Sensory Feedback for Upper-Limb Prostheses: Opportunities and Barriers

— The addition of sensory feedback to upper-limb prostheses has been shown to improve control, increase embodiment, and reduce phantom limb pain. However, most commercial prostheses do not incorporate sensory feedback due to several factors. This paper focuses on the major challenges of a lack of deep understanding of user needs, the unavailability of tailored, realistic outcome measures and the segregation between research on control and sensory feedback. The use of methods such as the Person-Based Approach and co-creation can improve the design and testing process. Stronger collaboration between researchers can integrate different prostheses research areas to accelerate the translation process. 1

There have been recent reviews that focus on the technological advances in sensory feedback systems (implanted devices in particular) [8], [9] and the impact of sensory feedback on control [10]. Thus, this paper aims to highlight specific barriers and opportunities drawn from the literature concerning the impactful incorporation of sensory feedback in upper-limb prostheses. We define impactful implementation of sensory feedback as one that ultimately leads to an improved quality of life. This can be achieved through increased user satisfaction and prosthesis use and, thus, reduced abandonment rates and overuse injuries. The paper will start with presenting the literature on user needs (section II), followed by a brief overview of the clinical results obtained with the different feedback methods, with a focus on conflicting or unclear results (section III). Section IV will then discuss the challenges and opportunities associated with producing translatable sensory feedback research that matches the user needs and results in effective clinical implementation.

II. USER NEEDS
Many recent surveys have focussed on trying to understand the needs of individuals with upper-limb difference, many of which focused on device abandonment, and while reported rates vary there is broad agreement that current prostheses need improvement. The reported abandonment rates of passive, body-powered and myoelectric prostheses range between 6 and 100%, 80 and 87% and 0 and 75%, respectively. The variation in the reported rates can be linked to the different recruitment methods and the intrinsic differences among the population [11], [12].
The Atkins survey reported in 1996 is one of the largest conducted, representing the views of 1,575 participants. The survey summarises the problems faced by the users of both body-powered and electric-powered prosthesis, and defines the areas of improvement, split into near-term and long-term. The near-term were focused on improving reliability and comfort, with the cables and harness being key body-powered prostheses and batteries and electrodes for electric-powered. Glove material was also a common concern. The proposed future work was similar for both, focusing on wrist movement and reduced visual attention [11].
Twenty-five years later and the challenges facing prostheses users do not seem to have changed; comfort and function are still considered innovation priorities. The reasons for abandonment are linked to those priorities and include temperature, weight, pain, poor fit, difficulty of control, slow response speed, lack of durability and functionality as well as lack of sensory feedback [12]. What is particularly surprising is that glove durability and reliability of electrodes, what Atkins et al. described as "near-term considerations", remain a source of frustration among upper-limb prosthesis users [12]- [20].

A. The Need for Sensory Feedback
Varying survey methodologies and demographics have led to conflicting results about the importance of sensory feedback. The need for sensory feedback is often considered more important for myoelectric than body powered prostheses. The Atkins survey showed that sensory feedback (described as reduction in visual attention required) was given an average priority of 3 rd and 5 th out of 10 by myoelectric and bodypowered prostheses users, respectively, with a lower number indicating greater preference [11]. Similarly, Biddis et al. showed that, out of 10 options, sensory feedback was ranked as the 4th most important by myoelectric users, but only 8th by body-powered users [13]. This difference in perceived importance can be linked to the ability of the mechanical nature of body-powered prostheses to provide some level of sensory feedback which improves the control of the device.
Biddis et al. also observed that sensory feedback is considered more important by rejectors than wearers, with 85% of rejectors and 44% of wearers ranking it as a significantly important feature [15]. Whereas this could be a result of many different factors, surveys have showed that users who spend more time learning to use their prostheses are less likely to reject it [21]. This could suggest that the users have learnt to compensate for the loss of sensory feedback through methods such as pressure on the stump or the sound of the motor [22]. The survey by Pylatiuk et al. found the highest rate of support for sensory feedback at over 90% [23].
Surveys that use open-ended questions surrounding imagined sensory feedback often report participants finding it difficult to imagine. Having not tried any form of sensory feedback, they tend to answer with "not really sure" or "literally have no clue" [17]. However, more guided questions such as those proposed by Lewis et al. enabled the most important sensations to be identified by asking the respondents to rate their importance. Grip force was identified as the most important, followed by movement, position, first contact, end of contact and touch. In another question, the participants were asked about the preferred feedback modality. This was done after prompting the respondents to self-assess the sensitivity of their stump to pressure, vibration, and temperature. Surprisingly, their preferred modality was chosen to be temperature, even though the average sensitivity score for temperature was the lowest. Vibration, electric and pressure were the next preferred modalities with the acceptability of visual and acoustic feedback being markedly lower. No details were provided on how each of the modalities were described to explain why an "electric" feeling was chosen over a seemingly more natural sensation of pressure [24]. This presents the limitation of relying on surveys to understand user needs, particularly when using speculative questions. Moreover, surveys provide a limited insight into why participants selected their specific answers. For example, Lewis et al. showed that grip force is the most important sensation amongst the options given. However, participants did not express if that is important to enable dexterous manipulation or simply to hold objects without dropping them. Further engagement with users through interviews will enable researchers to understand how sensory feedback is useful in a practical sense and thus where to focus research and development to maximise impact.

B. Acceptance of Invasive Approaches
With many of the sensory feedback approaches depending on invasive devices (i.e., surgical implants), it is important to understand potential user acceptance. Engdahl et al. found that participants were most interested in myoelectric control, followed by peripheral nerve interfaces (PNI) and targeted muscle innervation (TMR) and least interested in cortical interfaces (CI). Engdahl et al. also reported the comments received on each of the different methods which showed that, based on prior negative experience of reliability, users are hesitant to try new technology, especially if some of the parts that may need repair are implanted. Furthermore, most of the respondents prefer not to undergo surgery again and are more interested in having reliable basic features than advanced ones [25].

III. OVERVIEW OF RESULTS TO DATE
Section II has demonstrated that there is limited literature on the need and efficacy of sensory feedback. Improvement in control of the prosthesis is a primary metric used to assess sensory feedback, using both existing and bespoke outcome measures (see Table 1). Other metrics include the naturalness of the sensation (in terms of being somatotopic and homologous), as well as the effect on embodiment and phantom limb pain. Emotional benefits, from direct quotes, may also be considered if available. This section will highlight the range of outcome measures available, identify if results are conflicting or in agreement, and identify challenges and opportunities. The reader can refer to recent review papers for more details on different systems' design and functionality [6], [9], [10] and to [26] for a historical perspective.
Starting with recent review papers and branching out to cover new and related research within the period 2010-2020, the approaches with consistent or increasing interest in the literature were identified, as shown in fig 1. Vibrotactile and electrotactile feedback are common non-invasive methods that rely on stimulating the mechanoreceptors and afferent nerve endings in the skin to relay different pressures and/or hand positions through spatial, frequency or intensity mapping [27]. Invasive techniques include electrical stimulation of the peripheral nerves using implanted electrodes to elicit action potentials (mimicking those generated by mechanoreceptors), and thus referred sensation. Interfaces at the Dorsal Root Ganglia (DRG) ensure that only afferent fibres are excited [28], whereas CI enable the selection of the stimulation location based on the architecture of the somatosensory cortex [29]. Both DRG interfaces and CI are still in their early development phase, with a limited number of clinical trials, and will not be covered in this review [30]. Extended Physiological Perception (EPP), first proposed in 1974, is an alternative feedback approach that utilises the body's inherent proprioceptive abilities by creating a mechanical link between the prostheses joint and body joint. The physiological appropriateness of EPP is thought to enhance its intuitiveness, as subconscious pathways are used to process the feedback [31]. EPP has been recently extended in an approach that suggests replacing the mechanical link to the prosthesis with implanted devices connected to the muscles to apply the appropriate forces [32].

A. Improving Control of Prostheses
Research on sensory feedback for upper-limb prostheses gained momentum after the commercialisation of powered prostheses [34]. Unlike their body-powered counterparts, powered (myoelectric) prostheses are controlled using electromyographic (EMG) signals, reducing the user's ability to assess the force applied. This, combined with the inherent uncertainty of measured EMG signals, are the main drivers for implementing sensory feedback as it allows the user to adjust prediction errors, thus facilitating learning and ultimately improving control [35], [36].
Humans rely on sensory inputs to perform even the simplest actions. Consider drinking water from a cup as an example; the brain uses vision, tactile inputs from the fingers and proprioception to enable us to drink without dropping the cup or spilling. The brain weighs those inputs and noisy and delayed sensory signals tend to receive lower weightings as they are considered less trustworthy [10] [37]. Over time, the brain builds an internal model that enables more efficient use of the different sensory inputs. Considering the availability of rapid implicit feedback (such as vision, the sound of motors or actuators and pressure on the socket), sensory feedback must operate with high reliability and minimal latency if the brain is to integrate the additional sensory inputs into the internal model [10]. Moreover, the uncertainty of the controller also influences the measured improvement; ideal controllers (only available in virtual environments) might underestimate the importance of feedback as the user can rely on pure feedforward control with practice. The presence of uncertainty increases the reliance on sensory feedback [38], [39]. However, highly uncertain Overview of effect of feedback on control. The plot compares the results obtained using different types of tasks (x-axis) when carried out with and without other sensory (vision and/or auditory) cues being available to the participant (y-axis). The radius of the bubbles represents the number of studies, the colour represents the results of those studies, and the outline represents the type of sensory feedback used. Tactile and proprioception feedback were not differentiated in the classification. The results were classified based on the best outcome (e.g., increased accuracy at the expense of time and improved performance limited to lower force levels were considered an improvement). The studies included in each of the bubbles are as follows. A1: [73], [82], [123], C1: [65] C2: [60], [66], [71], [124]- [126],D1: [38], [82], [123], D2: [38], [124], [126], [127], E2: [73], [123], [124], [127], [128]. Note: Multiple experiments within one paper are represented as different studies. This is an overview of the results to highlight the variation in results obtained, for a more comprehensive analysis the reader can refer to [10].
controllers (e.g., high forces using EMG control) reduce the benefit from feedback as adjustments of the control signal do not always translate into adjustments in the resulting action [40], [41].
It is also important to consider the type of tasks used to measure improvement, they should represent ADL and capture the benefit of sensory feedback. Simple tasks (e.g. level reaching) can underestimate the benefit, as participants are more likely to depend on feedforward control, and complex tasks may overestimate the benefit during ADLs (e.g. picking a cherry stem) because, until a sufficiently reliable and dexterous prosthesis is developed, the user is likely to handle the more delicate part of the task using the intact limb. Fig. 2 shows results obtained using different sensory feedback systems split based on the type of tasks assessed (Supplementary material, Table 1) and the availability of other sensory cues. The figure shows that most studies confirm that sensory feedback improves control. However, those representing more realistic assessments tend to show little or no improvement. While this does not necessarily mean that sensory feedback does not improve control, it does highlight room for improvement in the type of assessment used. This will be discussed further in the challenges and opportunities section.

B. Providing Somatotopic and Homologous Feedback
Producing homologous (of the same type as the original) and somatotopic (in the same location) sensation would serve to increase the intuitiveness of a sensory feedback system [42].
The deafferented cortex, initially connected to the arm, can be activated through two pathways: hand maps and the original nerve structure. Those pathways can be utilised to elicit sensations that are referred to the phantom hand. Hand maps are skin areas where tactile stimulation is perceived as originating from the phantom limb. They are highly variable between participants and are usually found on the stump and face [43]. Although time-consuming to locate, hand-maps can enable perception of specific fingers with reduced stimulation thresholds [44]- [46]. The evoked sensations are stable in terms of stimulation threshold and location for at least 11 months [47].
Stimulation of the peripheral nerves (Ulnar, Median and Radial) using implanted electrodes has been shown to, in most cases, produce referred sensations matching the innervation areas even 20 years post-amputation. The stability of innervation areas confirms that the original somatosensory pathways remain intact following amputation, despite the apparent reorganisation present as hand maps [48], [49]. Congenital amputees do not report phantom sensations as the required somatosensory pathways have never developed [50].
It is also possible to elicit referred sensation non-invasively via transcutaneous stimulation of the peripheral nerves [42], [51]- [53]. However, a disadvantage of this non-invasive approach is the poor stability of the somatotopic sensation with limb movement and persistent sensation under the electrodes [52]. Methods to reduce the sensation under the electrodes include interferential stimulation and channel-hopping [54], [55]. Moreover, this concept does not seem to work for digital amputees due to the stump's different cortical representation [56].
Studies have shown that delivering more natural and pleasant (not painful or uncomfortable) sensations during electrical stimulation results in stronger embodiment, although a natural sense of touch has yet to be achieved [57] [58]. Experiments performed in the 1970s using implanted electrodes resulted in unnatural sensations such as paraesthesia and throbbing [59]. Since then, new encoding methods have enabled more natural sensations to be achieved using peripheral nerve interfaces. Fig.  3 compares the studies that describe the type of evoked sensation and indicates the invasiveness of the electrodes (xaxis) and the chosen encoding method (colour). Tactile sensing is encoded through spatial and temporal activation of afferent fibres. Therefore, the high-order processing of the different sensory inputs controls the sensory modality experienced [60]. The low selectivity of transcutaneous stimulation limits its ability to elicit natural sensation with sensations of paraesthesia or vibration reported [61]- [63]. In fact, vibrotactile stimulation appears to be to be more comfortable [64]. Implanted electrodes enable higher selectivity and, therefore, the stimulation waveform parameters affect the sensory modality experienced [65]. Thus the fine-tuning of the waveforms through biomimetic approaches can produce the most natural sensations [58], [66]- [69]. Duration of use also influences the naturalness of the elicited sensation, with prolonged use being associated with improved quality of sensation [70]- [73]. Invasive approaches tend to elicit a range of sensation qualities and this figure focuses more of the best quality elicited. The colour coding indicates the different encoding methods [65], [66], [70], [71], [73], [87], [124], [129]- [133]. Small-scale modulation uses a slow sinusoidal envelope to modulate the pulse width of a 100Hz square pulse train. The lower limit of the pulse width is limited to ensure some activation of the neuronal population at all times (see [65] for details).

C. Increasing Embodiment
Embodiment relates to the feeling that the prosthesis is an extension of the body rather than a tool. It is usually measured subjectively through questionnaires that are based on the rubber hand illusion [74] and objectively through measuring the temperature of the residual limb [75] or using the phantom hand location [76]. Communicative Hand Gestures have also been suggested as an implicit measure of embodiment [77]. Experiments that consider the effect of sensory feedback on embodiment tend to report an improvement [75], [78], [79] even with modality mismatched feedback [80]. However, this increased embodiment is higher when natural sensations are delivered (biomimetic feedback) [66]. The definition of embodiment is inconsistent between reports, particularly concerning whether or not agency intrinsic within embodiment. Middleton and Ortiz-Catalan warn against relying on lab tests for an assessment of embodiment: "We must be careful not to extrapolate a sense of ownership and agency (or both) that occurs in a cultivated moment or instant to an irreversible, sustained phenomenon" [81].
Home use studies provide an avenue for a more complete assessment of the benefit of sensory feedback on embodiment, and user experience as a whole. Studies with both implanted and TSR feedback methods have shown that home use increases the measured benefits of the feedback. Graczyk et al. reported that while in-lab functional tests did show increased embodiment, only home use resulted in significant improvement that was most evident in the first month and stabilised afterwards [73], [82]. Schofield et al. showed that embodiment became more specific to certain conditions after home use. Participants that reported increased embodiment with time-delayed and non-somatotopic feedback at the start of the experiment only reported increased embodiment with timely somatotopic feedback by the last trial [83].

D. Reducing Phantom Limb Pain
Most adults with acquired amputation report phantom sensations, of which 80% are painful. When persistent, phantom limb pain has a negative effect on the quality of life and adaptation to amputation [84]. Sensory feedback has a demonstrable positive effect on phantom limb pain, as has been demonstrated using methods including the Visual Analogue scale [71], [85]- [87], the McGill pain questionnaire [87], the Trinity amputation and prosthesis experience scale [65], the present pain intensity scale [87], the neuropathic pain symptom inventory [71] and verbal scoring [88]. The reduction in phantom pain is usually linked to a change in the pain sensation to pressure/vibration sensation elicited by the stimulation [76]. However, this reduction in pain was not evident once feedback was removed (at the three-month follow-up) [87]. The reduction in phantom limb pain was sometimes described as the hand "opening up" [65], [76]. Additionally, the phantom sensation change was shown as an increase in the length of the phantom limb to align with the prostheses [73], [82], [89], [90].

E. Emotional Benefit
Sensory feedback is primarily discussed in relation to increased performance of the prosthetic, with less focus given to the emotional benefits (perceived or observed) despite their importance for psychological adjustment [91]. Given the increased rates of depressive symptomatology amongst individuals with acquired upper-limb loss [92], promoting positive coping mechanisms that elevate their self-worth and minimise their sense of isolation is vital for improving quality of life [91]. This suggests that the need for more systematic methods of assessing the emotional benefits of sensory feedback. Participants of experiments on sensory feedback tend to report such emotional benefits. One described it as "I felt what I was doing, do you understand, it was exactly as it was my own fingers. What a feeling!". Further, sensory feedback is often linked to benefits with interactions with children and loved ones. "I hope for some sort of sensory feedback for the grip function, when you touch things. When I use my hands for touching I want to avoid pinching my children by mistake". This emotional benefit seemed less relevant to individuals with congenital limb deficiencies who have never experienced the sense of touch through the missing hand [22].

IV. CHALLENGES AND OPPORTUNITIES
Despite the development of a wide range of technologies for sensory feedback, and their promising results, clinical translation is poor. The following subsections reflect on the challenges hindering translation and highlight the key opportunities to overcome them. The specific challenges associated with translation of neurostimulation technologies within the associated ethical and regulatory environments are outside the scope of this manuscript, and have been addressed in recent reviews [93]- [95].

A. Gaining a Deep Understanding of User Needs
While surveys are highly valuable and do highlight common problems, they often fail to communicate the underpinning behaviours and needs, leading to a mismatch between user expectations and reality and, potentially, device abandonment [96]. Fortunately, the challenges associated with understanding user needs in the context of healthcare and rehabilitation are well known, and many qualitative research tools (such as interviews and focus groups) can be used to provide richer data. In the context of sensory feedback, such data might involve what participants expect sensory feedback to feel like, what sensor locations they think are most useful, and what tasks they expect sensory feedback to help with. While these questions can be asked via a survey, using qualitative methods enables a more in-depth explorations of such participant expectations and motivations, providing an understanding of how sensory feedback will fit into the users' life beyond responses to preexisting questions defined by the researcher. It is this in-depth insight into the user's behaviours (i.e., how they interact with the environment and intervention) that enables the design of acceptable and engaging interventions that 'fit' into the user's life, subsequently increasing uptake and ultimately effectiveness. Indeed several qualitative studies have been conducted in relation to upper-limb difference (see [97] for meta-synthesis). There have been a few qualitative studies evaluating the benefit of sensory feedback systems but they are more summative in nature, focusing on how the developed sensory feedback systems affected the participants' lives [72], [81].
We propose using the Person Based Approach (PBA), which provides a systematic method to integrate formative qualitative studies, using guiding principles that are then modified throughout the planning, design, development and evaluation stages [98]. Formative studies, as opposed to summative, focus on understanding user needs in a broader manner rather than evaluating the performance of the developed system. The PBA relies on detailed evidence synthesis and iterative qualitative research to develop an in-depth understanding of the behaviours and lives of potential users in order to understand how they engage with the intervention [98]. It values autonomy and empathetic understanding, which are particularly important for personal devices, such as prostheses [99], [100]. This approach, with early integration of qualitative studies, captures the users' lived experience and their expectations of sensory feedback, emphasizing this without being limited by the constraints of an already-developed system. Removing these constraints ensures that the in-depth understanding of user needs remain at the heart of the process throughout the stages of intervention planning, development, optimisation, and evaluation. Furthermore, once the initial intervention has been developed, the iterative nature of the process means that the guiding principles and design features are regularly refined to reflect the users' experience trying the intervention -thus ensuring that the intervention remains as acceptable and effective as possible. The reader can refer to [101] for an overview of how the structured and systematic PBA can be applied by research teams at each research stage. A summary of this process can be found in Fig.1  of the supplementary material).

B. Meaningful Assessment Measures
The measures used to assess prostheses technology are critical to the translation of research and the derivation of demonstrable impact; it provides the evidence required to convince both investors and users of its benefit. It has been shown in the literature considered in Section 3A that sensory feedback improves control. However, tests that attempt to quantify this improvement in representative use cases tend to show insignificant improvement, if any ( fig. 2). The one case that showed improved performance was a subjective measure where the participants reported the improved performance in the tasks that they found most difficult [73]. While it is possible that this assessment reflected increased confidence rather than improved control, it resonates with Schaffalitzky et al.'s recommendations to consider functional improvements from the user's perspective rather than physical functioning per se [102]. Moreover, it highlights the importance of developing tailored ADL tests. Given that prostheses users are constantly evaluating the benefit of using a prosthesis against its problems [4], the assessment measures should aim to align to the users' personal metrics. This has significant potential implications on predicting prostheses usage, as current lab measures show no significant correlation to prostheses usage [103]. (See [104] for more information on outcome measures for upper-limb prosthesis). The influence the user input has on outcome measures can be illustrated by using the cups relocation task as an example. This task tests the user's ability to handle fragile objects (disposable cups). If user engagement highlighted that holding a cup while walking is the trickiest scenario in day-today life, the assessment could, therefore, incorporate a walking activity. This will enable the feedback system to be tested to ensure reliability while walking while also evaluating how effective sensory feedback is in such situations.
Qualitative research approaches, such as the PBA, can be the first step in developing meaningful measures as they enable researchers to capture the behaviours and unmet needs of the population of people with upper-limb difference [100]. Further involvement of a subset of participants in the design process, through co-creation, can ensure that the qualitative data collected is interpreted and utilised appropriately. Co-creation design frameworks have been used to rapidly ensure health interventions are accepted and useful for users by making stakeholders (such as the users, clinicians, industry and policymakers) part of the design process [105]. Each stakeholder provides input at various stages of the work [106]. When designing outcome measures, users and clinicians can provide valuable insight into what has worked in the past and what has not to guide the ideation process. Industry partners have more technical experience and can spot potential failures during home use, ensuring that the chosen designs are robust. Finally, including policymakers ensures that the chosen assessment measures can provide sufficient evidence of value and impact, and paves the way for them to be translated into clinical assessments.
If the assessment measures are developed alongside the sensory feedback system, then there is the potential to collect detailed and tailored usage data. This data should aim to capture the different factors discussed in section III. Comparisons between the home-use and lab-use measures could provide valuable insight into differing use patterns and could guide modifications to maximise at-home benefit. Such longitudinal home-use data would also support future clinical decision making.

C. Integrating Forward Control and Feedback
The interaction between the feedforward (control) and feedback systems is often not considered, despite it affecting both basic functionality and overall performance.
Electrical stimulation used to elicit sensation poses challenges arising from stimulation artefacts if recordings are made at the same level for the purposes of control, with the risk that control could be lost during stimulation [107]. Current closed-loop experiments using implanted interfaces rely on surface EMG recordings, rather than recordings from the implanted interface, to control the prostheses [108], [109]. This approach reduces the effect of stimulation artefacts due to both the spatial separation of the interfaces and the reduced stimulation magnitude required for direct nerve interfaces. Experiments involving surface stimulation (transcutaneous) and recording have been designed by placing the stimulating and recording electrodes on different arms [110], [111]. However, this is thought to reduce the system's intuitiveness, making it unsuitable for use outside the lab environment [110]. Several methods exist to overcome the interference between stimulating and recording electrodes, including using concentric electrodes [112], blanking [113], [114] and timedivision multiplexing [115], [116].
The design of current sensory feedback systems aims to enhance the performance obtained with standard myoelectric control. However, several new decoding algorithms are being developed on an invasive and non-invasive level [117]. With a few of those methods reaching maturity and clinical trials, it would be interesting to test how sensory feedback integration differs with such methods as well as data acquisition [118]. The current home tests being planned for different control methods can provide an opportunity for concurrent testing with sensory feedback. The resulting increased complexity of experimental design will pay off in the long run, as the two systems will be expected to work synergistically. The recent work of Marasco et al. provides a key step towards the integration of feedforward and feedback through the development of metrics that assess the contribution of each of the parts of the interface on the overall performance [119].

V. IMPORTANCE OF INTERDISCIPLINARY COLLABORATIONS
The theme arising in the previous three sections is the need for an interdisciplinary, integrated qualitative and quantitative, and long-term research approach to the design of sensory feedback systems. Established research groups may have the capabilities and expertise required to perform this integrated interdisciplinary research but it may become more difficult to maintain this capability as more advanced technologies near translation. Indeed, the breadth of expertise required may prove to be a significant barrier to many research groups, or worse still may negate any potential clinical impact due to a lack of expertise in one or more areas. The existing regulations in place to reach participants and conduct (both invasive and noninvasive) experiments are rightly robust but are themselves often a significant barrier for smaller research groups. The shift to more interdisciplinary research necessitates the development of new means of collaboration to improve research efficiency. An optimal collaboration structure would be supported by national centres of excellence supported and funded by government. The role of such a centre would be: to define outcome measures and develop strategies for implementation, to manage the engagement with device user groups and stakeholders, to shape policy and regulation, to maintain registers of expertise, and ultimately to enable interdisciplinary collaborations and co-creation for the design of long-term studies. Co-design requires goodwill and engagement from user groups, and thus must be effectively managed at a national level to ensure that the maximum benefit can be derived.
While the ultimate aim may be to produce prostheses that are as functional as the human arm, a national centre would enable the identification of grand challenges that solve existing problems faced by users while paving the path towards the ultimate aim. Those challenges can be used to guide research efforts. This centre can also facilitate the development of a shared platform for participants to register interest in different research activities outlining their background and preferences to allow for both convenience and purposive sampling.
Surveys and interviews will be designed based on input from multiple stakeholders to ensure similar areas of investigation are grouped, reducing repetition while increasing the engagement per study. Research on specific technical advancements will continue independently by different groups, with scheduled meetings providing an opportunity for the exchange of ideas and suggestions. Long term experiments, however, can and should be done collaboratively to enable different technologies and theories to be demonstrated. The development of assessment measures must be synergistic with the technologies being developed, and tools such as the PBA can be used to strengthen the links between quantitative and qualitative research.
Moreover, the collaboration of different research groups to develop shared libraries of commonly required resources eliminates the time spent replicating them and increases their reliability. This collaborative library can include software used to collect sensation information (such as [120]), the design of home use tracking devices (critical for assessment measures) as well as a shared list of lessons learnt, to name a few. A collaborative network is one route forwards, although there are well known challenges associated with highly interdisciplinary research [121]. However, the relatively small scale of research on sensory feedback for upper-limb prostheses means that the development of this framework can provide an interesting case study that can be adapted to other areas of research.

VI. CONCLUSION
This paper has highlighted the available literature on the sensory feedback needs of individuals with upper-limb difference and the outcome measures used to assess the developed feedback systems in terms of improving control, naturalness, embodiment and reducing phantom limb pain. This has shown that developing user-focused outcome measures is a potential area of improvement, particularly in relation to control and emotional benefits. Sensory feedback is progressing to include cutting-edge technology such as cortical stimulation and Dorsal Root Ganglia interfaces. However, only the simplest form of sensory feedback is found in a few high-end commercial prostheses [122]. This shows the inherent lag in the commercialisation of the research. Establishing stronger interdisciplinary collaborations will accelerate translation by 1) enabling user needs and experiences to be better captured through co-creation and person-based approaches, 2) ensuring the assessment measures used are meaningful for those who use them, and 3) testing the integration sensory feedback and control. This shift in the way research on upper-limb prostheses is conducted is expected to enable the basic user needs, that have been mostly consistent for 25 years, to be finally met.