Background
For individuals living with upper limb loss or difference, myoelectric prostheses have the potential to restore lost functionality and improve independence. Significant advancements have been made in myoelectric control methods, but sensory feedback is still a missing component from commercial prostheses. Sensory feedback is one of the most commonly requested features of state-of-the-art prostheses [
1], and is critical to able-bodied limb control [
2]. Consequently, artificial sensory feedback has received much attention over the past decade [
3,
4]. Typically, this takes the form of sensory substitution feedback, where the information provided from missing sensory organs is communicated to the user via an alternative method such as vibrotactile [
5‐
10] or auditory stimuli [
11‐
13], or via direct nerve stimulation [
14‐
16].
Despite this attention, artificial sensory feedback has not yet achieved commercial availability for prostheses, which may be related in part to the experimental conditions in which these systems are tested. Frequently, artificial feedback is tested with subjects blindfolded and not connected to the prosthesis. Although these studies consistently show the benefit of sensory feedback, they omit the incidental sources of feedback that prosthesis users rely on every day, such as vision, sound, and prosthesis vibration. This incidental feedback often serves the same purpose as the artificial feedback being tested (i.e. informing the user about the state of the prosthesis), and studies have shown this incidental feedback is sufficient for some tasks [
17]. Therefore, when artificial feedback is tested
alongside incidental feedback, results become inconsistent—some studies suggest discernable benefits of artificial feedback alongside incidental feedback, such as improved time to target prosthesis position [
18], ability to perform object manipulation tasks [
19], and coordination of grasping with the prosthesis [
20], however the same and other studies also show no changes in other aspects of prosthesis use [
9,
19,
21‐
24].
One theory explaining this discrepancy stems from the degree of precision of each feedback source. When we receive the same information from multiple sources, we merge them in accordance with their uncertainty: sources with less uncertainty are favored over those with greater uncertainty [
25,
26]. Therefore, if incidental feedback (particularly vision) is more precise than the artificial feedback being tested, then the tested feedback may not meaningfully improve the users understanding of their prosthesis movements.
One candidate for sensory feedback which is not well estimated by incidental vision is kinesthesia. Prior work has suggested that limb speed, and in particular joint speed, has high visual uncertainty and can be supplemented with audio feedback to greatly reduce this uncertainty [
27]. Knowledge of limb speed may aid in the formation of internal models of biological and prosthetic limb movements, and a previous study has indeed suggested that supplemental joint speed feedback may improve reaching accuracy during instances of perturbed myoelectric control [
28]. However, the benefits of such feedback requires additional investigation within the context of impaired proprioception following amputation. Proprioceptive organs including muscle spindles and Golgi tendon organs are activated differently in an amputated limb than they are in intact limb; agonist–antagonist muscles pairs stimulate these organs during movement [
29], but this pairing is generally absent from amputated limbs. Instead, standard surgical procedure for upper-limb amputations involves myodesis of the muscles to the end of the distal bone, preventing the normal passive stretching of antagonist muscles during movement and negatively affecting proprioception [
30].
The purpose of this study was to investigate the effect of joint speed feedback on prosthesis control and adaptation to errors during reaching. Transradial amputee subjects controlled a virtual 1-DoF myoelectric limb and completed center-out reaching tasks under steady-state and perturbed dynamics conditions. We quantified control by measuring trial-by-trial adaptation to self-generated and perturbation-generated errors to learn how quickly myoelectric control users can update their understanding of the dynamics and adjust accordingly.
Discussion
This study expanded upon our previous work by investigating transradial amputee performance during center-out reaching tasks. These tasks require coordination of elbow angle and wrist EMG to complete the reach. This paradigm differed from other similar studies into proprioceptive feedback by using a ballistic reach paradigm, which prevented subjects from incorporating feedback into their task performance and allowed us to investigate solely the impact of feedback on improvements to feedforward control [
10,
13,
41,
42]. Our results provide some insight into how artificial joint speed feedback may be used to improve control of a myoelectric prosthesis. We found evidence that subjects were able to reduce their average reaching errors when provided audio feedback encoding the joint speed of a myoelectric limb (Fig.
3). We also found evidence suggesting the feedback may help prosthesis users reduce compensatory movement bias (Fig.
4a). However, no significant differences were found between feedback conditions for adaptation behavior after abrupt perturbations to the controller (Fig.
6).
In some aspects, our results agree with those from our previous study with non-amputee participants. Transradial amputee participants were able to complete ballistic center-out reaches requiring simultaneous control of positional- and myoelectric joints, in a manner similar to the how they may use their prosthesis in a home environment. Additionally, the same compensatory behavior was observed in both studies, where subjects would strategically under-reach with the wrist and compensate by overreaching with the elbow to minimize the distance to the target. This manner of compensatory movement is common for upper-limb prostheses (though normally demonstrated for the trunk and shoulder [
43]), and aligns with the expected optimal reaching strategy to minimize endpoint error—given that only one limb configuration can reach each target, any wrist error can be optimally compensated with an opposing elbow adjustment of half the magnitude. It should be noted that, if reaching tasks required wrist flexion instead of wrist extension, the optimal reaching strategy to compensate for excessive flexion would be to overreach with the elbow. We also observe a more positive wrist error when reaching towards Target 3 (involving simultaneous extension of both the wrist and the elbow) than towards other targets, though interestingly the endpoint errors were lowest.
2 This tendency may suggest that fine control of wrist extension is more difficult when coupled with simultaneous elbow extension. Alternatively, it may suggest that visual estimation of the requirement movement to achieve the target is more difficult. Interestingly, while we showed no impact of sensory feedback on the average errors in the previous study[
28], amputee reaches in the present study demonstrated lower elbow bias, and a trend towards lower wrist bias, with feedback available (Fig.
4a).
The present study differs from the previous study with respect to steady-state errors; while no significant differences were observed in endpoint, elbow, or wrist errors for non-amputee reaches, transradial amputee endpoint and wrist angle errors were significantly improved with joint speed feedback. Furthermore, the stochastic analysis reveals an interesting difference between non-amputee and transradial amputee reaches: while elbow control noise is roughly equivalent between populations, the control noise of the myoelectric wrist can be more than twice as high for transradial amputees compared to non-amputees [
28] (Fig.
5b).
However, where non-amputees demonstrated improved reaching errors after adapting to perturbations while reaching towards changing targets, transradial amputees showed no significant differences in perturbation adaptation behavior. One possible explanation for these inconclusive results stems from the heightened control noise. With myoelectric control noise for transradial amputees nearly double that of non-amputees, likely due to lack of or damage to proprioceptive organs, it is possible that this increased control noise led to increased internal model uncertainty, decreasing the capacity to adapt to perturbations. These trends may extend to adaptation behavior after control system perturbation. It should be noted that no individuals with congenital limb difference were included in this study; we expect control noise to be between those of transradial and non-amputees due to the more natural insertion of residual muscles, however this remains to be investigated.
The absence of incidental feedback may also contribute to the increased control noise observed in this study. When controlling a myoelectric prosthesis, users will generally rely on cues such as the sound and vibration of the hand as an indirect indicator of speed; the presence of these incidental cues in daily life may reduce control noise and consequently give rise to differences in adaptation behavior. In contrast, the auditory feedback modality used in this study provides a “best-case” scenario for low feedback uncertainty [
27]; consequently, one might expect increased control noise when using vibrotactile or electrotactile feedback modalities with higher sensory uncertainty. This may even be the case when discrete vibrotactile feedback is fused with auditory feedback, as was seen in a recent study by Engels et al. [
13].
Analyses in our current study were limited by the analysis methods available and the data collected for each. Our protocol required subjects to reach for several targets arranged throughout the reaching space, which ensured reaching performance was not localized to any one particular region. However, this also required splitting up reaches into smaller blocks of consistent reaches to prevent subject fatigue. As a result, adaptation models for self-generated errors were fit on relatively small amounts of data; this was especially the case for the stochastic signal processing analysis. Furthermore, this analysis requires a stationary target, thus reaches towards changing targets had to be omitted from this analysis. Analyzing self-generated error adaptation using two different methods allowed us to partially account for the limited data and build a fuller picture of adaptation behavior at steady-state.
The hierarchical model used our previous study requires sufficient data to fit all parameters across all included perturbation conditions [
28]. Although the intent was to use the same model in this study, the smaller number of subjects prevented this model from converging. Furthermore, constraining the model parameters using insights from steady-state errors did not alleviate issues with model convergence [
44]. In its place, we took an approach previously used in our pilot study [
38]. In this approach, an individual exponential decay model is fit to each subject, for each condition. The coefficients from these models were then analyzed using a linear mixed effects model. To supplement this analysis, post-hoc comparisons were made on the initial and final errors achieved during perturbation trials. However, no significant differences were found during perturbation trials, whereas differences were found for non-amputee reaches.
The outcomes from the stochastic signal processing techniques also warrant additional attention. The non-improvement in the adaptation rate of the EMG-controlled wrist internal model is opposite of what is expected from reduced wrist noise. A possible explanation is that the high EMG control noise for transradial amputees, more than double than that of non-amputees at times, was more substantial than effects of volitional adaptation, which may have influenced the internal model adaptation rate as calculated using analytical methods [
33]. It should be reiterated this analysis was conducted on relatively small amounts of data, which may disproportionately affect the variability or biases of calculated internal model adaptation rate.
The findings in this study corroborate those in a recent study on the clinical relevance of artificial feedback [
19]. They conclude that the benefit of sensory feedback depends on the complexity of the task and the proficiency of the feedforward control. Our study involves a simple task—center-out reaching—made complicated by the control scheme. Our pilot study with trans-humeral amputees used a more difficult control scheme, and the high control noise made control (and adaptation) difficult [
38] However, in our present experiment with trans-radial amputees, we show that improved feedback can reduce the control noise, thereby improving feedforward control [
34]. This outcome suggests a need to test artificial sensory feedback systems with amputee patients of different levels to determine how beneficial feedback is to each population. Developing a more complete understanding of which factors determine the degree of benefit for prosthesis feedback can help researchers develop clinically impactful artificial sensory feedback which improves quality of life for people with amputations.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.