Neural Prosthesis Control Restores Near-Normative Neuromechanics in Standing Postural Control

Neural Control of Robotic Ankle Enabled Symmetrical, Synchronized Bilateral Limb Mechanics

To evaluate participants when using a dEMG-controlled prosthesis (dEMG) compared with when using their daily passive device (passive) under expected perturbation (an untrained task), we developed and used a custom perturbation platform (Fig. 1, Fig. 2a). In both expected pull and push perturbation conditions, we observed noticeable improvements in stability measures of participants with transtibial amputations for the dEMG condition compared with the passive condition. During these conditions participants generated anticipatory adjustments of their center of mass (CoM) against the direction of the perturbation. During the “push” condition, the able-bodied participants moved their CoM in the anterior direction (negative) before the weight was released and continued to hold their CoM forward well after perturbation (Fig. 2b). At the moment of release (t = 0 sec), AB participants had more anticipatory CoM excursion than the passive (p=0.012; p<0.05 is considered significant) but not the dEMG condition (p=0.092) (Fig. 2b). During the compensatory “hold” time (average value during t = 1–2 sec), we observed similar CoM excursion between AB and TT participants with dEMG control (p=0.516) but not with their passive daily device (p=0.004).

TT participants also improved center of pressure (CoP) excursion in the anterior-posterior (AP) and medial-lateral (ML) directions using dEMG control compared with their daily passive device for the pushing perturbation (Fig. 2c, d, respectively). In their daily device, TT participants generated anticipatory CoP excursion with their intact foot at t = 0 s in the posterior direction; however, the CoP excursion under the passive prosthetic foot remained significantly less (release: 31.48(±14.62 mm) vs. 6.94(±8.46 mm), p = 0.005). To evaluate compensation from the perturbation, the CoP excursion was averaged during a “hold” period after the perturbation, defined as t = 1 to 2 sec. The difference of CoP excursion between the intact and prosthetic foot was not significant after the perturbation (hold: 61.17(±26.72 mm) vs. 38.17(±13.33 mm), p = 0.210). Using dEMG control, TT participants generated symmetrical CoP excursion between their intact and prosthetic foot for both time points (release: 24.96(±18.29 mm) vs. 20.05(±16.12 mm), p = 0.509; hold: 48.66(±14.96 mm) vs. 33.61(±23.40 mm), p = 0.210). We observed similar symmetry of CoP excursion in AB participant behavior (Fig. 2c, d). In the ML direction with their daily passive device, TT participants demonstrated anticipatory CoP ML excursion toward their intact limb (release: −5.68(±4.18 mm)) and similar compensatory CoP ML excursion (hold: −12.08(±7.02 mm)). However, with the dEMG control participants moved CoP ML toward their prosthetic in anticipatory (release: 1.93(±7.99 mm)) and compensatory time points (hold: 2.69(±11.60)). These CoP ML excursions were significantly different between passive and dEMG conditions for the compensatory time point (hold: p = 0.036) but not for the anticipatory time point (release: p = 0.154). Using the dEMG control, participants had similar magnitude CoP ML excursions compared to our AB control participant (Fig. 2d) (release: p = 0.855, hold: p = 0.721). Results for the pull condition can be viewed in the supplemental materials (Fig. S1).

Before the perturbation, able-bodied individuals had a resting plantarflexion torque (Fig. 2e, t = −0.5 sec) and generated an anticipatory change in torque before perturbation. Shortly after perturbation onset (t = 0 sec), individuals generated dorsiflexion torque primarily with their right leg. Using dEMG control, TT participants generated similar levels of torque from both their prosthetic and intact limbs. However, using their passive device, individuals did not generate dorsiflexion torque with their prosthetic limb and had delayed generation of dorsiflexion with their intact limb (Fig. 2e). Representative EMG from TT2 participant also demonstrated differing activation patterns between the residual and intact tibialis anterior (TA) muscles (Fig. 2f). Using their passive device, TT2 generated anticipatory activation of intact shank and thigh muscles but did not noticeably activate their residual TA or rectus femoris (RF). Using dEMG control, TT2 significantly activated both residual and intact limb shank and thigh muscles (Fig. 2f).

Neural Control of Robotic Ankle Enabled Matched Intact and Residual Limb Joint Motions

We observed that participants with transtibial amputation modified postural configuration and overall joint strategies when using dEMG control to match their intact limb more closely in both “Push” and “Pull” conditions (Fig. 3 and Supplementary Material Tables S1 and S2). At the release of the push perturbation, the participants actuated their intact and prosthetic ankles in opposing directions while using their daily passive device (Fig. 3a, p = 0.0007). However, we observed no difference in joint angle excursions between intact and dEMG controlled prosthetic ankles (p = 0.487). During the perturbation hold (t = 1–2 sec), we observed a significant difference between intact and prosthetic ankle joint excursions in the passive but not the dEMG control condition (passive: p = 0.005; dEMG: p = 0.195). We did not observe a significant difference between intact and residual knee joint excursions in the passive condition at the release point (Fig. 3a, p = 0.060); however, this result approached significance. We did not observe interlimb differences in the dEMG control condition for the same time point (Fig. 3b, p = 0.8345). We observed a significant difference between the intact limb knee joint and residual limb knee joint during the perturbation hold for only the passive condition (passive: −1.57°(±4.03°) vs. 3.96°(±1.97°), p = 0.037) and not the dEMG control condition (0.23°(±5.15°) vs. 3.20°(±6.00°), p=0.1437, and these values indicate mean (±standard deviation)). For hip joint comparisons, we only observed no significant difference between limbs for the passive condition or dEMG control condition at release or hold time points.

dEMG Control of Robotic Ankle Normalized Postural Control Strategy and Reduced Falls and Compensatory Steps in Participants

We observed distinct changes in balancing strategies as TT participants used dEMG control compared with their passive device (Fig. 4). While using dEMG, three TT participants used predominantly an ankle joint strategy, as evidenced by relatively small excursions from the knee and hip joint and a large, synchronized excursion from the ankle joints (Fig. 4a). We observed that two TT participants used a mixed strategy, with simultaneous excursion from all joints (Fig. 4b). These participants still improved synchronization between prosthetic and intact limb ankle joint excursion compared with their passive device. Overall, nearly all participants failed to maintain static standing balance at least once (participants took a step or allowed handlebar to reach the end of the track) while using their passive device. Although one participant still failed to maintain balance in some repetitions using dEMG control only, all remaining participants were able to maintain static standing balance (Fig. 4c).

dEMG Control Resulted in Near-Normative Neural Coordination Patterns

To study neural coordination, participants were also asked to perform postural sway task while EMG signals were recorded from bilateral muscles of lower limbs. To illustrate neural coordination changes on an individual basis, we plotted muscle module structures and activation coefficients from one representative AB participant (AB1) and one representative TT participant (TT5) using their daily, passive device and the dEMG control (Fig. 5). Using TT5 as our representative participant, we observed noticeable contribution from a single anterior muscle when TT5 used their passive device (intact rectus femoris had the highest contribution in muscle module, although predominantly posterior muscles were activated; Fig. 5a), and this contribution was notably less when using the dEMG control prosthetic ankle. For ‘backward’ muscle modules, TT5 generated two muscle modules (Fig. 5b); however, TT5 used only one module with the dEMG control. This muscle module used in dEMG condition was notably similar to that observed in AB participants (AB1 as a representative example in Fig. 5) for both the muscle module structure and the activation coefficient.

In fact, we observed that nearly all TT participants employed muscle module structures similar (r value above 0.735 significance level) to those of AB participants when using the dEMG-controlled prosthetic ankle (Fig. 6). Using their passive, daily device, only 3 of 12 muscle module structures across participants were significantly correlated to average AB module structure. Except for TT4, all participants demonstrated muscle module structures using dEMG control that were similar to AB behavior. Six of 11 muscle modules were above the significance threshold (r>0.735), and 3 modules were close to significance (TT1: 0.69, TT3: 0.70, TT3: 0.72). Three muscle modules were not included in this analysis because they were not labeled as a ‘Backward’ or ‘Forward’ sway (TT1 Passive, TT2 Passive, TT1 dEMG Control). One participant, TT4, did not show muscle module structure similar to AB behavior in either passive or dEMG control conditions. Each muscle module from ‘Backward’ Sway from TT4 primarily contained muscles from a single leg (one module for residual limb muscles, and one module for intact limb muscles). Thus, although the individual correlations for each of these muscle modules for TT4 was low compared with AB ‘Backward’ modules, TT4 grouped residual muscle activity with activity from other muscles on the residual limb. All muscle modules for TT participants and the averaged muscle module structure for AB participants can be seen in supplementary figures (Fig. S4–S6).

Participants

We recruited five individuals with unilateral transtibial amputation to participate in this study. All participants were male. This experimental protocol was approved at the University of North Carolina at Chapel Hill Institutional Review Board. Before participating in this study, each participant provided written consent after the nature and potential consequences of the study were explained. Information about the five TT participants is provided in Table 1. The average age of amputee participants was 43.4(±11.72) years. We also tested five able-bodied participants to provide a reference for both biomechanics and neural behavior (average age: 39.2(±11.43) years. All TT participants were active, community ambulators who wore their prosthesis at least 12 hours a day. No participants reported nerve damage or issues with sensation. TT5 had a notable comorbidity of dysvascular disease; however, no other participants reported any comorbidities. One participant was also an upper-limb amputee (TT2, shoulder disarticulation) and held the perturbation platform with a modified grip.

Pneumatic Actuated Ankle Prosthesis

For this study, we fit a pneumatically actuated ankle prosthesis for each TT participant. Each participant fit their daily prosthetic socket to the pneumatic ankle, and the alignment for each participant was set by trained prosthetist while the prosthesis was locked in a 90° position. This prosthetic ankle contained four pneumatic artificial muscles (PAM), two placed anteriorly to generate dorsiflexion moments and two posteriorly for plantarflexion moments. These PAMs were used to produce similar force-length dynamics of intact musculature and were actuated using EMG magnitude taken from the envelope of residual muscle activity. More details on the ankle prosthesis can be found in [58].

Direct EMG Control

In this study we took the envelope of previously antagonistic residual muscle activity to drive prosthetic ankle dynamics directly and continuously. We calculated the envelope of the residual tibialis anterior (rTA) and residual lateral gastrocnemius (rGAS) activity in real time (dSPACE, CLP-1103, 0 to 10 V output) to generate a smoothed control signal. We processed the residual muscle activity in real time by first applying a high-pass filter (100 Hz, 2^nd Order Butterworth) to reduce the effect from potential signal artifacts. Our real-time filtering then rectified the EMG signal and applied a low-pass filter (2Hz, 2^nd Order Butterworth). Our system then sent the filtered control signal from each muscle to the pressure regulators responsible for generating pressure proportional to the input voltage (0 to 10 V for 0 to 90 psi). We applied a gain to each input control signal at the beginning of each session such that a maximum contraction from each muscle resulted in a control signal between 9 and 10 V. To establish a set ankle stiffness, we applied baseline air pressure of approximately 2 V for each participant. This process is detailed in [52] as well as Supplementary Material.

Experimental Protocol

Similar to our procedure in previous work [52], we introduced TT participants to the pneumatic ankle with several acclimation trials before the start of training and evaluation. After acclimation, our study consisted of a minimum of three training sessions and two evaluation sessions. The details of this acclimation and training can be found in the Supplementary Materials.

To evaluate the ability of amputees to produce anticipatory postural adjustments under expected perturbation using dEMG control, we developed a platform capable of delivering expected perturbations in two directions: push and pull (Fig. 1, Fig. 2a). In this setup, we asked each participant to hold a handlebar placed at shoulder level while standing in a relaxed posture. We placed a light-emitting diode (LED) strip at the eye level of the participant. For the participant to know the perturbation timing, we switched off the LED strip 0.6 second before the electromagnet was released via two relays (5 V, Parallax, CA, USA) controlled using LabVIEW (NI, Austin, TX, USA). In this setup, the perturbation was delivered by a hanging weight that was connected to the handlebars by two cords via a series of pulleys such that the vertical force of the weight was translated into a horizontal force on the handlebars. We placed an electromagnet in series with one of the cords to control its connection with the weight. Thus, when the electromagnet was released, all weight was transferred horizontally, causing a force in the direction of the cord that remained attached to the handlebar. The location of the electromagnet was switched based on the desired perturbation direction (push or pull). We placed one load cell on each side of the handlebar to measure the force each participants applied before and during perturbation. A trial was removed from the analysis if participants applied force to the handlebar before the light switch. Before each repetition, we visually confirmed the participant’s stance to ensure they had a vertical, relaxed posture with both feet placed symmetrically and shoulder-width apart. We repeated each perturbation five times for each condition. We provided each participant with five practice trials for each condition before the beginning of testing to familiarize them with the task. The order of testing each weight and direction was randomized before the start of the evaluation session. The magnitude of the perturbation was chosen based on pilot testing because the perturbation magnitude was sufficiently difficult that amputee participants would potentially be unable to maintain balance during either the push or pull condition. We conducted this evaluation using 10% of the participant’s body weight as the perturbation magnitude in both a pull and a push direction. On separate days, we then evaluated dEMG control with both perturbation trials and the trained activities. We conducted the same evaluation on a separate day with participants using their daily device. The order of evaluation between devices was randomized, where TT1, TT2, and TT3 conducted the passive evaluation after the training and evaluation with the dEMG control. TT4 and TT5 conducted the passive evaluation before training and evaluation with the dEMG-controlled ankle.

Measurements

During all sessions, we collected EMG activities from intact and residual limb thigh and shank muscles. We placed EMG sensors (Motion Lab Systems, MA-420, Gainx20, Baton Rouge, LA, USA) on the following muscles: intact tibialis anterior (iTA), intact lateral gastrocnemius (iLGAS), intact soleus (iSOL), intact and residual limb rectus femoris (iRF, rRF), intact and residual limb vastus lateralis (iVL, rVL), intact and residual limb biceps femoris (iBF, rBF). We determined the location of muscle via palpation and anatomical landmarks [59]. For rTA and rLGAS, which are inside of the prosthetic socket, we placed low-profile EMG sensors (Neuroline 715, 1mm height, Ballerup, Denmark). We located muscles via palpation while participants contracted and relaxed muscles [60]. When residual muscle flexion was difficult to locate via palpation, we placed EMG sensors based on anatomical landmarks. We then iteratively shifted EMG sensor location to determine whether other locations provided improved the observed signal-to-noise ratio. We routed cables to avoid bony landmarks and connected them to pre-amplifiers (Motion Lab Systems, MA-412, Gainx20, Baton Rouge, LA, USA) outside of the prosthetic socket. We connected all sensors to an amplifier (MA300-XVI, Gain x1000).

In each session, we collected ground reaction forces (GRF) and moments under each foot using an instrumented split-belt treadmill (1000 Hz, BertecCorp, Columbus, OH, USA). CoP values were calculated under each foot using GRF and moments. We calculated overall CoP using the weighted average under each foot, weight by GRF [52]. We collected full-body kinematics (100 Hz, 43 markers, VICON, Oxford, UK). We calculated CoM position using the analysis software (Visual 3D, C-Motion, Inc., Germantown, MD, USA). We low-pass filtered data from the treadmill (GRF, CoP), (20 Hz, 4^th Order Butterworth). From the perturbation setup, we collected loading data from the handlebar via two tension load cells (Omega, Stamford, CT, USA) placed on the anterior and posterior side. We recorded the control signal sent to each relay for light switch and electromagnet from the same sampling rate as our split-belt treadmill.

Expected Perturbation Data Analysis

We processed data offline using MATLAB (MathWorks, Natick, MA, USA). For perturbation trials, we extracted and windowed each perturbation to 600 ms before load release and 2000 ms after release. To evaluate participant stability during each release perturbation, we selected the CoM excursion, CoP excursion, and ankle joint torque during each condition. For each repetition, we calculated the average of each measure from 600 to 500 ms before perturbation. We then subtracted this average value from each measure during each perturbation. We took the average of the first five recorded trials for each condition, each device, and each participant. We plotted the time series of each selected measure to compare behavior between the AB control, TT Passive, and TT dEMG control condition behavior. We plotted the push condition response because this was the most challenging condition for all participants (as evidenced by frequent stepping responses required by TT participants with their passive device). For quantitative analysis, we selected two time points from each repetition to understand the anticipatory and compensatory postural responses. For anticipatory responses, we extracted the value of each measure at the moment of load release (t = 0 sec). Because the perturbation required participants to hold the load after release, we calculated the average value of each measure for 1000 to 2000 ms after the load release, termed “hold.” We selected this time in order to capture the long-term behavior of participants, where most TT participants using their passive device initiated and finished compensatory movements (steps, using guard rails) during this time window.

To understand joint level strategies of participants when using their passive prosthesis compared with dEMG control, we evaluated the ankle, knee, and hip joint excursions during all perturbation repetitions with the same windowing described above. We subtracted the average joint angle of 600 to 500 ms before load release from the full time series. For evaluation, we then extracted values from “release” and “hold” time points.

For qualitative analysis of full-body strategies to maintain balance, we qualitatively evaluated the combination of joint angle excursions from each participant using the dEMG control by observing two distinct joint excursions strategies. For three participants, we identified primarily an “ankle strategy” when responding to the disturbance (where the hip and knee joint excursions did not exceed 2° in either direction). For the remaining two participants, we noticed a “mixed strategy” where participants moved all joints more than 2° for each joint simultaneously. We empirically defined these boundaries of excursions for strategies based on previous study of standing balance under perturbation [61]. During each condition, we observed the number of instances where participants could not maintain balance. They either held onto the handlebar until it reached either the forward or backward stop or took a compensatory step (Fig. 4c). We counted the number of these situations for each condition.

Voluntary Postural Sway (Muscle Module) Data Analysis

For the voluntary postural sway, we processed all data offline using MATLAB. We selected the first trial from the final evaluation day. We first low-pass filtered all data from the instrumented split-belt treadmill (2^nd Order, 20 Hz, Butterworth). We calculated CoP in the anterior-poster (CoP_AP) direction using GRF and moment data from the treadmill [62]. We windowed each individual sway, selecting the forward-most extremes of CoP_AP as the beginning and end of the sway. For each trial, we first removed any sways where the participant was unable to maintain balance and required a step forward or backward. For the muscle module analysis, we selected the middle 12 sways [45] from the remaining data. For evaluation of kinematic and kinetic variables, we included all sways from the remaining data.

We post-processed EMG data by high-pass filtering (3^rd order, 35 Hz, Butterworth), rectifying, then low-pass filtering (3^rd order, 40 Hz, Butterworth) [63]. We down-sampled the collected EMG by averaging data into 50-ms time bins [63]. We then concatenated EMG data from all sways for each condition and participant, generating data matrices of m (# of muscles) × b (# of bins) [63]. We did not time normalize EMG data, which resulted in data matrices of different sizes between participants [63]. EMG matrices for each condition and participant contained a minimum of 700 data points. We normalized EMG values of each muscle vector by a maximum voluntary contraction (MVC) for each muscle trial gathered at the beginning of each testing session. If we observed a muscle activation during the voluntary postural sway task that was larger than the collected MVC, we normalized the given muscle vector by this value instead. We did this separately for each condition and participant. We then scaled each muscle vector to have unit variance in order that equal weighting was applied to each muscle during the muscle mode analysis. After muscle module extraction, we removed this scaling for each muscle vector.

Statistical Analysis

We conducted statistical analysis of the data using statistical software (JMP, SAS, Cary, NC, USA). We used two-way ANOVAs for CoP excursion and joint angle excursions at “release” and “hold” time points with device (passive vs. dEMG control) and limb (intact vs. tesidual) as the main and interactions effects. We compared CoM and mediolateral CoP behavior between AB and TT participants using each device (AB vs. passive, AB vs. dEMG control) using one-way ANOVA. We further compared those same outcome measures with TT participants between device conditions (passive vs. dEMG control) using one-way ANOVA. For each analysis, we included participant as a random effect. We conducted pair-wise comparisons using Tukey’s honest significant difference test. We evaluated whether each analysis failed normality assumptions using a Shapiro-Wilk test of studentized residual distribution. We also evaluated whether each analysis failed equal variance assumption using Levene’s test for equal variance. When an analysis failed the normality or variance assumption, we conducted non-parametric Wilcoxon tests. We set a threshold for significance at a = 0.05 for all tests. The tests that failed assumptions required for parametric tests can be seen in supplementary material (table S3) as well as excluded trials and justification. We did not remove any outliers from our data.