Phase-Based Haptic Feedback for Improved Neural Control of a Robotic Prosthetic Ankle: A Preliminary Study

Abstract

While neural control of robotic prosthetic legs through direct electromyographic (dEMG) control has shown promise for restoring human-prosthesis coordination, many persons with amputations struggle to generate the appropriate electromyographic (EMG) control signals for locomotion. The objective of this study was to develop a novel, phase-based haptic feedback system to train persons with transtibial amputations (PWTA) to produce appropriate muscle activation patterns for dEMG control of a robotic ankle while walking. The system provided (1) instructional feedback to guide the timing of residual muscle activation and (2) corrective feedback indicating whether the target activation pattern was achieved. The feedback system was preliminarily evaluated with three PWTAs. Biomechanics and EMG activation patterns were assessed before, during, and after feedback was provided. All participants altered their EMG activation patterns and stance duration was increased on both limbs. These changes led to all three participants increasing the step-to-step transition work done by the robotic prosthesis and two of the three increasing the peak robotic ankle power. Notably, all these changes persisted even after feedback was removed. These results demonstrate the feasibility of phase-based haptic feedback as a training tool for dEMG control of robotic prosthetic legs, improving human-prosthesis coordination.

I. Introduction

ROBOTIC prostheses can generate positive mechanical power, mimicking the function of a biological ankle and foot during walking [1]–[3]. Most robotic prosthetic ankles utilize autonomous control to produce positive power and restore ankle push-off function during the step-to-step transition in walking [3]. Some previous studies have demonstrated potential benefits for persons with transtibial amputation (PWTA) using an autonomous powered prosthesis compared to a passive prosthesis, such as an increase in the mechanical work of the prosthetic side [2], [4], [5], a decrease in the work of their intact leg during collision [2], [4], and an improvement in walking energetic economy [2], [5]. However, these potential benefits of walking using an autonomous powered prosthesis compared to a passive prosthesis are not consistently reported in the literature. Some studies found no change in the collision work of the intact leg [6] and no improvements in walking energetic economy when walking with robotic prosthetic ankles [6], [7] compared to a passive prosthesis.

One potential explanation for this inconsistency in the energetic benefit of using robotic prosthetic ankles is a lack of coordination between the computerized prosthesis control and human motor control, two separate controllers. One challenge with autonomously controlled robotic prostheses is defining appropriate control parameters, such as the magnitude and timing of applied joint torque [4], [8], [9], that need to be coordinated with each individual user’s motor control and biomechanics. Lack of human-prosthesis coordination may lead to insufficient mechanical energy transfer from the prosthesis to human user and hinders the potential benefits of the robotic prosthetic device [10]. For example, studies showed that when the prosthetic ankle plantarflexion torque was applied too early at the mid stance, the trailing limb posture was still vertical to the ground, and the push-off torque, therefore, propelled the limb upward rather than forward reducing the mechanical work done by the prosthetic leg during the step-to-step transition [9]. Hence, many researchers have devoted their efforts to developing effective methods to tune autonomous prosthesis control parameters to optimize energetic efficiency [6], [8], [9], [11].

Another approach to ensure human-prosthesis coordination is to use the human nervous system as a unified controller. Advancements in powered prosthetic devices and control methods have enabled persons with amputation (PWAs) to use their residual muscle activity to control prostheses. One method is direct electromyographical (dEMG) control, where efferent neural signals measured from residual muscles directly modulate prosthesis joint dynamics [12]. In dEMG control, the prosthesis is integrated with the human’s nervous system, which may enable seamless human-prosthesis coordination, making the method versatile to different walking and postural control tasks [13]–[19]. However, this method requires people with amputation to generate appropriate muscle activation patterns necessary for producing biomimetic mechanical ankle power during walking [13], [14], [18], [19], which is challenging for many individuals with lower-limb amputations. This may be due to alterations in muscle physiology that occur after amputation and may disrupt the patient’s internal model [20] or could also be related to the lack of proprioception from the amputated limb [21]. Additionally, adaptation occurs in individuals with lower limb amputation who have used passive devices for years. These adaptations lead to altered gait patterns in prosthesis users who rely more on their intact limb for gait and balance [22], [23], and these altered gait patterns may persist even when using a robotic prosthesis. For example, previous studies found that even after 2–6 hours of walking with a dEMG-controlled prosthesis, PWAs still displayed similar peak prosthetic ankle power as passive prostheses [14], [18], [19], despite the device being capable of generating greater mechanical power. This suggests a need for effective approaches to re-train PWTAs to use residual muscles to power the robotic prosthetic ankle in walking.

Several studies have investigated different approaches to improve the ability of people with lower-limb amputation to utilize dEMG-controlled prostheses [14], [17], [19]. For example, physical therapy-guided training can improve stability and postural control in PWAs while using a dEMG-controlled prosthesis [17]. This approach, however, requires the presence of physical therapists and verbal and physical feedback. Song et al. demonstrated a novel surgery, referred to as the agonist-antagonist myoneural interface (AMI), which evokes natural proprioceptive sensation in the amputated limbs of PWAs [14]. Through this procedure, PWAs were able to walk using a dEMG-controlled prosthesis with greater push-off power, compared to participants without the surgery [14]. However, the invasive nature of this procedure limits its accessibility to many PWAs, highlighting the need for a non-invasive alternative. Another approach is to provide feedback displays of EMG signals to the users to improve their ability to activate their residual muscles during walking. Huang et al. found that when given visual feedback of their residual muscle activity, compared to the desired EMG activity profile, PWTAs using a dEMG-controlled prosthesis increased their prosthetic ankle push-off power in walking compared to without visual feedback [19]. However, it was not demonstrated whether this effect would be retained once feedback was removed [19]. In addition, visual feedback requires the participants to continuously allocate visual attention, which may interfere with the natural use of vision for walking.

Inspired by the need for an effective training system for PWTAs in using a dEMG-controlled prosthetic ankle and the promising results derived from training with visual feedback of the EMG [19], the objective of this work was to develop a novel haptic feedback system to train PWTAs to produce appropriate EMG activity from the residual gastrocnemius (rGAS) for direct control of a robotic prosthetic ankle to achieve effective push-off in walking. Compared to existing literature, the primary contributions of this work are as follows. (1) We designed a feedback system using a high-density vibrotactile haptic vest to provide instructional cuing to aid patients walking with a dEMG controlled robotic prosthetic ankle. Vibrotactile feedback was chosen because it is less obtrusive and causes less interference than visual or auditory feedback. In addition, haptic technologies are lightweight and wearable [24] and can encode various feedback information easily and effectively [25]. Furthermore, haptics systems have shown effectiveness in rehabilitation [26], [27] and assistive device (including prostheses) training and use [28]–[30]. (2) We integrated continuous gait phase estimation with the haptic encoder to train the prosthesis users in rGAS contraction timing. Continuous phase estimation has become popular recently in phase-variable control of wearable robotics [31]. To our knowledge our study was the first attempt to use a phase-variable for a haptic feedback design. A preliminary evaluation of this system was done with PWTAs, investigating both the influence of feedback on walking gait as well as the after effects once feedback was removed. The outcomes of this work may inform the design of a novel haptic system to train individuals with lower-limb amputations to effectively use continuous dEMG-controlled robotic prostheses, enabling seamless human–prosthesis coordination during locomotion.

II. Methods

A. Feedback System

To deliver vibrotactile feedback, this study used the Tactsuit x40 (bHaptics Inc., Daejeon, South Korea), a high definition haptic vest with a 4×5 grid of individually programmable motors on both the front and back of the device.

The feedback strategy was designed to deliver two types of feedback from the vest simultaneously. The front of the vest was used to deliver phase-based feedback, instructing participants when to flex their rGAS, while the back delivered corrective feedback, informing the participant after each step whether their rGAS was activated at the correct time.

Phase-Based Feedback:

Feedback on the front of the vest was designed to instruct the participant when to contract their residual muscles to plantarflex the device. As illustrated in Fig. 1, the haptic stimuli began at the bottom of the vest and progressed upwards. Participants were instructed to flex their rGAS muscle once the top row of motors activated.

Fig. 1.

Experimental Setup: A) Real time phase-based feedback was provided, with the stimuli moving from the bottom to the top of the vest. Participants were instructed to flex their rGAS when the top row of motors activated. B) The participant walked on the treadmill with the pneumatic powered ankle at a constant speed of 0.75 m/s. C) Corrective feedback was given as a 500 ms pulse from all motors in the top 2 rows on the back of the vest whenever incorrect push-off timing was detected.

The target push-off time was set to 56 ± 5% of the gait cycle. Each gait cycle was defined as the period from residual-side heel strike to the subsequent residual-side heel strike. We chose 56 ± 5% of the gait cycle as the target push-off time based on a previous study of walking with an autonomous robotic ankle prosthesis [9]. Malcolm et al. found that setting peak prosthetic ankle torque during push-off to occur between 52% and 58% of the gait cycle improved metabolic efficiency and increased prosthetic leg mechanical work during walking compared to earlier push-off timings [9].

Each row of motors activated one at a time. The top row was active for 10% of the gait cycle, between 51% and 61%, after which the stimuli would reset to the bottom row of the vest. The remaining 90% of the gait cycle was divided evenly between the bottom 4 rows, with the stimuli ascending one row for every 22.5% of the gait cycle that passed until it once again reached the top row at 51% of the gait cycle. In this way, the bottom 4 rows acted as a feedforward mechanism, allowing the participant to sense how long until the next push-off should occur.

Corrective Feedback:

The back of the vest was used to deliver corrective feedback, informing the user after each gait cycle whether or not they activated their rGAS at the correct time and with sufficient magnitude. To determine a “successful” gait cycle, two separate criteria had to be met. First, the EMG control signal needed to reach at least 80% of its maximum value, and second, the peak EMG must occur between 51% and 61% of the gait cycle. If either of these criteria were not met, corrective feedback was applied as illustrated in Fig. 1C.

B. Phase Estimator

To estimate gait phase in real-time, this study deployed the method proposed by Mituniewicz et al. [32]. This approach first performs a rigid body estimate of a limb’s anterior-posterior position and velocity, using the thigh and shank. The phase angle created from this translational phase portrait naturally resets itself when the limb is at its maximum anterior position and about to begin moving posteriorly. A more linear phase estimator is then computed from the initial phase angle, the time passed since the cycle reset, and a real-time velocity estimate via a linear regression trained on open-source data of people without impairment [33]. This regressed estimator is then weighed with the initial phase angle to ensure its ends align with the initial phase angle’s. With respect to a gait cycle normalized by foot-contact from the same limb, this phase angle reset event typically occurs between 2–4% before the ipsilateral contact event in people without impairment, depending on their speed [32]. This discrepancy was accounted for by constantly offsetting the phase estimator by approximately 4%, such that 60% of phase estimated gait cycle corresponds to 56% of foot-contact normalized gait cycle.

C. Robotic Prosthesis

In this study, a powered prosthetic ankle with pneumatic artificial muscles (PAMs) was used to test the dEMG control strategy. The prosthesis had two PAMs in the front to control dorsiflexion and two in the back to control plantarflexion (Fig. 1B). Dorsiflexion and plantarflexion were controlled through proportional pressure regulators (MAC 36 series valve and PPC36A Pressure Regulator, Wixom, MI, USA) which modulated the PAM internal pressure from 0 up to 90 PSI as the control signal reached it’s maximum output of 10V. This internal pressure produced a joint torque at the ankle. To provide a baseline stiffness to the ankle and ensure the ankle would return to an anatomically neutral position once the participant relaxed their residual muscles, constant offsets were added to both sets of PAMs between 2–3 V or 18–27 PSI [15]. These offsets allowed for dorsiflexion to naturally occur during the swing phase when the participant relaxed their muscles.

Participants were given volitional control over plantarflexion. As the feedback strategy used in this study was designed to aid with push-off mechanics during walking, only the rGAS muscle was used for dEMG control. To achieve this, bipolar neo-natal surface electrodes (Ambu Neuroline, Columbia, MD, USA) were placed on the rGAS. EMG amplitude was recorded in real-time at 1000 Hz (Motion Lab Systems MA-400, Lake Elsinore, CA, USA). This signal was high pass filtered (100 Hz), rectified, and then low-pass filtered (2 Hz) in Simulink (MathWorks Inc., Natick, Massachusetts, U.S.). The 2 Hz cutoff frequency was chosen to match the force bandwidth of the PAMs [13]. This EMG control signal was then sent to the pressure regulators to control plantarflexion. A gain was set such that the EMG control signal would reach approximately 10V when the participant reached approximately 50% of their maximum voluntary contraction (MVC) found during the start of each session.

D. Experimental Protocols

We recruited 3 PWTA participants (P01–P03) and their demographics are summarized in Table I. All participants gave their informed written consent prior to participation. The study was approved by the institutional review board (IRB) of NC State University (IRB No. 24436).

Table I.

Participant Demographics

Participant	Gender	Height (m)	Weight (kg)	Age	K-Level	Cause of Amputation
P01	M	1.83	97.98	55	4	Trauma
P02	M	1.73	115.65	33	4	Trauma
P03	F	1.63	71.17	30	3	Cancer

The experiment required three sessions for each participant, each session taking place on a different day. In the first session, the participant was fitted with the robotic prosthetic ankle by a licensed prosthetist. Then, in the second session they were trained on how to walk on a treadmill with the device without the phase-based feedback. Finally, the third session was used to evaluate how the phase-based feedback system altered participant’s gait and EMG activation strategies. This evaluation session took place approximately 1 week after the training session.

Each session began with an MVC trial. While sitting, participants were instructed to “point your toes down as hard as you can” and hold for 1–2 seconds, repeated 3 times. The MVC was defined as the maximum value of the filtered EMG signal recorded during this trial.

Training:

Participants began training by walking on the treadmill with the dEMG-controlled device, starting at a speed of 0.4 m/s. While walking on the treadmill, they wore a safety harness and were told to hold onto the handrails. After each 1 minute walking trial, participants were allowed a 2 minute break, then the next trial would begin with the speed increasing by 0.05 m/s. This process continued until the speed reached 0.5 m/s. At this point, participants were instructed to hold the handrails only if necessary to maintain balance. Afterwards, the speed would increase by an additional 0.05 m/s after each trial only if the participant walked for 1 minute at that speed without holding onto the handrail. This process was repeated until the participant was able to walk at 0.75 m/s for 1 minute without holding onto the handrail, completing the training. To minimize the effects of fatigue, a minimum of 2 minutes of rest was given between each trial with more time allowed at the participant’s request.

Evaluation:

During the evaluation session, the same EMG placements were used as the training session. The evaluation session consisted of 8 walking trials, where participants walked on the treadmill for 2 minutes at a constant speed of 0.75 m/s. For the first 2 trials, participants walked normally with no feedback provided. Then, 4 trials were performed in which participants received the phase-based feedback. Finally, feedback was turned off and participants were asked to perform 2 more walking trials. After each walking trial, participants were given at least 2 minutes of rest to minimize the effects of fatigue.

E. Data Collection and Pre-processing

EMG signals from the rGAS were recorded at 1000 Hz. For analysis, these recordings were demeaned, passed through a bandpass filter (4th-order butterworth, 10–350 Hz), rectified, and then passed through a low-pass filter (4th-order Butterworth, 6 Hz) to obtain the filtered EMG envelope.

Motion data were captured at 100 Hz from a 12-camera motion capture system (Vicon, Oxford, UK), which recorded the position of reflective markers placed around the pelvis and lower extremities. In addition, a dual belt treadmill with built-in force plates (Bertec, Columbus, OH, USA) recorded the ground reaction forces (GRFs) under each foot at 1000 Hz. All the recordings were synchronized. We used Visual3D (HAS-Motion, Kingston, Ontario, Canada) to calculate the joint kinematics and kinetics during walking based on the measurement of marker positions and GRFs.

To obtain the thigh and shank angles in real-time for the phase-estimator, segment angles were streamed from Vicon into MATLAB through the Vicon SDK at 100 Hz. Segment velocities were calculated using a first-order backwards finite difference method that was filtered using a first-order low-pass filter (cutoff 10 Hz) prior to phase angle calculation.

F. Data Analysis and Evaluation Metrics

For each participant, we analyzed data from the final trial of each evaluation condition: before feedback (BF), with feedback (WF), and after feedback (AF). For each of these trials, a segment of 20 consecutive gait cycles was selected, in which the participant generated the most consistent EMG signals. This segment was found by using a 20 gait cycle long sliding window with 15 gait cycles of overlap calculating the mean filtered EMG signal. To determine similarity between gait cycles, the mean cross-correlation at zero lag for the EMG profile of each gait cycle in the window against the 20 gait cycle average was calculated. The segment with the highest average cross-correlation was taken for data analysis. This was done to limit the effects of fatigue and motor exploration on the variability of our data [19]. Then, the EMG and gait performance were evaluated across the 20 gait cycles in each evaluation session.

To evaluate the effects of the phase-based feedback system on participants’ gait and EMG activation strategy, we used the following metrics. To quantify the EMG activation strategy, we determined the timing and magnitude of the peak filtered EMG signal. Since the feedback aimed to alter the prosthetic push-off timing and magnitude, we quantified stance duration and ankle power of both prosthetic and intact limbs. Stance duration was defined as the duration during each gait cycle in which the vertical GRF was > 50 N. Additionally, we defined a stance duration symmetry ratio metric, defined as the stance duration of the prosthetic side divided by the stance duration of the intact side. Prosthetic ankle power was calculated as the dot product of the ankle moment and the ankle’s angular velocity normalized by participant body mass (W/Kg). The peak prosthetic ankle power was used to quantify the joint power at push-off.

Finally, the step-to-step transition work was calculated using the individual limb method described in Donelan et al. [34] in order to understand how the haptic feedback impacted body-level kinetic measurements. Here, the step-to-step transition period was defined as the double-support phase during each gait cycle when both force places recorded > 50 N of force. During this period, we calculated the contribution of each leg to the center of mass (COM) power as the dot product between GRFs for each leg and the COM velocity [34]. We calculated the positive work done by the prosthetic leg when it was trailing and the negative work done by the intact leg when it was leading by integrating the center-of-mass power during the prosthetic to intact leg step-to-step transition [34]. Step-to-step transition work was normalized by body mass and reported in units of Joules per kilogram (J/Kg).

III. Results

A. Change of rGAS EMG Activity in Walking

Fig. 2 shows the normalized rGAS EMG profile, averaged across 20 gait cycles in each evaluation session for each participant. The 20 gait cycles were taken from the final trial for each condition (2nd BF trial, 4th WF trial, and 2nd AF trial). On average, the chosen gait cycles began at 43.5 ± 25.7s and ended at 72.4 ± 27.7s through the 2 minute trial. Peak EMG timing and amplitudes for individual participants are summarized in Table II. Fig. 2 indicates not only did participants change their muscle activation pattern once feedback was provided, but also that they maintained the new pattern after feedback was removed. However, the change in muscle activation varied among the three participants.

Fig. 2.

EMG profile in a gait cycle from the rGAS of each participant for each feedback condition, where the mean EMG signal magnitude across 20 continuous gait cycle is shown as a solid line, and the standard deviation shown as the shaded area around the line. The two vertical black lines bound the target region when participants were instructed to contract their rGAS muscle for prosthetic ankle control.

Table II.

Mean values and standard deviations for EMG outcome measures

	Peak EMG Time (% Gait Cycle)			Peak EMG Amplitude (% MVC)
Condition	BF	WF	AF	BF	WF	AF
P01	55.55 ± 3.44%	56.69 ± 4.36%	55.91 ± 3.47%	44.48 ± 8.48%	64.18 ± 10.92%	54.26 ± 6.17%
P02	47.25 ± 10.02%	51.27 ± 4.25%	52.79 ± 3.12%	36.53 ± 7.55%	34.76 ± 5.58%	43.18 ± 11.05%
P03	21.94 ± 14.04%	46.97 ± 8.19%	46.43 ± 7.35%	68.32 ± 10.37%	91.95 ± 32.35%	79.94 ± 23.36%

P01 did not shift the peak EMG contraction timing because the peak EMG signal was already within desired push-off timing in the BF condition. However, once feedback was applied, P01 increased their peak EMG amplitude by 19.7% in the WF condition and 9.78% in the AF condition compared to the BF condition.

P02 did not change the peak EMG magnitude between the BF and WF conditions, rather they reduced the activation of rGAS during early- and mid-stance. Peak EMG magnitude did, however, increase in the AF condition. In addition, P02 shifted their mean peak EMG timing to be within our target range of 56 ± 5% of the gait cycle.

P03 showed notable EMG activity change with reduced activation in early- and mid-stance and increased magnitude during push-off. Peak EMG amplitude increased by 23.63% in the WF condition and 11.62% in the AF condition compared to the BF condition. In addition, for P03, the peak timing shifted towards the target push-off time, occurring 25.03% later in the gait cycle in the WF condition compared to the BF condition.

B. Changes in Gait Temporal Pattern

Fig. 3 shows the stance duration for each participant across the three conditions. After feedback was applied all participants increased stance duration on both the prosthetic and intact limbs.

Fig. 3.

Boxplots illustrating individual results for the contact time for both intact and prosthetic limbs for each gait cycle across different conditions. Shaded boxes show contact time for the intact limb while unshaded boxes show contact time for the prosthetic limb. The central bar indicates the median, the black diamond indicates the mean, the edges of the box represent the 25th and 75th percentiles, whiskers extend to the minimum and maximum values of the data, excluding data points which extend beyond 1.5 × the inter-quartile range, which are represented as red dots.

P02 and P03 walked with more symmetric stance duration between their legs in the WF condition compared to the BF condition. In the BF condition, the stance duration symmetry ratios were 0.96 ± 0.04 and 0.89 ± 0.02 for P02 and P03, respectively. In the WF condition, the symmetry ratios were 0.99 ± 0.04 and 0.93 ± 0.03 for P02 and P03, respectively. P01 did not notably change stance duration symmetry across any conditions.

C. Changes in Gait Kinetics

The peak prosthetic ankle push-off power during stance is shown in Fig. 4. Two of the three participants, P01 & P02, increased the amount of joint power produced by the robotic ankle after feedback was introduced (i.e., BF vs. AF). In the AF condition compared to BF, ankle power increased by 1.11 W/Kg and 1.17 W/Kg for P01 and P02, respectively. P03 did not follow this trend; their peak ankle push-off power did not see any notable change across the 3 conditions.

Fig. 4.

Boxplots illustrating individual results for the peak power from the robotic ankle across the before feedback (BF), with feedback (WF) and after feedback (AF) conditions.

Fig. 5 shows how the step-to-step transition work changes for each participant across conditions. All participants increased the mean positive step-to-step transition work for the trailing prosthetic limb. Compared to the BF condition, positive trailing limb work increased by 0.027 J/Kg, 0.064 J/Kg, and 0.059 J/Kg in the AF condition for P01, P02, and P03, respectively.

Fig. 5.

Boxplots illustrating the step-to-step transition work done by both limbs when the intact leg was trailing and the prosthetic leg was leading. These results show the positive work done by the prosthetic limb and the negative work done by the intact limb during these transitions during the before feedback (BF), with feedback (WF) and after feedback (AF) conditions. The solid dark horizontal lines represent the estimated transition work for an individual without amputation weighing 75 kg walking at 0.75 m/s [35].

No such consistent trend was found between participants for the leading intact limb work, with P01 increasing the negative work magnitude, P03 decreasing the negative work magnitude, and P02 showing no notable change. This indicates that our feedback system was unable to change the work done by the intact limb in a consistent way.

IV. Discussion

In this study, we developed a novel, continuous phase-based haptic system meant to train individuals with transtibial amputations to produce appropriate rGAS activation strategies to operate a dEMG-controlled robotic ankle for efficient push-off function in dynamic walking. We adopted a high-density haptic vest that delivers haptic sensation via an array of vibrotactile motors. We developed a novel encoder that mapped continuous gait phase to motor vibration patterns to provide (1) cues for the prosthesis user to contract their rGAS in each gait cycle, and (2) confirmation feedback on whether the desired rGAS activity was produced (i.e., knowledge of results to reinforce motor learning). This system was designed to be lightweight and portable, and could be used as a training tool both in and outside of a clinical setting.

A. Training Effects

Existing literature on dEMG-control for lower limb prosthetics has shown that without some source of feedback of the EMG control signals, such as visual feedback [19] or proprioceptive sensation [14], amputee users struggle to produce appropriate residual muscle activation patterns to control the robotic prosthesis mechanics during locomotion. However, whether the effects of augmented sensory feedback are retained after the feedback is removed has not been explored. In locomotion for individuals without amputation, automaticity allows humans to walk without relying on continuous sensory cues. Ideally, amputee users should also be able to use this feedforward neural control to operate the prosthesis, coordinated with the rest of body. Hence, in this preliminary validation, we examined the effects of haptic feedback training after feedback was removed. We showed that participants maintained the learned EMG activation pattern and stance duration and were capable of walking with increased step-to-step transition push-off work even after feedback was removed.

Haptic feedback as a training tool for dEMG-control has been studied in upper limb prosthetics [36], and has reported similar results to those presented here. De Nunzio et al. showed that using tactile feedback to train grasping force helped participants develop an internal model for feedforward prosthesis control, and that this model persisted once feedback was removed [28]. They also found that at higher grasping forces, requiring larger EMG amplitudes, accuracy gradually decreased once feedback was removed. Our study found similar results using haptic feedback for lower-limb dEMG training. The ability of the participants to maintain the altered push-off timing implies that an internal model may be trained while walking with feedback that persisted after the feedback was removed. Another potential explanation may be related to entrainment, in which participants became synchronized with the periodic stimuli from the vest. Other studies using periodic haptic stimuli for gait rehabilitation have observed how these entrainment effects can persist even after the stimuli is removed [37]. However, in our study the effects of the feedback system and its after effects were all tested within the same experimental visit. It would be interesting in future work to apply the feedback system over a longer period of time, possibly even allowing participants to take the system home and use it in daily life.

B. Gait and EMG Activation Strategies

While all participants began with different baseline EMG activation strategies, once feedback was provided all participants notably changed their strategies to follow the guidance of the feedback. In doing so our feedback system trained participants how to control the timing and magnitude of the torque applied by the robotic device during push-off. It has been shown in autonomous robotic ankles that applying torque between 52% and 56% of the gait cycle can significantly improve metabolic efficiency compared to walking with a passive device [9]. After applying our feedback both P01 and P02 got within this range for their peak EMG timing, and P03 was notably closer to this range than they were before feedback was applied.

P02 and P03 shifted their peak EMG timing in the WF and AF conditions compared to the BF condition which altered their gait. P02 and P03 increased the symmetry of their stance duration, with P02 spending an almost equal amount of time on the prosthetic limb compared to the intact side. Amputees often walk asymmetrically, spending more time on their intact limb. This asymmetry has been reported to be a major determinant of overall physical function and mobility, with more symmetric gaits linked to higher mobility [38]. By using the feedback system to change push-off timing with a robotic prosthesis, we saw these asymmetries reduced for all participants besides P01, who already walked with our target push-off time before feedback was provided. This further supports that this system could be a powerful training tool for encouraging amputees to rely more on their robotic prostheses and less on their intact limbs.

P01 and P03 increased their EMG magnitude in the WF and AF conditions compared to the BF condition. The increased EMG magnitude was related to an increase in peak prosthetic ankle power for P01. However, P03 did not increase peak prosthetic ankle power in the WF and AF conditions even though their EMG magnitude increased. This is likely due to the ankle being calibrated to produce a maximum torque between 40–50% of MVC. Since P03 had a higher initial EMG amplitude than other participants in the BF condition, they likely maxed out the device’s torque and did not increase peak ankle power even though they increased their peak EMG magnitude in the AF condition (Fig 2).

C. Step-to-step Transition Work

We confirmed the feedback system was consistently able to guide participants to alter their gait and EMG activation strategy to increase the prosthetic side positive work during step-to-step transitions to be closer to levels expected from individuals without amputation (Fig. 5). Previous studies using biomechanical models have demonstrated how a decreased prosthetic side push-off work is related to greater energy dissipation on the intact side [34], with this being confirmed in experimental results [39]. Thus, an increase in prosthetic side push-off work may be related to a decrease in intact side collision work during the step-to-step transition which has been associated with beneficial outcomes such as improved metabolic efficiency [2], [40] and reduced mechanical loading of the intact leg [41]. However, in this study, no decrease in intact side work was found even though prosthetic side mechanical work increased, which may mitigate potential beneficial outcomes.

The observation that increased prosthetic leg work did not correspond to reduced intact side work during the step-to-step transition may be explained by a constrained walking environment. The fixed length and speed of the treadmill may have constrained the possible gait strategies available to participants. For example, the pneumatic ankle torque provides positive work to the human-robot system which could increase gait speed. However, as the length of the treadmill and speed are confined, the user may be required to dissipate the increased energy from the robot during the intact limb collision to maintain a constant walking speed on the treadmill.

D. Limitations

The feasibility study reported in this manuscript was preliminary. A future study with a larger participant pool is needed to confirm the effectiveness of the biofeedback and to further explore inter-subject variations. One potential limitation of this work is the lack of a control group to account for the effects of familiarization with the device. However, prior literature has demonstrated that even after 2–6 hours of familiarization with a dEMG-controlled prosthesis, PWTAs still walked with peak ankle power similar to a passive device [14], [18], [19]. Thus, we expect that the effects observed in this study are due to the feedback system and not familiarization.

As discussed, treadmill walking is another potential limitation. One of the main advantages of implementing the continuous phase variable estimator is that it could work at any speed walking both on the treadmill or overground walking. Due to current hardware limitations we could only test the feedback strategy and phase estimator on the treadmill, but future work should focus on developing a more portable system for overground walking.

V. Conclusion

This study aimed to develop a haptic feedback system to train PWTAs how to activate their residual muscle to neurally control a robotic ankle prosthesis while walking. We focused on providing instructional feedback to cue the appropriate rGAS contraction timing in order to enhance push-off power for human-prosthesis coordinated walking. This was achieved by a combination of both continuous phase-based feedback to instruct participants on when to push-off with the device, and corrective feedback to inform participants whether their dEMG control was successful. This feedback system was pre-liminarily validated on 3 PWTAs, all of whom adjusted their residual muscle control strategy and their stance duration of both limbs, with these changes persisting even after feedback was removed. Through these changes all participants increased the positive step-to-step work from the trailing prosthetic limb. These results indicate our novel haptic system has potential to serve as a powerful training tool for PWAs in dEMG control of robotic ankle prostheses, improving their ability to coordinate with the robotic device for efficient walking.