleus muscle [12], [13]. However, previous studies evaluated individual muscles, while impairments in coordination of activity across muscles, and the underlying neurobiological mechanisms remain elusive [14]. Previous studies have shown that the nervous system simplifies movement control by a low-dimensional modular organization of muscle activation called muscle synergies (i.e. motor primitives) [15], [16], [17]. By analyzing muscle synergies, we could gain insight into how human coordinate different muscles to complete the STS tasks. Ting et al. proposed muscle synergies are shaped by evolutionary and learning processes through neuromechanical pathways for individual movement [18]. Neurologically-unimpaired participants adaptively changed their muscle synergies to accommodate changes in limb biomechanics during growth and development [19], to adjust to environmental perturbations like split-belt locomotion [20], or accomplish specific goals such as reducing knee contact forces [21]. After neurological disorders including stroke, cerebral palsy, and multiple sclerosis, muscle synergies are disrupted, the complexity of muscle recruitment patterns and neurological coordination of movement (i.e. neuromuscular complexity) is decreased, and temporal activation patterns are altered [22], [23], [24]. Although there have been a few studies on muscle synergies during STS in healthy adults [25], [26], [27] and stroke subjects [27], [28], [29], [30], it is unclear how muscle synergies are altered after rehabilitation. Synergy analyses have the potential to provide an objective and more in-depth evaluation of intervention effects [18].
Exoskeletons enable repetitive and intensive task-practice in individuals after stroke to restore normal STS motion. Sousa et al. showed that intensive STS training combined with conventional therapy improves STS ability after stroke [31]. Robotic-assisted training is an efficient and effective treatment for gait recovery [32], [33], and may also be effective in improving STS performance [34]. Recently one study found that exoskeleton-based rehabilitation helps to regain independent walking earlier for non-ambulatory subacute stroke survivors [35]. However, wearable lower-limb exoskeletons for rehabilitation are in the early stages of testing so far [36], thus more studies investigating effects of exoskeletons may provide a diverse and promising solution for stroke rehabilitation.
The objectives of this study were to: 1) compare muscle synergies during STS motion between healthy subjects and individuals with early subacute stroke on both non-paretic and paretic sides; 2) investigate how a 3-week exoskeletonbased intervention changed neuromuscular coordination during STS movements; 3) evaluate relationships between muscle synergies and clinical measures of balance, lower-limb function, and activities of daily living. We hypothesized that organization of muscle synergies would be improved after exoskeleton-assisted rehabilitation, and the improvement facilitates better performance of lower-limb motor tasks for stroke participants.

A. Participants
Ten healthy adults and ten post-stroke individuals participated in this study. Inclusion criteria for the stroke participants were: (1) age from 18 to 80 years, male or nonpregnant female; (2) first, early subacute stroke (2 weeks to 3 months), (3) non-ambulatory and with unilateral hemiparesis; (4) able to complete STS motion independently; (5) able to understand and follow directions and give informed consent; (6) meet requirements for wearing the exoskeleton: height between 1.5 meters and 1.9 meters, weight below 100 kg. Exclusion criteria were: (1) excessive spasticity in the lower limbs (≥3 on modified Ashworth scale); (2) presence of wounds or pressure ulcers; (3) recent lower-limb fracture or osteoporosis; (4) having other neurological conditions such as cerebral palsy, spinal cord injury, and Parkinson's disease; (5) co-existence of severe comorbidities: severely impaired cardiopulmonary function, malignant tumor, infections, etc. Table I summarizes the baseline characteristics of participants. However, after three weeks of training, participants #8 and #9 dropped out due to early discharge, and #10's muscle analysis data was unable to be precisely divided into time periods of STS due to device malfunction. Thus, we finally included surface electromyography (sEMG) data for analysis pre-intervention (Pre) and post-intervention (Post) on 7 stroke participants. All participants provided written informed consent before data collection. All ethical and experimental procedures were approved by the medical ethics committee of Tongji Medical College of Huazhong University of Science and Technology (no. [2020] S296-1).

B. Exoskeleton Device
The ProWalk (Chwishay Intelligence Technology Co., Ltd, Shenzhen, China; Figure 1(a)) is a battery-operated, comfortable, and powerful unilateral wearable lower-limb exoskeleton. ProWalk utilizes Inertial Measurement Unit (IMU) distributed at the waist and non-paretic side, as well as torque and angle transducer at the knee joint on the paretic side to monitor the angular velocity and acceleration of the user's hip joints on both sides, as well as the angle variation and torque information of the paretic knee joint. ProWalk primarily relies on a neural network that fuses these multichannel motion data to recognize the user's motion intentions during STS transfer and walking. It has an STS training  mode and a gait training mode, which can be selected by the physical therapist through software on a tablet and specify the assistance level.
For example, during the STS task, the therapist can choose an assistance level from rank 1 to 5, with a maximum assistance torque of 25 Nm. Once the expected motion is recognized by the exoskeleton using the aforementioned motion data, it applies the appropriate assistance torque to the affected knee joint at the appropriate time (Figure 1(b)). This torque is transmitted through the thigh frame to help the patient compensate for insufficient knee extension during the motion.

C. Experimental Procedures
For the STS test, each participant was asked to sit on a height-adjustable seat with their hips flexed at about 90 • , and knees flexed slightly more than 90 • to enable them to independently complete the STS movement. Subjects repeated STS movements 10 times at a self-selected comfortable speed without help either from their upper limbs or the experimenter, and rested 5∼10 s after each movement to minimize fatigue.
During the STS test, lower-limb sEMG data were collected for muscle synergy analysis. The STS test was conducted both Pre and Post intervention for stroke participants.
Each stroke participant received eighteen training sessions of exoskeleton-assisted physical therapy in addition to conventional therapy over the period of three weeks. During each training session, the participant completed 40-minutes of exoskeleton-assisted therapy, including 20 minutes of STS training, then 20 minutes of standing balance training, stepping practice, or gait training according to individual-specific goals. Clinical performance was evaluated using the Berg Balance Scale (BBS) to measure balance ability, the Fugl-Meyer Assessment of Lower Extremity (FMA-LE) to measure lowerlimb motor impairment, and the modified Barthel Index (MBI) to measure activities of daily living. All measurements were performed Pre and Post intervention.

D. Data Acquisition
Eight muscles from bilateral lower limbs were measured at a sampling frequency of 2,000 Hz using the wireless Ultimu EMG system (Noraxon USA Inc., Scottsdale, AZ, USA): the gluteus medius (GM), rectus femoris (RF), vastus medialis (VM), biceps femoris long head (BF), semitendinosus (SEM), tibialis anterior (TA), medial gastrocnemius (MG), and soleus (SOL). All EMG setup procedures followed the SENIAM guidelines [37]. Kinematic data were synchronously collected to segment STS motions with IMU system (Noraxon USA Inc.). Each IMU sensor assessed three-dimensional orientations and accelerations at a 200 Hz sampling frequency. Seven IMU sensors were placed on body segments, including pelvis, bilateral thighs, shanks, and feet.
E. Data Processing 1) Segment STS Motions: STS events detection was consistent with methods used in previous literature [38], [39], [40]. The start of STS was detected when the trunk angular velocity exceeded 10 deg/s from baseline, and the end of STS was detected when the trunk angular velocity was below 10 deg/s from baseline [38], [40]. STS movements for which the movement duration was more than twice the average or less than half the average compared to the mean movement time duration were excluded. Moreover, abnormal STS movements were ruled out using the visually identifying the reconstructed 3D avatar animation automatically generated by MR3 software that was matched with the Noraxon hardware. The remaining STS motions were subdivided into four phases as follows [25], [26], and [39]: (a) the flexion momentum phase began with the first horizontal shoulder movement to generate forward momentum, (b) the momentum transfer phase which began at seat unloading (SU), which has two parts divided by seatoff (SO), (c) the extension (EX) phase which began when the ankle dorsiflexion reached its maxima, (d) the posture stabilization (ST) phase began when the shoulder position reached its maximum height to keep the posture stabilization ( Figure 2). Also, the total duration of STS, maximum angle of anterior pelvic tilt, the percentage of start of STS to the time of SO in the total duration and joint range of motion (ROM) were calculated by IMU.
2) EMG Processing: Raw sEMG signals were band-pass filtered (20-300 Hz, 4th order zero-lag Butterworth filter) and then full-wave rectified, afterward a low-pass Butterworth filter (4 Hz, 4th order) was applied to get the sEMG envelope. The amplitude of sEMG envelope was normalized to the peak value in each channel, then the sEMG envelope of each STS motion was time-normalized to 100% of the STS motion [41]. Finally, to fully account for inter-trial variability, the motion matrix was generated by concatenating five cycles of data from each participant instead of averaging them [42].
3) Muscle Synergy Analysis: Muscle synergy analysis was performed by non-negative matrix factorization (NNMF) for each limb [43], [44]. In NNMF, an EMG data matrix (E o ) can be linearly decomposed to spatial structure (W ) and temporal structure (C) according to the equation: where Eo is a N-by-M matrix (N is eight in our study stands for channel number, and M is the time points of sEMG data, o stands for original), W represents relative activation weighting of each muscle synergy module (N-by-K matrix, K is the number of modules, varied from 1 to 8 in our study), C is a K-by-M matrix, represents the temporal activation pattern of each motor module, and D is the root mean square residual between E o and W C To minimize D and avoid the local minima, we calculated the synergies with two steps: first try a few iterations (max to 5) at 10 replicates using the multiplicative algorithm, then continue with more iterations (set to 1000) from the best of the results using alternating least squares [41]. We defined K using the variance accounted for (VAF) criterion. VAF provides a way to quantify neuromuscular complexity of muscle synergies, and higher VAF indicates decreased complexity of motor control [41]. The V AF describes the degree of variance in E o accounted for by the reconstructed sEMG matrix (E r ), and n is the number of muscles (8 in our study): K was defined as the minimum number of muscle synergies required to achieve at least 90% of the overall VAF and 75% of the VAF in each muscle. After determining K in each participant at each session, to find the common pattern across participants, we obtained the common number of muscle synergies by using the consistent number of muscle synergies found in the majority of individuals within the group. Then, we re-extracted muscle synergies according to the common number of muscle synergies at each group. Because the order of muscle synergies extracted by NNMF was random among subjects and repetitions, in order to sort the similar modules extracted from different subjects and legs, we calculated the cosine similarity [20]. The cosine similarity values range from 0 to 1, and higher values indicate greater similarity. If two modules in one participant were extracted into one muscle synergy module set, we checked their cosine similarity and the peak time of temporal activation visually [27]. Finally, we checked the sorted results manually to make sure of accuracy. After all the modules from each leg were put into different categories, we calculated averaged muscle synergy patterns in healthy controls and stroke survivors Pre and Post for further analysis. The analysis was performed in MATLAB (version R2021 b).

4) Statistical Analysis:
To investigate the differences between the healthy group and stroke group before intervention, we analyze each individual muscle channel of spatial structures and activation amplitude at each point of temporal structures between two conditions. To investigate the effects of exoskeleton on neuromuscular coordination of stroke participants, the same analysis was conducted Pre and Post. Independent t-test was used to check the significant differences ( p < 0.05) and the normal distribution was determined by the Kolmogorov-Smirnov test. Paired t-test was used to compare clinical scales (BBS, FMA-LE, and MBI) and kinetic variables measured by IMU Pre and Post. Wilcoxon rank-sum test was used when data did not follow a normal distribution ( p < 0.05). The comparison of joint ROM between groups was conducted using independent t-test, while paired t-test was used to compare within-group differences Pre and Post. Moreover, we evaluated correlations between neuromuscular complexity (overall VAF) of muscle synergies and clinical test scales using Spearman's rank correlation coefficient. In addition, to verify the representativeness of group average muscle synergy patterns, the inter-subject similarity of each group was compared using Pearson correlation coefficient (r -value). Figure 3 shows the overall VAF of different muscle synergies numbers for healthy subjects. We found that three motor modules were necessary to meet the VAF criterion for most healthy subjects (9 of 10 subjects). Figure 4 shows the averaged motor modules extracted from the STS motions of ten healthy subjects. Each muscle synergy had a particular composition and activates at different phases of the movement, associated with different human movements and biomechanical characteristics. Muscle synergy 1 (M1) primarily activated ankle dorsiflexor (TA) and knee extensors (VM and RF). The activation of M1 started during the last part of phase 1 to help generate forward momentum, and remained activated during the whole phase 2, associated with momentum transfer, namely ankle joint dorsiflexion to move The relationship between the number of modules and reconstruction quality for healthy subjects. The red vertical dashed line marks the minimum number of modules required to achieve the overall VAF > 90%.

A. Muscle Synergy Patterns in Healthy Subjects
the body forward and knee joint extension to raise the hip. Muscle synergy 2 (M2) primarily activated knee flexors (BF and SEM), it was mainly activated in the second half of phase 2 (after seat-off) and most of phase 3. After the knee extensors provide sufficient upward momentum and velocity during the momentum transfer phase, the knee flexors play a major role in the extension phase to help individuals achieve a stationary standing position. Muscle synergy 3 (M3) mostly activated hip abductor/adductor (GM) and plantar flexors (SOL and MG). It was activated in the second half of phase 3 and the entire duration of phase 4, and associated with body posture control and ankle plantarflexion to decelerate the movement.
The similarity between the synergy pattern of each subject in a group and group average synergy pattern was computed using Pearson correlation coefficient (r -value), as shown in Table II. Except for module 2 of the non-paretic side postintervention, the similarity of spatial structure and temporal structure in each group was greater than 0.4, and > 0.7 for most participant, which illustrated that the average synergy pattern represented most individuals' pattern pretty well. So, although individual differences existed, these average synergy patterns were good representativeness for the patterns of most individuals. Figure 5 shows the changes of muscle synergies on the nonparetic leg of stroke subjects during STS motions pre and post intervention. In the top-left region of the figure, it can be observed that, in comparison to healthy individuals, two modules were required for early subacute stroke patients on the non-paretic side to satisfy the VAF criterion (8 of 10 subjects). Following exoskeleton-assisted STS training, three modules were required (6 of 7 subjects).

B. Changes of Muscle Synergies on the Non-Paretic Leg
As showed in the bottom-left region of the figure 5, the nonparetic side differed from the typical three modules observed in healthy subjects. Module 1 (C1) elicited activation primarily in the muscles activated in M1 and M2. Meanwhile, module 2 (C2) elicited activation primarily in the muscles activated in   5. Changes of muscle synergies on the non-paretic leg of stroke subjects during STS motions. The relationship between the number of modules and reconstruction quality showed in the upper left part. A simple schematic of the spatial structure changes after stroke and exoskeleton treatment showed in the upper right part. In the bottom half of the figure, spatial structure for each muscle is plotted as mean ± standard deviation, each column in spatial structure represents group average across legs. Thick lines in the shaded plot represent the average temporal structure across legs, the shaded ribbons show the standard deviation range. The dashed lines in the temporal structure ribbon plots, and the horizontal lines in the Gantt charts represent the averaged muscle synergy of healthy subjects.  The changes in muscle synergies after exoskeleton training were depicted in the bottom-right region of figure 5. The spatial structure and temporal structure of module N1, N2 and N3 were similar with M1, M2 and M3, respectively. The fractionation of the modules observed after intervention indicates the recovery of muscle coordination. Figure 6 shows the changes of muscle synergies on the paretic leg of stroke subjects during STS motions pre and post intervention. In the top-left region of the figure, it can be observed that, in comparison to healthy individuals, two modules were required for early subacute stroke patients on the paretic side to satisfy the VAF criterion (7 of 10 subjects). Following exoskeleton-assisted STS training, three modules were required (5 of 7 subjects).

C. Changes of Muscle Synergies on the Paretic Leg
As showed in the bottom-left region of the figure 6, the paretic side preserved 3 motor modules similar to healthy subjects, but there was a profound disparity between these modules and normal modules. The P3 module roughly preserved the spatial structure and temporal structure of M3. The temporal structure of the P2 module was similar to M2, but the muscles primarily activated by P2 are significantly different from those in healthy subjects. The P2 module had greater muscle activation of knee extensors and lower activation of SEM compared to M2. As for the P1 module, it mainly activated ankle dorsiflexor and knee flexors, showed a significantly lower activation of knee extensors compared to M1. As for the temporal structure of P1, muscle activation was increased in part of phase 3 compared to M1. These extensive changes suggested impairment of muscle synergies.
The changes after exoskeleton training showed in the bottom-right region. In the spatial and temporal structure, there were no significant differences pre and post-intervention, but a important tendency of decreased activation of knee flexors and increased activation of knee extensors was found in P1, whereas in P2, the opposite trend was observed. And compared with healthy subjects, there were no significant differences except GM in P2. These changes showed abnormal muscle activations during the phase of momentum transfer and extension were reduced. Overall, the three weeks of exoskeleton-based therapy improved the abnormal activation of knee extensors and flexors.

D. Correlations Between Muscle Synergies and Functional Improvements
To investigate the impact of ProWalk on stroke participants' functional improvements, we compared BBS, FMA-LE, and MBI Pre and Post. As shown in Table III, significant improvements in BBS ( p<0.05), FMA-LE ( p<0.05), and MBI ( p<0.05) were found post-intervention. That indicates the exoskeleton was effective in improving balance ability, lowerlimb functions, and activity of daily living.
Kinetic variables and joint ROM of STS Pre and Post also reflected the functional ability of stroke survivors. As shown in Table IV, stroke survivors require more time to complete the STS task and need to increase their forward leaning angle to generate sufficient energy. Meanwhile, there was no significant change in the proportion of time spent before and after leaving the seat relative to the total time, indicating that the ability to generate sufficient kinetic energy before standing up and maintain balance after standing up were both impaired after stroke. There was a decrease in joint ROM on the paretic side and compensatory increase in ROM on the unaffected side, resulting in increased asymmetry between both sides. After training, participants showed a certain degree of recovery in their ability to perform the STS task, with improvements observed in joint ROM on the paretic side and symmetry between both sides (Table IV, Figure 7).
To further investigate if the clinical improvements could be attributed to the improvement of muscle coordination, we compared the relationship between the changes of overall VAF in each module number on both sides and changes in clinical scores Pre and Post. The decrease in overall VAF could represent the recovery of neuromuscular complexity. When the number of motor modules is three, there were significant correlations between intervention-induced VAF decreases of the non-paretic side with change in BBS (r = 0.946, p = 0.003) and change in FMA-LE (r = 0.847, p = 0.024) (Figure 8). This suggested that an increase in the motor performance of balance and lower-limb functions post-stroke may have a correlation with improvements in neuromuscular complexity of muscle synergies on the non-paretic side.  6. Changes of muscle synergies on the paretic leg of stroke subjects during STS motions. In the upper right part, the red rectangular boxes highlighted the major abnormal changes in the spatial structures occurred after stroke. In the bottom half of the figure, the horizontal lines in the spatial structure bar plots, the dashed lines in the temporal structure ribbon plots, and the horizontal lines in the Gantt charts represent the averaged muscle synergy of healthy subjects. Bars with asterisks indicate a significant contribution difference between the stroke group and healthy controls in the particular muscle channel. The yellow rectangles in temporal structure plots highlight the phases when the activation intensity significantly differs from healthy subjects. * , * * : compared with healthy subjects.

A. Muscle Synergy Patterns During Sit-to-Stand Movements in Adults
We recruited ten healthy adults to obtain normal muscle synergy pattern, and their age distribution did not show significant differences compared to stroke participants ( p = 0.089). For our small sample study, our results showed that muscle synergies are similar across adults and do not change with age (between young and old healthy adults) [26]. Our results obtained three modules for healthy adults, different from the four modules of the previous study conducted by Yang et al. during STS movements [27]. In addition to lowerlimb muscles similar to ours, they also measured the trunk muscles. Our modules were similar to the last three modules of their study. The first module we lacked was mainly activated in phase 1, which was produced by rectus abdominis and abdominal external oblique muscle to generate the momentum for STS motion. We did not measure the trunk muscles, so we lacked muscle activation in most of phase 1, which can also be found in the Gantt chart of in figure 4. In another study that only investigated the muscles of the lower limbs, they got three modules similar to ours [26]. Therefore, the muscle synergy pattern in our observation basically outlined the lowerlimb muscles of healthy adults, the consistency with previous literature and compliance with human biomechanics further support the validity of the muscle synergy extraction method used in our study.
The selection of the appropriate number of muscle synergies plays a critical role in muscle synergy analysis, and currently there is no definitive method. We used a classical approach, which set a threshold for the overall VAF and constrained the VAF of each muscle to ensure accurate reconstruction of the original data [45]. In recent times, novel algorithms have been introduced, including an Akaike Information Criterion (AIC)-based method [46], and an algorithm that guarantees inter-trial consistency and low similarity across synergies [47]. These algorithms have been validated in healthy populations, showing higher accuracy than the VAF-based methods. In order to apply them in clinical trials, further validation is needed for stroke survivors.

B. Muscle Synergy Patterns After Stroke
The training of STS transitions is important during the early subacute stage of stroke, and learning muscle synergy patterns underlying the task helps us better understand the neuromuscular change after stroke. Previous studies have demonstrated many abnormal changes in STS tasks after stroke, including the lateral deviation of the trunk to the nonparetic side and asymmetrical weight bearing observed before seat-off [2]. That is to say, due to the damaged coordination of the paretic leg, the non-paretic leg is forced to reduce the neuromuscular complexity to compensate for the insufficiency of completing the STS task. In our study, we noticed abnormalities in muscle synergies on the non-paretic leg ( Figure 5). In C1, it seemed to be formed by merging M1 and M2. Co-activation at the onset of body extension compared to M1 may aid with faster momentum transfer followed by larger vertical momentum to help move the body upward, which is consistent with the earlier activation of knee flexors exhibited on the non-paretic leg [48]. Moreover, C2 seemed to be formed by merging M2 and M3. Simultaneously activating the knee flexors, plantar flexors, and hip abductor/adductor helped maintain balance during body extension and to avoid the possibility of falls. This corresponds to the early SOL activation time compared to healthy subjects [13].
On the paretic side (Figure 6), the number of muscle synergies is the same as for healthy people, but the temporal and spatial structures have been significantly changed. In our study, the most profound changes occurred in P1, whose activating level abnormally increased in phase 3 compared to M1, accompanied by higher activation of knee flexors and lower activation of knee extensors. Correspondingly, lower activation of knee flexors and higher activation of knee extensors occurred in P2. In other words, poor temporal coupling between the momentum transfer phase and the activity of knee flexors and knee extensors may cause failed or impaired STS motion [49]. Therefore, during the phase of momentum transfer, stroke survivors tended to bend their trunk and knees adequately to provide the momentum to raise body up independently and avoid falling. In addition, the abnormally increased activation of SOL in P2 may be the result of weak postural control of body extension and stabilization [50], [51]. Previous studies also reported premature and excessive SOL activation on the paretic leg [12], [13]. In summary, impairment of muscle synergies occurred on both sides.

C. Impact of Exoskeleton Training
After three weeks of exoskeleton-based therapy, positive changes of muscle synergy patterns were demonstrated on both sides. After training, we found a fractionation of modules on the non-paretic side, and the new modules were similar to healthy subjects. Our results indicated that with exoskeleton training of balance and lower-limb function, a more normal neuromuscular complexity was restored, so that participants not only have better ability to complete the STS motion, but also to complete it using better synergy patterns and more flexibly. A previous study also had similar results, Cheung et al. proposed that the merging and fractionation of muscle synergies may be a mechanism to adjust to changing limb biomechanics by investigating the synergies for running in preschoolers [19]. Also, one study used electromyographic biofeedback to change muscle coordination patterns and successfully reduced knee contact forces [21]. These studies highlighted the redundancy and plasticity of muscle synergies.
From figure 6, it can be seen that there is a relatively large variability in the temporal structure on the paretic side among different patients due to differences in the severity of the disease. However, we can still find common changes compared to healthy subjects, namely, abnormal time coupling of knee extensors and flexors. On the paretic leg, the abnormal activation in the middle part of phase 3 of P1 declined after the intervention, as well as a decrease in activation of the knee flexors and an increase in activation of the knee extensors in P1. The changes are consistent with the kinetic changes, that is, after exoskeleton training, stroke survivors do not need to bend their trunk excessively before standing up.
The changes on both sides showed the effects of the exoskeleton intervention on muscle synergies. Our study suggests that the performance of the knee flexor-extensor muscles on the paretic side is an principal factor in ensuring the independence of stroke patients in STS motion, which is consistent with previous literature [52], [53]. The exoskeleton effectively compensates for insufficient knee extension during the momentum transfer phase by capturing motor information and perceiving the intention to stand up, thereby providing continuous knee extension assistance during the extension and stabilization phases. This eliminates the need for stroke survivors to excessively bend their trunk and knees. After treatment, we observed that the paretic side displayed a restoration of the normal activation of knee extensors during the momentum transfer phase. On the other hand, the nonparetic side regained its coordination without sacrificing flexibility or redundancy while completing the STS task. The direct assistance provided by the exoskeleton to the knee extensor muscles on the paretic side is an important advantage of the exoskeleton compared to conventional physical therapy, and it can also provide more repetitive task-specific training. Through these methods, the exoskeleton improves the clinical performance of stroke survivors and has great potential for application.
Finally, we investigated the correlations between neuromuscular complexity of muscle synergies and clinical measures. Because the number of modules required Pre and Post provides a coarse measure [22], we used the overall VAF of each module number Pre and Post to account for the change of synergy complexity [22], [41], [54]. As a result, we found a significant correlation on the non-paretic side compared with the change in BBS and FMA-LE. This is very important in terms of clinical implications, which shows the mechanism of intervention-induced improvement is not only the restitution of more normal synergy patterns on the paretic leg but also the use of adaptation or reorganized patterns emanating from the non-paretic leg. Possibly due to the insufficient training intensity or fewer subjects, we did not find a significant relationship on the paretic side.

D. Limitations
There were some limitations in our study. First, we did not measure muscles in the trunk due to the limited EMG channel numbers of our device. Thus, the muscle synergy pattern in phase 1 was lacking. Second, the body of the exoskeleton would cover where part of the electrodes should be attached, so we did not measure muscle synergies wearing the exoskeleton. Finally, due to time constraints, we did not set a control group trained by traditional physical therapy. We cannot confirm or parse out to what degree the long-term effects were produced by the exoskeleton, but these data provide important results to inform future randomized controlled trials to further investigate the effects of exoskeletons on muscle synergies.

V. CONCLUSION
In our study, we investigated changes in muscle synergies after stroke both on the paretic and non-paretic sides, the effect of an exoskeleton intervention on muscle synergies (i.e. plasticity of motor synergies), and the correlations between muscle synergies and clinical outcomes (i.e. behavioral correlates of muscle synergies). We found that synergy patterns of the paretic leg were damaged after stroke, and the nonparetic leg accommodated for the stroke-induced abnormalities in biomechanics by merging the muscle synergies. After three weeks of exoskeleton-based training, synergies of the nonparetic side were fractionated into three modules similar to healthy controls; and on the paretic side, muscle synergies partially shifted towards the normal pattern observed in healthy subjects. The increase in motor performance poststroke could be attributed to the recovery of neuromuscular coordination. These results demonstrated the positive effects of the exoskeleton-based training on lower-limb function in subacute stroke survivors through modifications of muscle synergies.