Impact of Targeted Assistance of Multiarticular Finger Musculotendons on the Coordination of Finger Muscles during Isometric Force Production

Abstract

Neurological injuries often cause degraded motor control. While rehabilitation efforts typically focus on movement kinematics, abnormal muscle activation patterns are often the primary source of impairment. Muscle-based therapies are likely more effective than joint-based therapy. In this study, we examined the feasibility of biomimetic input mimicking the action of human musculotendons in altering hand muscle coordination. Twelve healthy subjects produced a submaximal isometric dorsal fingertip force while a custom actuator provided assistance mirroring the actions of either the extrinsic extensor or the intrinsic muscles of the index finger. The biomimetic inputs reduced the activation level of all task-related muscles, but the degree of change was different across the muscles, resulting in significant changes in their coordination (co-contraction ratios) and force-EMG correlations. Each biomimetic assistance particularly increased the neural coupling between its targeted muscle and the antagonist muscle. Subjects appeared to fully take advantage of the assistance, as they provided minimal level of effort to achieve the task goal. The targeted biomimetic assistance may be used to retrain activation patterns post-stroke by effectively modulating connectivity between the muscles in the functional context, and could be beneficial to restore hand function and reduce disability.

I. Introduction

Recovery of hand function following neurological injuries such as stroke remains limited even after extensive rehabilitation efforts [1] – [3]. The resulting deficits have a profound impact on functional recovery of the upper limb [4] and overall quality of life due to the functional importance of the hand [5]. The underlying mechanisms of impairment are primarily neurological, namely aberrant activation patterns. Stroke survivors often have difficulty fully activating a muscle [6], [7] and thus experience profound weakness [8]. They may have motoneuronal hyperexcitability, manifesting as spasticity [9], excessive coactivation, or difficulty with termination of muscle activation once initiated [10]. Additionally, they have difficulties with creating different activation patterns and thus modulating pattern with task [11], [12]. The level of deficit may vary greatly across muscles. For example, we have observed much greater relative deficits in extrinsic finger extensor muscles than in extrinsic finger flexors [7], [13].

Instead of targeting underlying cause of functional impairment (i.e., abnormality in muscle activation), however, much of stroke rehabilitation focuses on joint movement rather than muscle activation. Accordingly, a substantial number of assistive robotic devices for hand have been developed to provide joint-based training, with an emphasis on kinematics [14]. These devices can provide assistance to complete the movement by actuating the finger joints, such as for hand opening [15], [16] and power grip [17] – [19]. In the hands, however, many muscles may contribute to the motion of one joint in a complex manner [20], and the impact of the movement-based therapy on a given muscle may be difficult to predict. Note that, in human motor learning that occurs during movements under external force, dynamics of external inputs/forces, not kinematics of movements, are found critical during motor adaptation [21].

Additionally, the sensory feedback created by robotic manipulation of joints through torque production differs from the natural experience. Musculotendon forces create not only joint torque but also joint compression that increases joint stability and generates proprioceptive sensibility (i.e., awareness of movement: [22] – [24]). During this process, different types of somatosensory and/or proprioceptive feedback are provided via neural pathways involving sensory receptors measuring signals of different modalities, including Pacinian (movement) and Ruffini (position/movement) corpuscles within joint capsules (albeit to a limited degree; [25]). Many robotic systems, however, produce joint torque which create a very different sensory experience. Further, dynamic processes such as intersegmental force transmission via passive joint impedance, found as critical aspect of hand movement production by multiarticular musculotendons [26], cannot be replicated by current robotic systems.

To overcome these limitations, we have been exploring muscle activation-based therapy. Rather than controlling joint motion, we apply forces in a manner to simulate the action of a specific muscle. In this fashion, we can match the required assistance to the muscle. As stated, certain muscles, such as extrinsic extensors, may require more assistance than other muscles, such as extrinsic flexors. By providing muscle-specific therapy, we hope to improve muscle activation patterns and thus reduce impairment. Toward this end we have recently developed a biomimetic hand exotendon device [27], which is capable of replicating kinetic outcomes of multi-articular hand musculotendons (i.e., coordinated multi-joint moment patterns) by applying forces to exotendons that mimic the anatomy of those musculotendons.

The effects of such ‘biomimetic’ kinetic inputs on human motor control are of great potential importance in neurorehabilitation, but are yet unexamined. Note that different scenarios of motor adaptation under biomimetic assistance could emerge, depending on the capacity of the human central nervous system (CNS) to recognize biomimetic inputs, and/or to act upon such stimuli. For instance, if the CNS does not identify distinct patterns of somatosensory feedback created by the exotendon force (and perceive it as a change in the task goal, i.e., lower target force), biomimetic assistance may simply result in the reduction in the level of effort, i.e., decrease in activation level of all muscles (while maintaining a similar coordination pattern). Conversely, if the CNS does recognize the pattern of biomimetic somatosensory input (from sensory receptors in the joints and muscles) assisting a specific muscle, it may attempt to reduce the activation of the assisted muscle, while maintaining the activation of the other muscles at the similar levels. However, modulation of a single muscle activity (in isolation) may not be feasible, as it would be constrained by different neurophysiological characteristics of underlying neural control structure, such as the degree of cortical projection [28], [29], neural crosstalk/coupling with other muscles [30], and its system-level control structure such as modular control [31].

We aimed to elucidate the effects of targeted assistance of multiarticular musculotendons of the hand on motor control of a specified task. Here, subjects produced an isometric fingertip force in the dorsal direction. The use of an isometric force production task was selected to exclude potential effects of various movement-related factors (such as variability in movement speed and change in muscle length) on muscle coordination. Activation of three different task-related muscles (agonist/synergist and antagonist) were recorded during the trials: extensor digitorum communis (EDC), first dorsal interosseous (FDI), and flexor digitorum superficialis (FDS). Subjects received targeted assistance of the extrinsic extensors, targeted assistance of the interossei, or no assistance. We expected that the targeted assistance of a multiarticular hand musculotendon would influence the control of targeted muscle, while activation of the other muscles remained less affected.

II. Methods

A. Subject Characteristics

Twelve healthy, right-handed subjects (S₁ – S₁₂; 5 males; mean ± SD age = 24 ± 2 yrs) with no known history of neurological or orthopedic conditions participated in the study. The experimental protocol was approved by both the institutional review boards of MedStar Health and the Catholic University of America; written informed consent was obtained from each subject prior to participation.

B. Instrumentation

Subjects wore an exotendon device, a modified version of our recently developed biomimetic device (BiomHED; Lee et al., 2014) that provided assistance during the task (Fig. 1a). Briefly, for the index finger, four cables are routed through custom thermoplastic components attached to the dorsal and palmar aspects of the finger via Velcro straps. One cable (Finger ExoTendon 1; FET₁) is routed to replicate the anatomical configuration of the extensor digitorum communis (EDC) and extensor indicis (EI) tendons, another cable (FET₂) follows the flexor digitorum profundus, and the third and fourth cables (FET₃/FET₄) follow tendons for the first dorsal (radial) and palmar (ulnar) interossei (FDI/FPI). Tension applied to each exotendon has been shown to reproduce kinetic effects of the replicated multi-articular tendon (Lee et al., 2014). In the modified version of the device used in this experiment, we eliminated the fabric glove used to route the cables, as such fabric gloves were found very difficult – especially for patients with clenched hands – to don and/or doff; instead, we used customized 3D-printed plastic components to route the cables, each of which can be easily worn using Velcro straps (see Fig. 1a). Exotendons of the modified device were found, in our pilot experiment with a small number of subjects (n = 3), to generate movement patterns (i.e., multi-joint coordination) similar to our previous glove-based device. During experiments for this study, we provided targeted assistance to the EDC/EI muscles and/or FDI/FPI muscles by actuating the motors pulling FET₁ and/or FET₃/FET₄ (actuated together) (see Experimental protocol)

Fig. 1.

Biomimetic device providing targeted assistance of multiarticular musculotendons of the finger: (a) Exotendon routing; (b) Experimental setup. Here, the distal phalanx component of the device that makes contact with the device was extended distally, reducing the degree of tactile sensory to the finger.

Three pairs of disposable, self-adhesive silver/silver chloride surface electrodes (diameter 15 mm, center spacing 20 mm; Noraxon, Scottsdale, AZ, USA) were used for surface EMG recordings. One electrode was placed on the hand to record the activity of an intrinsic hand muscle, FDI, and two on the forearm to record extrinsic hand muscle activities from the 1^st compartments of the flexor digitorum superficialis (FDS) and EDC. To ensure accurate placement of each electrode, EMG signals from the electrodes were inspected while subjects performed several finger and wrist movements associated with the target muscle and adjacent muscles. The electrode location was adjusted if the EMG signal changed during isolated contraction of any neighboring muscle. However, it should also be acknowledged that it is still possible that the electrical activity of neighboring muscles (such as wrist flexor/extensor) may have influenced the recorded EMG signal to some degree. The EMG signals were sampled at 1000 Hz and processed with a 3^rd-order band-pass filter with 5Hz and 400Hz cutoff frequencies and a 3^rd-order notch filter of 60Hz.

C. Experimental Protocol

Subjects were seated in an adjustable chair with their elbow resting on a padded support. The right hand was placed on a table and immobilized by two Velcro straps with the wrist in a neutral posture (Fig. 1b). The position and orientation of a 6 degree-of-freedom load cell (Mini40; ATI Industrial Automation, Apex, NC) were adjusted such that each subject could maintain the following posture during force production: 30° flexion at the distal interphalangeal (DIP) joint, 45° flexion at the proximal interphalangeal (PIP) joint, and 30° flexion at the metacarpophalangeal joint (MCP). The contact surface was covered with sandpaper (320 grit) to prevent slipping.

In order to normalize EMG for subsequent analyses, subjects were first asked to create maximum activations. Subjects performed each of the following tasks for approximately two seconds to excite the targeted muscle: maximum finger extension (EDC), maximum index finger flexion with the DIP joint extended (FDS), and maximum index finger abduction (FDI). During maximum activations, subjects were guided (by experimenter) to maintain a finger posture similar to the aforementioned finger posture used during the force production. EMG data were recorded for four seconds. During the first second, subjects increased muscle activation to the maximum level (ramp-up phase), which they maintained for the next three seconds. Data recorded during the middle two seconds of the sustained maximum contraction were used to calculate the maximum EMG level. For this portion of EMG data, a moving window of 1 second was shifted in 100-millisecond steps to find the maximum EMG level for each channel.

During experiments, subjects were instructed to produce dorsally directed isometric force equal to F_tar (experimental task). Dorsal fingertip force production (or finger extension) was selected as a target task as it represents the critical impairment observed among stroke patients; recovery of finger/ thumb extension of patients is also highly correlated with their overall functional recovery [32][33].

In the Pre-experimental session (Fig. 2a), subjects were first instructed to produce maximum voluntary dorsal fingertip force (MVF) without assistance. Then, the target force level was set as 40% of the MVF. The level of assistance for each exotendon was then determined under assistance-only conditions. Assistance was produced by actuators (brushed DC motor with gearheads, A-max 16, GP 16A with reduction ratio 29:1, Maxon Motor AG, Switzerland), which created tension in the corresponding exotendon. During these trials, subjects were instructed not to voluntarily produce fingertip force in any direction, but simply to maintain the static finger posture while tension applied to the exotendons produced the fingertip force. From these assistance-only trials, the level of assistance for each exotendon was determined as the exotendon force that produced about 67% of the target fingertip force. Once the assistance levels for both exotendons were found, three trials of the ‘assistance-only’ condition for each exotendon were also recorded.

Fig. 2.

Experimental protocol, consisting of (a) pre-experiment session; and (b) three experimental sessions (A₀, A_EXT, and A_INT). The order of the three experimental sessions/conditions was randomized across subjects.

Then, in the Experimental sessions (Fig. 2b), three sets of trials were run. Each set consisted of one block (5 trials) of unassisted trials followed by two blocks (5 trials each) of one of three assistance conditions (no assist, extrinsic assist, and intrinsic assist). Here, the last two blocks, corresponding to Phase 1 and 2 of the trials, were administered to examine whether some type of motor learning (and/or adaptation) occurs under assistance over time.

In the first block (unassisted trials) of each set, the actuators of the device were turned off; subjects produced the target, dorsally-directed fingertip force for 4 seconds on their own. A customized graphical user interface provided subjects with information regarding the timing of the task performance (e.g., time to start and end the force production task) and the force magnitude. In the next two blocks (assistance condition), utilized one of the three assistance conditions. For no assistance, subjects continued to perform the task without assistance. For extrinsic assistance (A_EXT), the actuator pulled on the central exotendon on the dorsal side (FET₁) and the subject was instructed to “perform the task with the help of the device”. Again, the actuators created roughly 67% of the target dorsal fingertip force. For the intrinsic assistance (A_INT), the actuators pulled on the lateral exotendons (FET₃/FET₄) while the subject worked with the device to create the dorsal force.

Approximately 15 – 20 seconds of rest were given between the trials (within each condition), and a 10-minute break was provided between the conditions. The order of the three assistance conditions (A₀, A_EXT, A_INT) was randomized across subjects.

D. Data analysis

The following analysis techniques were implemented to process force and EMG signals.

1) Force data analysis

The force signal was low-pass filtered (3^rd-order Butterworth filter, cutoff frequency = 25 Hz) before being analyzed [34]. The following measures were computed from the fingertip force data:

1. Dorsal fingertip force: The magnitude of the dorsal fingertip force was recorded during the task performance.

2. Shear (distal/proximal) fingertip force: During dorsal fingertip force production, the magnitude of the shear fingertip force, produced in the distal/proximal direction (tangential to the contact surface), was also recorded. As information regarding the shear fingertip force was not provided to the subject during task performance, the shear fingertip force can be regarded as a motor output in the redundant (or ‘task-irrelevant’) dimension [35], whose variability data could provide information regarding how the central nervous system strategically controls multi-dimensional motor outputs (see below). For each subject, the average distal fingertip force during the ‘hold’ phase was normalized by the magnitude of the dorsal fingertip force to account for between-subject variability in the magnitude of the target fingertip force.

3. Motor variability: Coefficient of variation (CV; or normalized standard deviation) was computed for the dorsal and distal (shear) fingertip forces as a means of quantifying variability in the motor outputs (fingertip forces). CV measure has been commonly used to quantify fluctuation in the motor output [36], [37].

4. Task error: The task error was defined as the root-mean-squared difference between the target force and the dorsal fingertip force F_d produced by each subject during the ‘hold’ phase of the task performance (see Fig. 3a), normalized by the target force magnitude. Task error provided quantitative means to evaluate task performance.

Fig. 3.

Mean (SD) of the fingertip force vector under five experimental conditions: No assistance (A₀), EXT assistance only (AO_EXT), INT assistance only (AO_INT), EXT assistance (A_EXT), and INT assistance (A_INT). Ellipses at the fingertip force vectors indicate ‘between-subject’ variability of the force vectors.

2) EMG data analysis

We computed the following four measures from the EMG data, and examined whether these measures changed with type of assistance: a) mean activation level (MAL), b) co-contraction ratio (CCR), c) regression coefficient (RC), d) intermuscular coherence integral (CI). The first measure (MA) indicates how the activation of each muscle was affected by the assistance, while the next three measures (CCR, RC, ICI) quantify how the coordinated control between two muscles were affected. Specifically, intermuscular coherence quantifies the degree of common neural input across muscles [38], which also reveals their ‘functional connectivity’ (i.e., coordinated use to achieve a functional task; [39]). These three measures were computed between the three possible pairs of the muscles (EDC-FDS, EDC-FDI, and FDS-FDI).

1. Mean activation level (MAL): Mean activation levels of the three muscles (MAL_i; i = 1, 2, 3; 1: EDC, 2: FDS, 3: FDI) were estimated by averaging the normalized activation levels of these EMG signals during the hold phase (1 – 4 sec) of the fingertip force production.

2. Pairwise co-contraction ratio (CCR): Co-contraction ratio values of the three muscle pairs, EDC-FDS (CCR₁₂), EDC-FDI (CCR₁₃), and FDS-FDI (CCR₂₃) were estimated from their activation levels during the hold phase.

   $${\mathit{\text{CCR}}}_{\mathit{\text{ij}}}=\frac{{\mathit{\text{MAL}}}_i}{{\mathit{\text{MAL}}}_j}\times 100(\%)(i,j-1:\text{EDC},2:\text{FDS},3:\text{FDI})$$

3. Regression coefficient (RC): For each muscle pair, the regression coefficient, the slope of the regression line of the activation profiles of the two muscles, was also computed from the normalized activation profiles of the ramp phase (0 – 1 sec), which quantified the ‘dynamic’ rate of change in the activation level of one muscle as the activation of the other muscle changes/increases.

4. Intermuscular coherence integral (ICI): Entire EMG data of each block (4 trials) were first concatenated, then EMG-EMG coherence between muscle-pairs during the hold phase was estimated using non-overlapping segments (rectangular window) that resulted in a frequency resolution of 2 Hz within the MATLAB environment (MathWorks, Inc.; Natick, MA), employing a script developed by Neurospec (www.neurospec.org; [40]). Given two EMG signals x and y, let the power spectra of the two signals be denoted as f_xx(λ) and f_yy(λ), and their cross-spectrum as f_xy(λ). The coherence at frequency λ, R_xy(λ), is then computed as:

   $$|R_{\mathit{\text{xy}}}(\lambda )|=\frac{{|f_{\mathit{\text{xy}}}(\lambda )|}^2}{f_{\mathit{\text{xx}}}(\lambda )f_{\mathit{\text{yy}}}(\lambda )}$$

   The coherence value at a given frequency was considered only if it was greater than the 95% confidence limit. The coherence estimates were then z-transformed as follows:

   $$Z(\lambda )=\frac{1}{2}\text{ln}\frac{1+R_{\mathit{\text{xy}}}(\lambda )}{1-R_{\mathit{\text{xy}}}(\lambda )}$$

   The ‘z-transformed’ coherence values will be normally distributed with a standard deviation of approximately 1 (Rosenberg et al., 1989). The three frequency bands of interest were α (8 – 12 Hz), β (13 – 35 Hz) and γ (36 – 55 Hz). The integral of z-transformed coherence (or intermuscular coherence integral; ICI) within the three bands was calculated for the three muscle pairs (ICI_{i j}; i, j – 1: EDC, 2: FDS, 3: FDI).

3) Statistical analysis

For each of the force variables (dorsal and shear forces, CV, and task error), a repeated-measures analysis of variance (rmANOVA) was performed with the assistance condition (A₀, A_EXT, A_INT) as an independent variable (within-subject factor) (SPSS Statistics, Ver. 22; IBM Corp., Armonk, NY, USA). The p-values were corrected for multiple comparisons with a Bonferroni adjustment.

For the variables calculated from the EMG data (MAL, CCR, RC, ICI), we computed the relative change in each measure (from unassisted to assisted) under the three different assistance conditions for each subject in order to account for a relatively large between-subject variability in coordination patterns of hand muscles. Here, we used the log change in percent (L%; Δ_L) to resolve the problem of asymmetry and non-additivity of the relative measure [41], [42].

$${\mathrm{\Delta }}_L{\mathit{\text{MAL}}}_m(A_i,B_j)=\text{ln}\frac{{\mathit{\text{MAL}}}_m(A_i,\text{block}j)}{{\mathit{\text{MAL}}}_m(A_i,\text{block}0)}\times 100(L\%)$$

Here, m (muscle) = 1 (EDC), 2 (FDS), 3 (FDI); i (assistance) = 0 (none), EXT, INT; j (block/phase) = 0 (before assistance), 1 (assistance phase 1), 2 (assistance phase 2). Similarly, for CCR and RC values,

$${\mathrm{\Delta }}_L{\mathit{\text{CCR}}}_{\mathit{\text{mn}}}(A_i,B_j)=\text{ln}\frac{{\mathit{\text{CCR}}}_{\mathit{\text{mn}}}(A_i,\text{block}j)}{{\mathit{\text{CCR}}}_{\mathit{\text{mn}}}(A_i,\text{block}0)}\times 100(L\%)$$

$${\mathrm{\Delta }}_L{\mathit{\text{RC}}}_{\mathit{\text{mn}}}(A_i,B_j)=\text{ln}\frac{{\mathit{\text{RC}}}_{\mathit{\text{mn}}}(A_i,\text{block}j)}{{\mathit{\text{RC}}}_{\mathit{\text{mn}}}(A_i,\text{block}0)}\times 100(L\%)$$

$${\mathrm{\Delta }}_L{\mathit{\text{ICI}}}_{\mathit{\text{mn}}}(A_i,B_j)=\text{ln}\frac{{\mathit{\text{ICI}}}_{\mathit{\text{mn}}}(A_i,\text{block}j)}{{\mathit{\text{ICI}}}_{\mathit{\text{mn}}}(A_i,\text{block}0)}\times 100(L\%)$$

where m, n = 1, 2, 3; i = 0, EXT, INT; j (block) = 1, 2.

An rmANOVA was performed on each of these variables, with the assistance condition and phase as independent variables. Pillai’s Trace was used to determine p-value for each independent variable, and a significance level was set to 0.05. Again, for the pairwise comparisons, the p-values were corrected for multiple comparisons with a Bonferroni adjustment.

III. Results

All subjects achieved the task goal in all assistance conditions. However, an increased level of difficulty in completing target task and discomfort under assistance was reported by two female subjects (S₃ and S₁₀), possibly due to the size mismatch (i.e., the device did not fit their hands well). They required a longer time to complete the task, mainly due to slippage. This also led to the need to perform more trials, as well as an increased activation level of some muscles (over 90% of the maximum activation level in many cases) during the assisted task performance. Therefore, their data were excluded from the data analysis.

For all EMG measures, differences between the two phases of assisted conditions were not found to be statistically significant. Therefore, for all EMG measures (Δ_L MAL, Δ_L CCR, Δ_L RC, and Δ_L ICI), values averaged across the two phases were used throughout the analysis.

A. Task Performance: Fingertip Force

No significant difference was found in task performance (i.e., task error) between the three assistance conditions (p > 0.14 for all comparisons; see y-axis components of three solid arrows in Fig. 3). All subjects achieved the task goal in all conditions. Across all subjects, the force magnitude was highest in the dorsal direction (49.8N ± 12.1N), at a midlevel for the distal/proximal direction (30.0N ± 16.6N), and lowest in the medial/lateral direction (7.1N ± 6.1N).

While the assistance type did not affect the motor output in the task-relevant dimension (dorsal fingertip force), there was a significant ‘between-condition’ difference in the task-irrelevant dimension, shear fingertip force (x-axis in Fig. 3; p < 0.001). Under no assistance (A₀), the fingertip force was directed towards the distal direction with an average angle of 38.9°, the shear fingertip force magnitude being 80.7% of the dorsal fingertip force magnitude. However, under extrinsic muscle assistance (A_EXT), the fingertip force direction was reversed, pointing towards the proximal direction with an average angle of 20.6° (shear force magnitude equal to 37.5% of dorsal force magnitude), and thus rotated almost 60° from the no-assist condition. In contrast, under ET_INT assistance, the force direction was directed even further towards the distal direction, and its average angle was 43.7° (shear force magnitude 95.8% of dorsal force magnitude, or almost equivalent).

A significant difference in the within-trial variability of the force components was also observed between conditions. The CV for shear fingertip force significantly increased under extrinsic muscle assistance (A_EXT: mean ± SD = 0.326 ± 0.274; p = 0.032) when compared to the unassisted condition (A₀: mean ± SD = 0.149 ± 0.128), but did not increase under intrinsic muscle assistance (A_INT: mean ± SD = 0.118 ± 0.053; p = 0.664). In contrast, no between-condition difference was observed in the CV for dorsal fingertip force (0.060 ± 0.0.014 under A₀; 0.066 ± 0.020 under A_EXT: 0.062 ± 0.010 under A_INT; all p-values > 0.10).

B. Change in activation/control of individual muscles

1) Individual muscle activation

During unassisted task performance, muscle activation level was highest for the EDC muscle (28.8% ± 11.3%; averaged across all subjects), and lowest for the FDS muscle (10.5% ± 11.3%). In general, activation of the FDI muscle (21.5% ± 10.3%) was slightly lower than that of the EDC muscle, but with higher between-subject variability (i.e., higher SD value).

The activation levels of all muscles were found to significantly decrease under both assistance conditions (Fig. 4a), but the degree of reduction varied across the muscles. The relative log change in percent of muscle activation level (Δ_L MAL) was greatest for the intrinsic finger muscle in all conditions (Δ_L MAL₃; p < 0.001), while the degree of change was similar between the extrinsic muscles (p = 0.002 for Δ_L MAL₁; p = 0.001 for Δ_L MAL₂) (Fig. 4b). For all muscles, no statistically significant difference was found between the assistance conditions (all p-values > 0.99 for A_EXT vs. A_INT).

Fig. 4.

Change in muscle activation pattern: (a) muscle activation level; relative change in the muscle activation level, assessed by log percentage. For all three muscles, both types of biomimetic assistance induced significant decrease in their activation level.

2) Force-EMG correlation

Biomimetic assistance was found to interrupt/degrade the EMG-force correlation, assessed by r-values, for the intrinsic muscle (FDI) but not for the extrinsic muscles (EDC and FDS – see Fig. 5 for a representative case). Correlation between the fingertip force and the intrinsic muscle (FDI) was found to significantly decrease under assistance (p = 0.008), while correlations between the force and EDC and FDS remained unaffected by the assistance (Table 1). Here, note that the two assistance types produced different outcomes, as the effects of A_INT on FDI-force correlation was significantly greater (as assessed by p-values) when compared to those of A_EXT.

Fig. 5.

Representative force-EMG plots for (a) EDC; (b) FDI muscles (subject 5). The degree of change in both the correlation (r) and slope (s) was much greater for the FDI muscle. Here, different colors denote different trials.

TABLE I.

Relative change in the slope of the EMG-F_d correlation under three assistance types (unit: L%)

	Assistance type†
	A0	AEXT	AINT
EDC-Fd	−0.49±3.03	−5.63±8.09 (p = 0.28)	−4.46±7.43 (p = 0.57)
FDS-Fd	2.49±9.28	−5.55±17.95 (p>0.99)	−6.50±14.04 (p > 0.99)
FDI-Fd**	0.42±2.87	−14.77±14.71 (p = 0.08)°	−28.29±19.85 (p = 0.001)**

^°p < 0.1;

^*p < 0.05;

^**p < 0.01

^†The p-values for A_EXT and A_INT denote significance levels of the difference from A₀ condition (post hoc pairwise comparison), respectively. The p-values were adjusted for multiple comparison (Bonferroni correction)

C. Change in coordinated use of different muscle pairs

The effects of the biomimetic assistance on co-contraction ratios were found to be significantly greater for the extrinsic-intrinsic muscle pairs (EDC-FDI, FDS-FDI) when compared to the extrinsic muscle pair (EDC-FDS). There was a greater degree of increase in the Δ_L CCR values for the extrinsic-intrinsic muscle pairs (p = 0.008 for EDC-FDI; p = 0.004 for FDS-FDI) under assistance, while the effects of assistance on the Δ_L CCR of the extrinsic muscle pair (EDC-FDS: CCR₁₂) was not significant (p = 0.620) (Fig. 6). For the extrinsic-intrinsic pairs, significant differences were found between A₀ and A_EXT and between A₀ and A_INT, but not between A_EXT and A_INT (Fig. 6).

Fig. 6.

Change in the co-contraction ratio of the muscle pairs, assessed by log percentage. For all three muscles, both types of biomimetic assistance induced significant decrease in their activation level.

D. Change in common neural input to hand muscles

Overall, intermuscular coherence between the extrinsic muscles (EDC-FDS), assessed by ICI, was greater than for extrinsic-intrinsic muscle pairs (FDI-FDS and EDC-FDI) across the conditions (Table 2).

TABLE II.

Mean (SD) ICI values of the three muscle pairs under three assistance conditions (Unit: Hz)

	A0	AEXT	AINT
EDC-FDI	9.15 (7.80)	10.82 (6.89)	11.19(5.92)
FDI-FDS	3.27 (0.53)	3.59 (0.59)	3.65 (0.44)
EDC-FDI	4.01 (0.96)	3.93 (0.87)	3.99 (0.84)

The biomimetic input significantly increased the degree of neural coupling between the agonist (EDC, FDI) and antagonist (FDS) muscles, as assessed by the relative change in the intermuscular coherence integral (Δ_L ICI) values (Fig. 7). Note that, for each assistance type, the agonist-antagonist pair including the muscle targeted by the assistance was more greatly affected by the biomimetic assistance. The targeted assistance of the EDC muscle (A_EXT) was found to significantly increase the ICI value of the EDC-FDS pair (p = 0.04), while the assistance of interosseous muscles (A_INT) increased the ICI between the FDI and FDS muscles (p = 0.05).

Fig. 7.

Log change in the intermuscular coherence integral (Δ_L ICI) of the three muscle pairs. Increase in the ICI value of the EDC-FDS pair was more significant under A_EXT (p = 0.010) than under AINT (p = 0.072). Conversely, the ICI value of the FDI-FDS pair increased significantly under A_INT (p = 0.020), but not under A_EXT (p = 0.203).

IV. Discussion

The main purpose of this study was to examine the effects of targeted assistance of multi-articular hand musculotendons on the motor control of hand musculature. Our results confirmed that the biomimetic input, which was designed to provide targeted assistance to multi-articular hand musculotendons, induced significant changes in the activity of hand muscles during isometric force production at the fingertip. Furthermore, the degree of its effect was different across the muscles, which subsequently led to significant changes in the ‘coordination’ of these muscles, as indicated by the pairwise co-contraction ratio (CCR) values (Fig. 6). Change in the CCR values of different muscle pairs, induced by the biomimetic assistance, showed that the biomimetic assistance mainly affected the coordination between the extrinsic and intrinsic muscles (FDI-FDS, EDC-FDI: significant decrease in their CCR values under assistance), while the coordination between the extrinsic muscles (EDC-FDS) remained relatively unaffected (EDC-FDS: no significant change in its CCR value).

Subjects were able to fully take advantage of the externally-applied forces in both cases. Namely, they provided minimal extra force to achieve the task. For both assistance cases, the volitional force provided by subject was largely in the dorsal direction, with minimal shear force, and thus the resultant force the subject provided was close to the minimum needed to complete the task. The total resultant fingertip forces for A_EXT and A_INT, however, was quite different. This is evidenced by the striking similarity in the changes in activation patterns across all three muscles for A_EXT and A_INT (Fig. 4). Here, it appears that the difference in the task-irrelevant dimension (i.e. shear force) between the two assistance types was disregarded as it did not contribute to the task goal; it is possible that the two types of assistances may result in distinct muscle coordination patterns when a task of increased complexity, such as multi-joint movements, is performed.

While the effects on activation pattern were very similar, the different types of assistance had very different impacts on other parameters extracted from the EMG data. For example, there was a clear difference in the effects of the two types of assistance on intermuscular coherence (Fig. 7); the extrinsic assistance (A_EXT) significantly increased the intermuscular coherence (ICI) between EDC and FDS, but not between FDI and FDS muscles. In contrast, the effects of A_INT on ICI was significant for the FDI-FDS pair, but not for the EDC-FDS pair. The correlation analysis also showed that force-EMG correlation of both extrinsic muscles did not change under assistance, while the correlation between the intrinsic muscle and the fingertip force significantly decreased. These findings indicate the ‘differential’ effects of the biomimetic assistance on the control of intrinsic and extrinsic hand muscles, which may eventually lead to modulation of muscle coordination patterns if used in a longitudinal training protocol. Such a difference may arise due to the role of the intrinsic muscle in force production, while the extrinsic muscles mainly provide stability during force production [43]. Overall, biomimetic input increased the common neural input between the targeted agonist (EDC or FDI) and antagonist (FDS) muscles, which indicates an increased level of ‘functional’ connectivity between the muscles [44]. As common neural input to hand muscles may be diminished in stroke survivors [45], potential increase through the applied assistance could be beneficial for stroke survivors. The impact on the high-frequency (γ -band) was especially strong, which is thought to indicate common neural input mediated through brainstem pathways [46] that are found to be more involved in the control of distal muscles following stroke [47].

Our results indicate that the targeted assistance of the multiarticular musculotendons of the finger affects the activation of the hand muscles differently. Biomimetic (somatosensory) inputs, therefore, may provide a new way to selectively affect the use of different hand muscles. Various motor neuron diseases – whether impacting upper or lower motor neurons – have been found to degrade functional use of hand muscles, specifically for the intrinsic hand muscles [8], [48], [49]. Among patients with Parkinson’s disease, excitability of an intrinsic hand muscle (FDI) was significantly degraded [50]. Following stroke, the degree of impairment in the control of intrinsic hand muscles was highly correlated with the clinical test score [51], which underscores their functional importance. In general, coordination of the hand muscles is greatly impaired in stroke patients [13], [52], [53], and the degree of the impairment is highly correlated with the degree of their functional impairment [11].

The biomimetic input employed in this study appears to selectively promote neural coupling between the targeted agonist muscle and the antagonist muscles, as suggested by the intermuscular coherence values. At the same time, it also assisted in the task performance so that the task goal could be achieved at a lower activation level of the targeted muscle. In addition, the outcome of intermuscular coherence analysis also suggests that the proposed targeted muscle assistance could promote connectivity between the muscles in the functional context (i.e., ‘task-oriented’ training; [54]). Therefore, if used in the rehabilitation training of people with neurological impairments (e.g., stroke), targeted assistance of the multiarticular hand musculotendons could compensate/reinforce the reduced capacity of impaired muscles of the patients to achieve the task goal, unlike conventional ‘joint-based’ assistances that mainly focus on restoring task kinematics.