Effects of Boundary-Based Assist-as-Needed Force Field on Lower Limb Muscle Synergies During Standing Posture Training

The boundary-based assist-as-needed (BAAN) force field is widely used in robotic rehabilitation and has shown promising results in improving trunk control and postural stability. However, the fundamental understanding of how the BAAN force field affects the neuromuscular control remains unclear. In this study, we investigate how the BAAN force field impacts muscle synergy in the lower limbs during standing posture training. We integrated virtual reality (VR) into a cable-driven Robotic Upright Stand Trainer (RobUST) to define a complex standing task that requires both reactive and voluntary dynamic postural control. Ten healthy subjects were randomly assigned to two groups. Each subject performed 100 trials of the standing task with or without assistance from the BAAN force field provided by RobUST. The BAAN force field significantly improved balance control and motor task performance. Our results also indicate that the BAAN force field reduced the total number of lower limb muscle synergies while concurrently increasing the synergy density (i.e., number of muscles recruited in each synergy) during both reactive and voluntary dynamic posture training. This pilot study provides fundamental insights into understanding the neuromuscular basis of the BAAN robotic rehabilitation strategy and its potential for clinical applications. In addition, we expanded the repertoire of training with RobUST that integrates both perturbation training and goal-oriented functional motor training within a single task. This approach can be extended to other rehabilitation robots and training approaches with them.


I. INTRODUCTION
S TANDING postural control is essential for performing everyday functional tasks. Reactive postural control involves feedback-driven adjustments to maintain/recover postural stability in response to unexpected external perturbations. Dynamic postural control requires anticipatory postural adjustments to maintain upright balance during voluntary movements [1]. However, these abilities may be impaired in individuals with neuromotor disorders such as with spinal cord injury (SCI) and cerebral palsy (CP), which severely challenge the subjects to safely carry out activities of daily living [2].
Robot-assisted rehabilitation is gaining popularity in standing postural control recovery due to its important features, e.g., repeatable training environments, adaptable supports, increased therapy intensity, and reduced physical burden on therapists [3]. The assist-as-needed (AAN) strategy is widely used in rehabilitation robots to maximize motor recovery in patients by providing adaptive assistance based on measurements of their functional ability [4]. A prevailing paradigm of the AAN strategy is the trajectory-based assist-as-needed (TAAN) controller, which generates assistive force fields to guide subjects to follow predetermined trajectories for motor tasks. The force field is constructed based on the error between the current and desired joint or task trajectories [5]. TAAN force fields have been used in the training of various human movements and were reported to be effective in improving limb function, motor control, and task performance [6], [7]. However, previous studies have also indicated some challenges with the TAAN strategy in transferring the training effects observed in healthy subjects to the patient population [3].
The boundary-based assist-as-needed (BAAN) strategy is another type of AAN strategy. BAAN controllers provide assistance to subjects when specific points on the human body, e.g., the center of mass (COM), move out of the posture limits. The assistive force field is developed based on the posture stability boundary, commonly represented by planar geometric shapes (e.g., rectangular [8], circular [9], star-shaped polygon [10], etc.). We have developed a cable-driven Robotic Upright Stand Trainer (RobUST) and a training study that uses a circular BAAN controller. During training, subjects practice multi-directional postural tasks within and beyond their stability boundary. Primary outcomes of the RobUST-intervention are the expansion of standing workspace and improvements of standing postural control in healthy subjects [11]. In subsequent studies, we have found that the training effects due to the BAAN force field are well maintained in SCI patients even after the training has stopped [12], [13]. This is supported by the literature with other rehabilitation strategies which also apply the BAAN force fields [10], [14], [15].
To expand potential clinical applications and impact on patients, effects of AAN force fields on the neuromuscular control system should be investigated [3], [16]. Muscle synergies are proposed to be the lowest level in the neuromuscular control hierarchy. The central nervous system (CNS) recruits muscle synergies to translate complex task-level goals into basic execution modules [17]. Several previous studies indicate that muscle synergy analysis can explain the effectiveness of rehabilitation in different patient populations [18]. Following this idea, Cancrini et al. [16] and Lencioni et al. [19] investigated muscle synergies during reaching task training while assisted by planar robots with TAAN controllers. They found that the effects of the TAAN force fields on upper limb muscle synergies are limited in healthy subjects and in post-stroke patients. Escalona et al. used a wearable robotic exoskeleton with different TAAN control modes to assist healthy subjects in overground walking [20]. They proposed that the TAAN force field does not affect the number of lower limb muscle synergies but significantly changes the muscle weightings in each synergy. However, their follow-up study found that such muscle weighting changes induced by TAAN force fields in healthy subjects are not replicated in SCI patients [21]. This might partially explain the gap in the training effect between healthy subjects and SCI patients using the TAAN strategy.
Although promising results have been achieved, the muscle synergy adjustments induced by AAN force fields, especially BAAN force fields, during training remain unresolved and can be further investigated. To the best of our knowledge, this paper is the first to investigate the effects of the BAAN force field on lower limb muscle synergies during standing posture training. We integrated virtual reality (VR) technology into the RobUST platform to develop a complex standing task requiring both reactive and dynamic postural control. Ten healthy subjects were randomly assigned to two groups. Each subject completed 100 trials of the standing task, with or without BAAN assistance provided by RobUST. Our results show that the BAAN force field significantly improves reaching and balance control, which agrees with our previously observed results [11]. We have also found that the BAAN force field reduces the number of lower limb muscle synergies while increasing the synergy density, i.e., the number of muscles recruited in each synergy during both reactive and voluntary dynamic posture training. This pilot study helps better understand how muscle coordination is impacted by the BAAN force field and its benefits in potential clinical applications. In addition, we introduced VR to expand robotic training space and integrated perturbation training and goal-oriented functional motor training in a single task. Our methodology might also work in other robotic rehabilitation approaches.

B. Experiment Preparation
RobUST is a cable-driven robotic platform that can apply forces on the trunk and pelvis. As shown in Fig. 1a, eight cables are attached to two belts at the level of the trunk and pelvis. Eight motors (Maxon Motor, Switzerland) instrumented with load-cells (LSB302 Futek, California) are mounted on a stationary frame to control the cable tensions. A motion capture system (Vicon Vero 2.2, Denver) provides real-time information on the position and orientation of the two belts to the robotic controller. Device setup details of the RobUST are described in our previous work [11].
Before experiment, a subject was blindfolded, wore two belts, and stood within the RobUST system. Cables were removed from the pelvic belt and only the trunk cables were kept. The subject was asked to maintain balance when receiving a sudden force perturbation at the trunk. A single force perturbation pulse lasted for 0.1 sec. The perturbation direction was randomly chosen from four directions (anterior, posterior, left, and right). For each direction, the perturbation force started from 40% body weight (BW) and increased by 1% in each attempt until the subject lost balance (i.e., lifted one leg off the ground). The highest perturbation force in a direction was recorded as the maximum perturbation force threshold in that direction. Besides, the moving trajectory of the pelvic belt was recorded for each subject.

C. Experiment Protocol
We integrated VR technology into the RobUST to develop a novel standing task. Fig. 2 is the schematic of the standing task. A subject stood in the RobUST and wore a VR headset (HTC VIVE pro). In the virtual environment, a target with ten circular zones was set in front of the subject at 5 m. Its bullseye was set at the sternum height of the subject. The target reciprocated horizontally at a constant speed of 30 cm/s within the range of 3 m. A ball was set above the target (Fig. 2a). Once triggered, the ball flew towards the subject at the speed of 3 m/s (Fig. 2b). As the subject was visually tracking the flying ball, at a random timepoint between 0 ∼ 0.8 s after the ball was released, RobUST delivered a trunk perturbative force at the maximum threshold in one of the four directions (Fig. 1b). The subject was asked to maintain balance without stepping when receiving the perturbation (Fig. 2c). The subject used the VR controller in his dominant hand to catch the ball (Fig. 2d), aim at the bullseye of the moving target (Fig. 2e), and then throw the ball (Fig. 2f).
A warm-up session was set before the experiment. Subjects performed the standing task several times until they became familiar with the whole procedure. During the experiment, each subject performed 100 trials of the standing task. A 5-min break was set after the 50th trial.
Ten subjects were randomly assigned to two equal groups: FF group and no-FF group. The only difference between two groups was that during the experiment, subjects in the FF group received pelvic assistance from the BAAN force field provided by RobUST, whereas subjects in the no-FF group did not. Fig. 1b shows the schematic of the BAAN force field. For each subject in the FF group, the postural stability boundary was the smallest circle covering all pelvic moving trajectories recorded in the experiment preparation session. During the experiment, when the pelvic center moved out of the virtual boundary, RobUST generated a planar assistive force directed towards the center of the circle, which denotes the neutral position. Calculation details of the circular BAAN controller are described in our previous work [11].
Task performance data were collected by the VR system. Specifically, the number of successful catch and throw trials were recorded for each subject. Aiming accuracy score was also recorded. The bullseye was 10 points and the score in each outer ring decreased linearly from 10 to 0.

E. Data Preprocessing
Custom MATLAB routines (MATHWORKS, Natick, MA) were used for data preprocessing. The raw force plate data were low-pass filtered using a fourth order Butterworth filter with a cutoff frequency of 6 Hz. EMG data was first band-pass filtered (20-300Hz), demeaned, rectified, and low-pass filtered at 50Hz [22].
Force plates and EMG sensors were recorded within the RobUST console. In each trial, when the perturbation force was activated (Fig. 2c), the RobUST console recorded the perturbation onset timestamps of the force plates and the EMG sensors to synchronize the collected data. Fig. 3 shows examples of filtered, synchronized EMG and force plate data.

F. Time Bin Selection
Our standing task includes reactive control period (Fig. 2c) and voluntary dynamic control period (Fig. 2d, e, f). We selected several time bins to investigate temporal variations in muscle activity during these periods.
Background (BK) bins were set as 75 ms before the perturbation onset to represent the resting activity of the muscles. Automatic postural response (APR) has been well characterized in the literature and occurs about 100 ms after the perturbation [23]. We divided it into three time bins. Each APR bin lasts for 75 ms and begins at 100 ms (APR1), 175 ms Examples of the filtered, synchronized EMG and force plate data with their APR and VPR windows (a. a subject from the FF group b. a subject from the no-FF group). The top three rows (TA, RF, and BF) are bilaterally averaged EMG signals from three exemplar muscles in one trial. The fourth row (ALL) is the scaled, integral EMG activity across all muscles. The bottom row (Fz) is the GRF along the vertical axis. Red vertical line is the perturbation onset and is set as time zero. BK, APR windows are in grey and VPR windows are in blue.
Due to the practice and fatigue effects, the duration of the voluntary postural response (VPR) period changes across trials during training [17], [24], [25]. To address this issue, we developed a dynamic procedure to select VPR time bins. The workflow is described below: 1) Extract EMG signals from each trial. Next, for each type of muscle, average EMG signals bilaterally and then normalize the data by the maximum to guarantee that all muscles have the same relative magnitude (e.g., Fig. 3 top three rows). Then, sum all muscle signals together to get the integral EMG activity signal which represents the overall muscle activation and its variation in one trial (e.g., Fig. 3 row ALL).
2) For the integral EMG signal, extract data from 325 ms (end of the APR3) to the end of the trial. First, find all peaks except which are too close to each other (i.e., peak-to-peak distance less than 225 ms). Next, calculate the ratio of the top two maximum peak values. If the ratio is less than 0.9 (i.e., one dominant peak situation, which mostly occurs in the FF group, see Fig. 3a), record the timestamp of that dominant EMG peak. If over 0.9 (i.e., multiple peaks situation, which mostly occurs in the no-FF group, see Fig. 3b), compare the corresponding peaks of the synchronized vertical GRF signals (Fig. 3 row Fz). The EMG peak with a higher Fz is set as the dominant EMG peak, and its timestamp is recorded.
3) Use the timestamp of the dominant EMG peak as the middle point to create a 75-ms bin (VPR2), which represents the maximum activation period in VPR. Use 5% of the dominant peak value as the threshold to detect the start and end points. Use these two points as middle points to create two 75-ms bins (VPR1 and VPR3), which represent the start and end periods of VPR. Fig. 3 shows two examples of the VPR bins (blue area) using the dynamic time bin selection method.

G. EMG Matrix Organization
We collected EMGs from 14 lower limb muscles for each subject. For each muscle, mean muscle activity during each time bin was calculated in each trial. These mean values were assembled to form two EMG data matrices for each subject: We horizontally concatenated the two matrices and calculated the maximum value for each row. APR and VPR matrices for each subject were then normalized by the corresponding maximum value in each muscle row [26]. Previous study suggested that EMG should be normalized by maximal voluntary isometric contraction (MVIC) for muscle activation comparison between periods and tasks [27]. However, research have also reported that the maximum value normalization method performs similar to the MVIC method in muscle synergy analysis [28], [29].
We horizontally concatenated the EMG matrix of each subject to construct the group EMG matrix. Specifically, we combined APR matrices (or VPR matrices) from 5 subjects in the FF group to create the FF-APR group EMG matrix (or FF-VPR group EMG matrix). Likewise, we combined APR matrices (or VPR matrices) from 5 subjects in the no-FF group to create the noFF-APR group EMG matrix (or noFF-VPR group EMG matrix).

H. Muscle Synergy Extraction
The non-negative matrix factorization (NNMF) method [30] was used to extract muscle synergies from four group EMG matrices. Based on the NNMF method, the group EMG matrix (M) is composed of a linear combination of several muscle synergy vectors (W i ), which are recruited by the synergy activation coefficients (c i ). Therefore, we have: Muscle synergy vector W i provides information about which muscles are involved in the ith synergy and their relative contributions (range 0∼1). Muscle synergies are time-invariant and stable across perturbation directions. Synergy activation coefficient c i reflects the activation level changes of the synergy W i over time bins and perturbation directions [17].

I. Muscle Synergy Number
Muscle synergy number is manually determined to adequately reconstruct the group EMG matrices. Specifically, in Equation (1), we can increase the muscle synergy number n to reduce the error between the left and the right side, i.e., the original and reconstructed matrix. The reconstruction level is quantified by the Variability Accounted For (VAF) [31]: We increased the muscle synergy number until it could account for at least 90% for the overall VAF and 75% for each muscle VAF [26]. The VAF vs. synergy number curves were plotted for each group EMG matrix.

J. Statistical Analysis
SPSS (IBM, version 27, 2020) was used for statistical analysis. The alpha rate was set at 0.05. Normal distribution of the data was examined with Shapiro-Wilk test and visually inspected with Q-Q plots. Motor behavior data (e.g., successful reaches and aiming accuracy scores) were not normally distributed and thus a non-parametric Mann-Whitney U was applied. COP-related kinematics between groups were examined with independent t-tests.
For the spatial-temporal muscle activation feature analysis, we first extracted all APR bins and VPR bins from the preprocessed and bilaterally averaged EMG signals. Next, data in each bin were averaged, and then normalized across trials, subjects, and groups. A three-way mixed Analysis of Variance (ANOVA) with two between-factors (group: FF and no-FF; and four perturbation directions) and one repeated measuresfactor (bin: 1, 2 & 3) was applied to examine significant spatial-temporal differences during APR and VPR. In case of a significant ANOVA model, we proceeded with post-hoc testing with Bonferroni's inequality procedure for multiple comparisons. We studied homoscedasticity and multicollinearity of the data with Levene's and Mauchly's sphericity tests, respectively. In case the assumptions of homogeneity and sphericity were violated, adjusted p-values to variance-corrected and Greenhouse-Geisser correction were reported.

B. Balance Control Performance
The FF group showed greater level of balance control during the experiment. Table I represents averaged group COP-related variables. The BAAN force field improved standing balance Motor performance during the standing task. Y-axis is the successful catching/throwing rate, which is the number of successful catching/throwing trials divided by the total number of trials. The application of the BAAN force field improved the success of the motor behaviors. ⋆ = p < 0.05.    Tables II & III, and were visually depicted in Fig. 5A.  We found that muscle activation level in the FF group was less in TA but greater in RF than the no-FF group (Table II).
In the FF group, muscle activation level remained stable across three APR time bins. However in the no-FF group, most of the muscles (LG, BF, ES, GM) showed highest muscle activation level in APR3 (Table III).
2) Temporal-Spatial Features of VPR: The statistical analysis results were summarized in Table IV, and were visually depicted in Fig. 5B.
The no-FF group showed greater muscle activation level than the FF group in TA, LG and GM during VPR2 (Table IV).
Since we applied dynamic time bin selection method for VPR bins, muscle activations in VPR1 and VPR3 were significantly lower than that in VPR2.

D. Muscle Synergy Number
For the FF group, four synergies reached over 90% overall VAF and could account for more than 75% VAF for each muscle (Fig. 6 a, c). However, the no-FF group needed at least eight synergies to meet the 90% overall VAF threshold and the 75% muscle VAF threshold (Fig. 6 b, d).
We found that the synergy number is stable between APR and VPR (Fig. 6 a vs. c, and b vs. d).

E. Muscle Synergy Analysis
Muscle syergies extracted from four group EMG matrices and their corresponding activation coefficients are visually depicted in Fig. 7.
1) Synergy Density Analysis: Elements of the muscle synergy vector are referred to as muscle weightings, which reflect the relative contribution of each muscle in one synergy [20].  Muscle weightings in each synergy are normalized by their maximum (range 0∼1). Muscle with a weighting above 0.4 is considered the main contributor to that synergy and is defined as the "predominant muscle" [32], [33]. Number of the predominant muscles in each synergy is a measure of the muscle synergy density. Table V shows the predominant muscle number in muscle synergies extracted from the four group EMG matrices. During both APR and VPR, the FF group recruited less but more dense muscle synergies than the no-FF group.
2) Synergy Activation Analysis: Muscle synergy activation level fluctuated in different directions. Synergy activation coefficients remained stable across time bins during APR (Fig. 7a, b). Muscle syergies were more activated during VPR2 than VPR1 and VPR3 (Fig. 7 c, d).

A. FF vs. No-FF
Our results show that the BAAN force field significantly improves motor performance and balance control, consistent with our previous findings [11]. Furthermore, we found that the BAAN force field reduces the total number of lower limb muscle synergies while simultaneously increasing the synergy density, i.e., predominant muscle number in each synergy when performing the task. This may contribute to performing the task with a different recruitment of muscles providing an avenue for motor training.
Muscle synergy number has been reported to be positively correlated with task complexity [34], [35], and this number remains stable across similar tasks [36]. However, complex actions may require additional "task-specific" synergies to handle the complexity of movements, resulting in an increase in synergy number [32], [35], [37]. Research has also shown that even using the same task, synergy number could change since task execution difficulty often fluctuates across training trials [25], [38], [39]. In this study, the FF group performed the complex standing task the same number of times as the no-FF group but showed only half the number of synergies. This implies that the BAAN force field reduces the difficulty of task execution. In other words, performing the same task was easier for the FF group and this is well supported by the higher task scores (Fig. 4).
Increased synergy density (Table V) may account for the reduction in task execution difficulty with BAAN force field.
Previous studies have proposed that the increased predominant muscle number in each synergy is a neuromuscular regulatory mechanism to raise the joint stiffness [40]. The increased stiffness resulted in transient joint locking and thereby could enhance joint stability during unexpected external perturbations [41]. Although this strategy may reduce posture flexibility [42], the neuromuscular system adopts this strategy to reduce the complexity of CNS control to gain stability and enhance limb support during complex tasks [32]. Our observations are consistent with the literature. During the standing task, we observed that the FF group tended to stabilize the knee joints moderately, reduce the amplitude of motion in the lower limbs (Table I), and complete the task with fewer movements (single EMG peak during VPR, shown in Fig. 3). The no-FF group, in contrast, had to segment the serial motor task into distinct phases, modulated posture sequentially, and completed the task with more movements.
Another possible argument can be made by taking EMG signals as data points distributed in a high-dimensional space. Synergies extracted by the NMF algorithm represent boundary vectors to describe a subspace capable of covering most data points [43]. In this study, the no-FF group showed more EMG signal peaks (Fig. 3) with higher peak values during VPR (Table IV) and more variable predominant muscles than the FF group. This implies that compared with the FF group, EMG data points of the no-FF group are distributed around more centers with large variances in a higher-dimensional space (Fig. 6, VAF curves of the no-FF group rise more slowly than the FF group). Therefore, the no-FF group recruited more muscle synergy vectors than the FF group to expand the descriptive space.
Prior studies have proposed that subjects with neuromotor disorders (e.g., CP, SCI, stroke) recruit significantly fewer muscle synergies than healthy subjects during task execution [44]. Due to neuromuscular control deficits and muscle weakness, such a reduction in synergy number is hard to change during training [45], [46]. Although the TAAN force field could alter muscle weightings in each synergy, the synergy number remains stable during training [20]. Compared with healthy subjects, fewer muscle synergies limit patients' posture flexibility and adaptability during task execution. Therefore, they are unable to alter the muscle synergy weightings during TAAN-based robotic rehabilitation. This might explain the challenge of the TAAN-based robotic rehabilitation in transferring training effect from healthy subjects to the patient population [21]. Unlike the TAAN force field, we find that the BAAN force field not only can change muscle weightings but also reduce synergy numbers. This finding implies that the BAAN-based intervention might be able to train patients to complete tasks with less but more dense muscle synergies. This strategy might lower the task execution difficulty and neuromuscular control complexity for them. To test the feasibility of the strategy, future studies will investigate the effect of the BAAN force field on muscle synergies in individuals with motor impairments. Our findings in this pilot study should serve as the stepping stone to investigate future applications of the BAAN force field in robotic rehabilitation.

B. APR vs. VPR
Previous studies have shown that differences between APR and VPR should be investigated to gain a better understanding of motor control [47]. Isa et al. [48] and Drew et al. [49] proposed that hierarchical neural pathways act parallelly to recruit a common set of muscle synergies to realize multiple motor behaviors. Chvatal et al. found that muscle syergies are robust during APR and VPR stages in a perturbed walking task [50]. Their subsequent study also shows that synergies recruited in standing perturbation responses and walking are similar [22]. Our findings are consistent with the literature. In this study, we found that muscle synergies are robust during APR and VPR stages of the standing task. Our results imply that different postural behaviors during the complex standing task (Fig. 2) were generated by using a shared pool of muscle synergies. Postural adjustments were realized by modulating the activation time and amplitude of the muscle synergies. However, due to muscle decline and neuromuscular deficits, patients might recruit different synergies when receiving unexpected perturbations and executing voluntary movements during training. Therefore, future studies are warranted to investigate muscle synergy differences between APR and VPR in patient populations to further and expand postural control training in standing.
APR has been proposed to occur about 100 ms after the perturbation onset [23]. However, a recent study reported that after multiple training trials, APR could transition into the VPR phase due to the learning effect secondary to motor practice. The consequence would be muscle activation that occurs earlier than the perturbation onset because anticipatory mechanisms would have taken over [51]. Another study, however, has shown that dual-tasking might reduce or remove this learning effect for better comparison between APR and VPR [37]. In this vein, we designed a postural standing task that comprised two sub-tasks (Fig. 2) by integrating VR and RobUST. For the reactive control sub-task (Fig. 2c), perturbation direction and onset time were randomized to minimize the anticipation effect. To activate APR, a flying ball was set to attract the attention of subjects during perturbation. Such setups guarantee that even after 100 training trials, subjects still activate their muscles after the perturbation (Fig. 3). For the dynamic control sub-task (Fig. 2 d, e, f), it has been reported that the duration of the VPR period change across training trials [17], [24], [25]. To address this, we developed a dynamic time bin selection algorithm to extract the overall trend in EMG variation during VPR. Our approach can be extended to other studies in robotic rehabilitation when requiring comparison between APR and VPR in a single task.

V. CONCLUSION
In this study, we investigate how the BAAN force field impacts lower limb muscle synergies during standing posture training. We conducted a control experiment and demonstrated that the BAAN force field significantly improved balance control and motor task performance. Our results also indicated that the BAAN force field reduced the total number of lower limb muscle synergies while increasing the synergy density during both reactive and voluntary dynamic posture training. This pilot study helped uncover the neuromuscular basis of the BAAN force field, which might promote its clinical applications. Besides, we effectively incorporated perturbation training and goal-oriented functional motor training into a single task by utilizing the VR technique. This approach can be extended to other rehab robotic platforms.