Sensory feedback for limb prostheses in amputees

Commercial prosthetic devices currently do not provide natural sensory information on the interaction with objects or movements. The subsequent disadvantages include unphysiological walking with a prosthetic leg and difficulty in controlling the force exerted with a prosthetic hand, thus creating health issues. Restoring natural sensory feedback from the prosthesis to amputees is an unmet clinical need. An optimal device should be able to elicit natural sensations of touch or proprioception, by delivering the complex signals to the nervous system that would be produced by skin, muscles and joints receptors. This Review covers the various neurotechnological approaches that have been proposed for the development of the optimal sensory feedback restoration device for arm and leg amputees. This Review highlights the approaches that have been utilized in the implementation of sensory feedback onto prosthetic devices to restore the sensation of touch and proprioception for amputees.

to different body 'sensors': cutaneous mechanoreceptors, proprioceptors and nociceptors. Cutaneous mechanoreceptors transduce the tactile stimuli applied on the skin into signals for the brain that travel along afferent fibres (thus also called tactile fibres).
There are four types of tactile fibre and they are connected to four different mechanoreceptors in the glabrous skin (Fig. 2b). Fast-adapting type I (FAI) and type II (FAII) fibres respond only during dynamic phases of tissue deformation (for example, when

Agonist-antagonist myoneural interface (AMI)
Two muscles-an agonist and an antagonist-are surgically linked to restore natural proprioception

Self-containedneuromusculoskeletal
The prosthesis is directly connected to the residual body via bone, nerves and muscles Artificial skin Sensors' readout is transduced into stimulation parameters

Targeted sensory reinnervation (TSR) with arm nerves
Upper-limb nerves, once directed to the hand, are rerouted to the chest skin    Fig. 2 | artificial sensory feedback is inspired by nature. a, Working principle of a sensory feedback restoration device. Sensors are added to the prosthesis, whose readout is transduced (encoded) into the electrical stimulation of residual fibres. This mechanism is inspired by mammals. In the optimal system, the encoding and the stimulation generate a sensory nerve activation identical to the one that would have been produced by cutaneous mechanoreceptors and proprioceptors. b, Tactile afferents are connected to four types of cutaneous mechanoreceptor, which adapt differently to mechanical stimuli and responses to electrical stimulation 17,109,110 . c, There are varying numbers of tactile afferents (with FAI being the most represented). Data are extracted from refs. 18,111 . d, The receptive fields of tactile afferents are smaller for type I fibres 18,112 . e, Tactile afferents have different innervation densities throughout the inner hand and the sole of the foot 18,111 . Credit: a, Bottom left (grey hand) and far right (brain), reproduced with permission from pixy.org; bottom right (remapping approach), adapted with permission from iStock by Getty Images/Francesco Maria Petrini; bottom, second from left (sensorized glove), reproduced with permission from ref. 75 , under a Creative Commons license CC By 4.0; top, second from left (sensorized insole), adapted with permission from ref. 84 , AAAS. b, Left, reproduced with permission from ref. 109 , Elsevier. d, Hands, adapted with permission from ref. 112 , Elsevier Biomedical Press; feet, adapted with permission from ref. 18 , American Physiological Society there is contact with a surface). Accordingly, intraneural electrical stimulation of these afferents usually elicits perception of intermittent tapping/flutter 17 . Slowly adapting type I (SAI) and type II (SAII) fibres respond to sustained deformations of the skin. SAI fibres encode static or low-frequency changes in tissue deformation (for example, maintaining contact with a surface) and evoke sensations of sustained pressure with electrical stimulation 17 . SAII fibres encode for skin stretches and generally, when stimulated, do not elicit a sensation 17 . FAI fibres are the most numerous (Fig. 2c). Type I fibres have smaller receptive fields-that is, the extent of skin activating the receptor in presence of a tactile stimulus-compared with type II afferents (Fig. 2d). Afferents innervate the skin progressively more towards the distal part of the extremities (Fig. 2e). In the foot, also the lateral part of the sole is more innervated than the medial one 18 . Tactile afferents deliver messages to the central nervous system, in two manners 19 : the rate code (the intensity of a stimulus is proportional to the firing rate of the fibre) and the population code (the intensity of a stimulus is proportional to the number of fibres that are activated). These strategies are used, for example, to convey the information of intensity of a stimulus 19 . Combinations or variations of these strategies are adopted to provide the brain with more sophisticated touch features. Embossed letters are coded through a spatial code: subgroups of fibres fire when they get in contact with a protuberance on an object, while others do not. Very small textures are communicated to the brain through a temporal code (that is, an exact sequence of spikes) 20 . Direct contact force or shape are encoded by the spike latency code: two groups of fibres fire with a specific latency 21 .
Proprioceptors are responsible for the senses of posture and movement of the trunk and limbs, as well as of the sense of force. There are three main types of proprioceptive fibre in the muscles that are connected to two main proprioceptors 22 . (1) Type Ia fibres are connected to muscle spindles (which are embedded within the muscle fibres and make up the bulk of the muscle) and give information on the length of the muscle and the velocity of its change.
(2) Type II fibres are similar to type Ia fibres but they only code for the length of the muscle. (3) Type Ib fibres are connected to Golgi tendon organs (embedded within the tendons) and give information on the force exerted by the muscles. Stimulation of these fibres does not often elicit a sensation 23 . These proprioceptors can be activated or inhibited through vibrations over muscle tendons 22 . Joint receptors are connected to slowly adapting afferents, and evoke a sensation of displacement or a deep sensation from the joint when targeted with electrical stimulation 23 .
Remapping is not an optimal solution. Cutaneous stimulation of amputees' remaining extremities, not innervated by the afferents once connected to the missing limbs, is an example of technology that restores non-somatotopic sensations. Devices based on this technique (also named sensory remapping) have been developed and tested with volunteers 24,25 . Users are usually provided with coded stimuli (such as modulation of pulse rate) and learn to relate these codes to specific sensory information, for example, pinch force 26 . However, intensive cognitive load is usually required for a user to correctly interpret the coded signal 27 . Other problems of the approaches using remapped mechano-, vibro-tactile or electro-cutaneous feedback include: artefacts on the recoding system used for the prosthesis control due to the stimulation 28 , the miniaturization of the systems, power consumption and the quality of the sensation elicited, which is not very pleasant. These issues probably prevent the clinical adoption of such technologies, despite the great advantage of not requiring any surgical intervention.
The device design is tailored with respect to the amputation type. The design of a sensory feedback restoration device varies for upper-and lower-limb amputees. The sciatic nerve, which innervates the foot and lower leg, is more than twice as large as the median and ulnar nerves, which run to the fingers and palms, and it is also difficult to reach through the big leg muscles. The density and placement of the receptors are different between upper and lower limbs. Upper-limb amputees can use their intact hand for almost all activities, while amputees that are missing part of the leg cannot ambulate without a prosthesis. A failed manipulation can lead to a broken glass, whereas a failed step could lead to a fall.
A more proximal amputation entails a greater disability 7,8 (different levels of amputations are depicted in Fig. 1), as persons with a more proximal amputation have less of a residual neuromuscular structure to interface via different surgery and implantable approaches. Also, more proximal levels of amputation require more complexity in the sensory feedback restoration, as larger amounts of information need to be transmitted; for example, the proprioception from the prosthetic knee is not needed for a transtibial amputee.

the role of materials in the development of neuroprostheses
Mimicking the nervous tissue through artificial materials. As the peripheral nervous tissue is composed of a matrix of optimally integrated axons and connective tissues, it is reasonably stiff; however, at the same time, it is soft, as axons and bundles can slide on each other guaranteeing accommodation to the movements of the body by bending and stretching. This is in clear contrast to the materials traditionally used for nerve interfacing, which are several orders sharper and stiffer, and can therefore be harmful to the tissue in which they are implanted. Simultaneously, the body may be a hostile environment for the implanted devices: to be used as materials for neural prostheses, implants have to withstand the body environment over the years. Typically, neural interfaces are made of a substrate, which gives structural stability and compliance with respect to the nerve, and active sites that are embedded within it, delivering electrical current into the excitable tissue. Different geometrical solutions have been tested (Figs. 1 and 3 and Table 1): implants can be placed around the nerves (such as cuff electrodes 29 and flat interface nerve electrodes (FINEs) 30 ), longitudinally through the nerves (such as wire and thin-film longitudinal intrafascicular electrodes (wire LIFEs) 31 and tf-LIFEs 32 ), through the nerves via multielectrode needles (such as Utah electrodes 33 ) or via shafts (such as transverse intrafascicular multichannel electrodes (TIMEs) 34 ).
Different shapes and materials. Polyimide, silicone and parylene are the most commonly used materials for nerve electrodes substrates 35,36 . Silicones are polymers made of siloxane, and polydimethylsiloxane or PDMS is one of its most common types. An example of PDMS processing is spin coating, which is executed twice above and below a metal foil. Polyimides are synthesized by adding dianhydride and a diamine to a dipolar aprotic solvent (which generates poly at room temperature). Thermal imidization (that is, the removal of water through vapourization) of poly is the most common procedure to obtain polyimide. In contrast, to produce polyimide wafers, poly is spin-coated and then treated in a nitrogen atmosphere at a high temperature (~350 °C). Vapour deposition or sputtering are used to deposit metals on these wafers. The metal is then usually encapsulated with a second layer of polyimide, because of the low adhesion among these materials. Parylene (or polyparaxylylene) is a polymer made of the repetition of para-benzenediyl. Parylene-C (poly(dichloro-p-xylylene)) is the most common type used for biomedical applications. Usually parylene is processed through vapour deposition polymerization. The dimer of parylene is split into a monomeric gas, through heating. It is then polymerized on the target, by decreasing the temperature.
The most commonly used electrode is the spiral nerve cuff electrode 29 , designed to be expandable so that it can be fitted around a nerve and accommodated to subsequent neural swelling. Cuffs (Case Western Reserve University and Ardiem Medical) have three active sites (cathodes), an anode and a screen electrode, presenting a good stability in human experimentation 37 . They are made of silicone, with platinum sites (Fig. 3a), and typically have an exposed area of 0.45 mm 2 . However, the active sites of cuff electrodes have a structural barrier to stimulate the fibres in deep fascicles, as fascicles themselves are transversally topographically located in a nerve 38 . To overcome this issue, the FINE reshapes the nerve into a flat geometry to increase the interfacing area by moving central axon populations close to the surface. It is fabricated by moulding a commercial silicone elastomer 30 and the stimulating contacts are evenly distributed in the top and the bottom half of the substrate (Fig. 3b). Contacts are produced using platinum foil, exposing a 0.4-mm-diameter  40 . e, Cuff electrode biocompatibility in cats 49 . f, C-FINE biocompatibility in cats 39 . g, tf-LIFE biocompatibility 54 in rats. The arrows point to the electrodes. Scale bar, 100 μm. h, Wire LIFE biocompatibility 54 in rats. The arrows point to the electrodes. Scale bar, 100 μm. i, A TIME 63 . GND, ground electrode. j, A uSEA 113 . k, Tip of microelectrode with iridium oxide sputtered 115 . l, TIME biocompatibility 116 in rats. The arrow points to the electrode. m, TIME biocompatibility in pigs 116 . The sample is a typical H&E (haematoxylin and eosin) staining microscopy image (×20) of the peripheral nerve, where the TIME electrode has been identified inside the nerve by the arrow. n, uSEA biocompatibility in cats 56 . One electrode misses the nerve's fascicles (left) and another electrode tip is inside a fascicle and close to viable neurons (right). Scale bars, 100 μm (left) and 50 μm (right). Images of the biocompatibility are histological analyses (over long-term periods). Panels e-h, l and n are light microscopy images. Credit: panels reproduced with permission from: a,j, ref. 113  window for the stimulation. C-FINE 39 is a version of FINEs optimal for anatomical locations near joints or organs, as it has a variable pattern of stiffness. Owing to the use of pliable silicone, C-FINEs guarantee stiffness along the width of the nerve to reshape it, while flexibility along its length allows for bending with the nerve. For a direct contact with deeper nerve fibres, LIFEs are used to pinch the nerve. They consist of a Kevlar fibre, metallized with sputter-deposited titanium, gold and platinum, and insulated with an approximately 1-μm-thick medical-grade silicone elastomer 31,40 . The stimulation zone consists of approximately 1 mm of a non-insulated portion of the metallized fibre and a tungsten needle is glued to the LIFEs, to penetrate the nerve's tissue and guide the electrode placement.
Recent widespread use of polymers, such as polyimide and parylene, has led to the use of flexible devices 36 . They can be produced with microfabrication methods, enabling reliable serial production, in contrast to prototype-based manufacturing. Polymer-based electrodes are not only flexible but also the arrays are generally much less thick (ranging between 10 and 20 μm) than rubber-based arrays. They also have a higher number of stimulating channels, as insulation can be provided even between only 10 μm close adjacent metallic tracks. FAST-LIFEs 41 , combining longitudinal intrafascicular arrays and fascicular cuff electrodes, consist of a series of thin-film electrodes and traces deposited on a reinforced silicone substrate. Eight platinum intrafascicular electrodes are placed on the thin extension of the array and are distanced 0.5 mm from each another. Six cuff electrodes are integrated into the thicker part of the array. A combination of cuff electrodes (Ardiem Medical) and LIFE-like intraneural (double-sided filament electrode (ds-FILE) 32 ) electrodes made from polyimide with active sites of platinum have also been also tested in humans.
To exploit further the transversal somatotopic disposition of peripheral nerves, TIMEs have been constructed 34 by micromachining and patterning polyimide substrate, insulation material and platinum electrode sites. A guiding needle is needed for LIFEs and TIMEs, to penetrate the resistant nerve surface, and guide the electrode placement, increasing the complexity and duration of the surgery required.
Utah slanted electrode arrays (USEAs) are selective wire-like multiple contact interfaces that tackle the nerve transversally, without the use of needles. They are developed starting with a silicon wafer and processed with common microelectromechanical system methods, have great spatial resolution and can be used to penetrate several regions in the nerve, via a single implant.
The platinum-plated tips are exposed with an area of about 0.005 mm 2 . The more advanced version of USEAs 33 (marketed by Blackrock Microsystems) includes titanium, titanium tungsten and platinum sputtering 33 . The tips of the electrodes are coated with iridium oxide to facilitate electrical to ionic transduction, and the entire array, with the exception of the tip of each electrode, is insulated with the biocompatible polymer parylene-C. These electrodes are implanted by a pneumatic insertion device.
Higher charge storage through coating. Platinum is the most used material for active sites, because of its good biocompatibility and conducting properties. After implantation, the fibrotic response shifts 42 intraneural electrodes (LIFEs, TIMEs and USEAs) away from the excitable tissue; therefore, these devices could have an insufficient 43 injectable current in the tissue, to elicit neural response. In fact, the foreign body reaction pushes the interface away from the fibres, requiring increased currents to activate them. Specific coatings over the active sites, increasing their roughness and relative surface, can improve the conducting properties and injectable charge 43,44 . In fact, after the failure of stimulation in the first in-human polymer-based approach 45 , this problem was addressed by developing a coating that increases the maximum injectable charge 44 . In the case of TIMEs, the iridium oxide layer acts as active site material with a sufficient charge injection capacity 46 . Since iridium oxide was introduced as an active coating via electroplating, maximal charge injection capacities have increased from 75 μC cm −2 for bare platinum 47 to 2.3 mC cm −2 for iridium oxide, which corresponds to a 30-fold increase 44 . Similarly, in USEAs, sputtered titanium and iridium oxide films (SIROF) were deposited, obtaining a higher charge injection capability 48 .
Material stability and biocompatibility. Among implantable materials, there is a trade-off between invasiveness and efficacy, softness and stability. Their effect on the tissues is measured in complementary ways: through animal (or occasionally human) tissue processing after electrode explants and by observing their long-term functionality via chronic measurements during stimulation tests. Indeed, cuff electrodes were tested in cat experiments 49 revealing their long-term usability up to 28-34 weeks, with a pattern of mild morphological abnormalities, and especially the formation of an external layer of connective tissue. During an extensive human study, up to 160 weeks of cuff electrode use was achieved in humans 50 . Recently, the same electrodes were proven to be stable in human use 37 for up to seven years.
FINE electrodes squeeze the nerves; therefore, their effect on the tissue was carefully studied in rat 51 and cat 52 chronic experiments. The effects were studied on the tissue at regular timings post-implantation, observing moderate damage and concluding that FINEs that apply a small force can reshape the nerve without substantial changes in its physiology or histology. These results were translated to humans, resulting in successful electrode use for several years 53 .
The guiding-needle insertion of intraneural electrodes can damage the nerve, similarly to the pistol-based injection of USEAs, and therefore their biocompatibility is an important factor for their future adoption. Long-term histological and functional studies have revealed a good response of tissue to both wire-based and polyimide tf-LIFEs 54 . Regarding polyimide, a longitudinal animal study was performed, in which different time points of tissue response to polyimide were observed 42 . The selectivity, stability and biocompatibility were assessed in the sciatic nerve of 23 healthy adult rats for up to six months (note that one rat month is comparable to three human years 2 ). A moderate increase in the stimulation threshold of the electrodes was reported during the first four weeks after implantation, remaining stable over the following five months. The time course of these adaptations correlated with the progressive development of the fibrotic capsule around the implants. The density of nerve fibres above and below the inserted implant remained unaffected. The impact of TIME electrodes on nerves was evaluated functionally and morphologically, through implants in the rats' sciatic nerve for two months 55 . The results indicated that implantation of devices in the nerve did not cause relevant axonal loss or demyelination.
USEAs have been studied histologically both in cats 56 and in humans 57 . During a six-month study, the morphology and fibre density of the nerve around the electrodes were found to be normal 56 . Implanted nerves were found to undergo a compensatory regenerative response after the initial injury, which was highlighted in new axons growing around microelectrode shafts. Similar findings regarding inflammation have been observed in humans implanted with an USEA in the ulnar and median nerves 57 .

Bionic limb applications
Limb amputation results in the loss of the extremity receptors and thus the sensory organs that interact with the environment. However, the somatosensory nerves and pathways conveying sensory information to the central nervous system remain functional 58 . Invasive procedures (that is, requiring surgical intervention) for    45 . The subsequent exploitation of the artificial feedback, delivered by wire LIFEs, was implemented in a prosthetic hand 67 , enabling amputees to identify the size and stiffness of different objects with a myoelectric prosthesis without visual or auditory cues. More recently, polymer-based LIFEs (FAST-LIFEs and ds-FILEs) were successfully tested in humans 41,68 . Using different geometries, TIMEs 64 , FINEs 53 and cuff electrodes 69 have been shown to be effective and functional neural interfaces, which provide sensory feedback in upper-limb amputees. Using the neural stimulation delivered through TIMEs, an amputee was able to identify the object's position, shape and compliance. In addition, with FINEs and TIMEs, an improvement in force control has been demonstrated 53,64 . These were the first proofs of motor control benefits thanks to a real-time sensory feedback embedded in a prosthesis. Notably, in hand amputees implanted with TIMEs, the ability to provide information regarding textures was shown 70 . In 2016, the first use of an USEA for the peripheral nervous system (PNS) in humans was shown through recording of peripheral neural signals and providing sensory feedback 71,72 . In the past decade, many other benefits have been reported, connected to the use of direct nerve stimulation using neural interfaces ( Fig. 4 and Table 2). Researchers have begun to demonstrate the benefits of the sensorimotor strategies related to prosthesis control 65,68 . The cognitive aspects of sensory feedback restoration have also been investigated considering prosthesis embodiment 65,73 and integration 74 . Health benefits have been related to the reduction in phantom limb abnormal representations 65,73,75 and phantom limb pain 10 . Sensory feedback restoration has also been proven to have an impact on how long a prosthesis can be used in everyday life 75 .
More recently, notable stability 69,76 and also the home use of bidirectional prostheses using FINE 75 and cuff electrodes 37 for several years have been investigated and demonstrated.
In parallel with direct PNS stimulation, another approach has been developed, in particular for people with a high level of amputation called targeted muscular reinnervation (TMR) 77,78 . TMR is a surgical approach, in which, the residual nerve branches that once innervated the amputated hand are 'redirected' to another area of the body, such as the pectoral region (Fig. 1). The 'redirected' nerves re-grow and reinnervate the new muscles. Similarly, specific sensory nerves can be transferred so that the skin of the chest or the arm is reinnervated, which is called targeted sensory reinnervation (TSR) 78,79 . An arm amputee undergoing TMR to the pectoral region will be able to feel tactile cues on the chest as though they were originating from the missing hand 79 . To evoke this somatotopic sensation, this chest skin could be mechanically or electrically stimulated. A fully portable system was recently extensively tested with amputees at home using both TMR control and TSR provided by vibrators 80 . A similar method was also implemented by exploiting transcutaneous electrical stimulation of the reinnervated area 81 .

Sensory feedback restoration for lower-limb extremities.
Sensory restoration is also possible in lower-limb amputees ( Fig. 5 and Table 2). By stimulating the sciatic nerve, it is possible to evoke the sensations referred on the phantom leg and foot: FINEs 82 and TIMEs 83,84 have been successfully tested in humans. Interestingly, FINEs have been adopted with transtibial amputees and TIMEs with transfemoral amputees, who are radically more disabled 7 . People implanted with FINEs perceived sensations on a few, spatially extended, areas under the phantom foot and on the phantom leg (Fig. 5), important for balance and locomotion. The neural sensory feedback delivered by TIMEs implanted in the tibial nerve were successfully exploited by transfemoral amputees. Distinct and spatially selective sensations of touch, pressure and vibration were elicited from more than 20 positions of the phantom foot, together with the contraction of the muscles of the missing leg (Fig. 5). Notably, this neural feedback was then exploited in motor tasks, which proved that this approach improved users' recognition of prosthesis movement and touch, mobility over stairs and obstacle avoidance 84 . When stepping out of the laboratory into an ecological environment, walking speed and self-reported confidence in the prosthesis increased, while mental and physical fatigue decreased for participants during neural sensory feedback compared with the no-stimulation trials 83 . This is an important health-related benefit since lower-limb amputees have a higher risk of having a heart attack 85 . Together with functional and health outcomes, the cognitive integration of the device into the participants' body schema was confirmed using the neuroprosthetic intervention 84 , by measuring the prosthesis embodiment and cognitive effort while using the artificial leg. Finally, participants reported radically decreased phantom limb pain when provided with neural sensory feedback.
In addition to PNS interfacing, a promising surgical approach has also been developed, using an agonist-antagonist myoneural interface (AMI) to restore proprioceptive sensations in a transtibial amputee 86 . The technique consists of connecting in series two opposing muscle-tendon ensembles (an agonist and an antagonist). The contraction and shortening of one muscle (by volitional intention or electrically activated) thus induces the stretching of the other muscle in series and vice versa. This linked motion enables the natural body sensors embedded in the muscle tendon to send signals to the brain, transmitting information on the muscle length, speed and force, which is then perceived as physiological joint proprioception. Multiple AMI pairs can be created for the control and sensation of multiple prosthetic joints.
All these findings provide the rationale for larger population investigations of the clinical utility of neuroprostheses that restore sensory feedback in leg amputees.

Restoration of proprioception using PNS.
Proprioception is of crucial importance for the manipulation of objects 87 . It is known 22 that vibrations can elicit/inhibit proprioception, and thus tendon vibrators have been used to elicit the sensation of hand movements and postures in transhumeral participants with TMR 88 . In contrast, electrical stimulation of the nerve rarely evokes a proprioceptive sensation without also activating a muscular one. Indeed, using microstimulation, it has been proven that sensations of movement or position cannot be directly evoked by stimulating single muscle spindles 23 . Proprioceptive illusions are more difficult to induce than tactile sensations, as they require more than the mere additive activation of nervous fibres. However, the activation of a larger group of muscle spindle afferents could induce a sensation of joint movement 89 , which probably explains the findings shown by adopting certain neural interfaces 67,72 .
The difficulty in eliciting proprioceptive percepts is also related to how the proprioceptive afferents are distributed inside the nerve. The muscle spindle fibres are not grouped together within the nerve according to their function as the tactile afferents 90 , instead they are clustered with motoneurons innervating the muscles 91 . Stimulating a group of only proprioceptive fibres is less likely than it is with tactile afferents: the active sites that induce muscle spindle fibres activation are also likely to easily evoke phantom muscle contractions. One relevant remark here is that the nerve fascicles could be not maintained at all levels, in particular in very proximal segments 92 . An alternative approach to conveying proprioceptive information is through sensory substitution, which consists of providing information on limb posture by electrically evoking a distinct cutaneous sensation 93 . This approach has not yet been tested in activities of everyday life.

Prospective and big picture
Technological limitations to overcome. Silicon-and polyimidebased cuff electrodes are currently in development for human certified use (for example, Neuros Altius System, USA, and Neuroloop, Germany). However, polyimide-based solutions (especially thin-film electrodes) could face major problems in terms of approval for use as active implantable medical devices, since they are associated with a risk of failure when implanted over the long term, and there has also been no previous use of medical-grade polyimide for this purpose 36 . However, the latest technological developments in iridium oxide coatings and in layers promoting adhesion between metal and polyimide have shown promising results with respect to long-term integrity and stability, and could facilitate the roadmap towards clinical approval for medical use 36 . The use of parylene-C and the coating of several medical devices could be a promising step forwards 94 . Indeed, USEAs already include parylene-C in their composition.
While developing these sophisticated technologies, the rudimental connection between the prosthesis and the stump is an often-overlooked limitation. At present, uncomfortable sockets are used, which in upper-limb amputees can compromise the stability of the prosthesis, and in lower-limb amputees can be a source of pain. Connecting the residual bone to the prosthesis via an implanted biocompatible screw, called osseointegration 69 , offers a solution. It improves the long-term stability of such devices and provides greater comfort for patients, especially as it alleviates the problem of the perceived weight of the prostheses 95 . It also provides the perfect space for implantable neurotechnology leads to pass through the body to connect with external electronics 69 .
Capturing sensory information external to the limb is one of the main technological drawbacks of the technologies described in this Review, since prosthetic limbs are not equipped with sensors. Future prostheses will need this data as input to optimize  The optimal neural stimulation should elicit a rich and natural perception generating a natural activation of the neural fibres. This can be achieved with a model-based approach in which, by interacting with the environment, the prosthesis can extract all the necessary information using advanced wearable sensors. Then, biomimetic sensory encoding algorithms convert sensor outputs and generate the optimal stimulation parameters based on the modelled fibre firing rate and recruitment. After this step, the electro-neural model identifies the optimal electrode configuration and the optimal electrical pulse trains are delivered using a multichannel neural stimulation. Finally, the electrical stimulation is injected into the somatosensory nerve through a fully implantable system. SA1, slow-adaptive fibre class 1; RA, rapid-adaptive fibre class; PC, Pacinian corpuscle fibre class; spks, spikes; IPG, implantable pulse generator; V, voltage; T, time; FEM, finite element method; CHn, channel n. N af , fast sodium; N ap , persistent sodium; K s , slow potassium; L k , linear leakage; C n , C m and C j , linear conductances; G m , G j , G ax , G an , G pn and G px , resistances representing the myelin sheath. Credit: dynamic skin indentation in sensing panel and all of biomimetic sensory encoding panel, reproduced with permission from ref. 65 , Elsevier; all of electro-neural modelling panel, adapted with permission from ref. 104 , under a Creative Commons license CC By 4.0 neurostimulation. To attain this goal, important attention has been paid to developing prosthetic electronic skin 96 : a high-density matrix of sensors implemented over a flexible material that can be applied over the hand or leg. While various sensory signals 97 have been acquired using this technique (such as pressure, movement or temperature), the robustness and real-life use of these technologies are yet to be proven. Indeed, studies 83,84 suggest that for leg amputees, a limited number of sensors (such as three for the foot and one for the knee) may be sufficient. In contrast, in the hand, a high density is needed to implement the biomimetic paradigms 65,98,99 .
Uniform performance metrics are required. Many studies report data in different ways, which hinders our ability to understand the relevance of the findings: sensations should be reported uniformly 100 together with their time evolution. In this way, clinical efforts globally could provide unified information regarding the behaviour of these devices and provide solutions to common problems.
In addition, common metrics should be defined (Table 2), which should be simple while meaningful, and therefore implementable in different scenarios. One example of such an indicator could be the time of use 75 of the device for upper-limb prostheses with and without intervention. This could help decrease the prosthesis abandonment rate 9 . Regarding lower-limb amputees, the most direct indicator of an intervention's success is an increase in mobility 101 . This then leads to improvement in the cardiovascular system 83 , and potentially also to social and economic reinsertion.
Despite their differences, several general rules can be drawn from the studies described in this Review: typically, extraneural electrodes (cuffs and FINEs) exhibit a higher stability, at the price of lower selectivity (and therefore efficiency), compared with intraneural electrodes. This trade-off between invasiveness and selectivity has already been reported in animals 102 (Fig. 6a). However, this is more evident in leg applications: in fact, when interfacing the nerves of the arm, the maps of the elicited sensations are similar across different interfaces (Fig. 4). This is most probably because leg nerves (tibial and sciatic) are much bigger than those in the arm (median and ulnar), making it more difficult for extraneural devices to stimulate inner fascicles (Fig. 5).
Transforming paraesthesia into touch increases device acceptance. Although direct neural stimulation through neural interfaces has extensively shown its ability to evoke a large spectrum of tactile feelings, these sensations, referred to the phantom extremities, have often been described as unnatural and more similar to paraesthesias 103 . Biomimetic sensory encoding strategies 99 (that is, that resemble natural tactile coding) hold the promise of providing more pleasant sensations and therefore of increasing device acceptance 14 .
Various encoding strategies have been used to translate the readout of sensors embedded or added to the prosthesis into stimulation parameters (amplitude, pulse width, frequency and duration of pulse trains). The linear modulation of injected charges 10,53,64,69 or frequencies 62,67,71 has been adopted to translate sensor information into neural modulation. This is in accordance with the physiology of afferent tactile fibres, which deliver information on the intensity of a sensation to the brain through population recruitment or changes in firing activity 19 . Recruitment and firing are controlled by modulating stimulation amplitude and frequency, respectively.
Despite the fact that these techniques are easy to implement in closed-loop systems, the perceived naturalness of the electrically evoked sensations is often rated as very low by the patients 63 (Figs. 4 and 5). In fact, the natural touch coding and the relationship between natural sensors and neural activity is more complex than intensity coding alone (Fig. 2). A major problem is that these types of stimulation induce highly synchronized neural firing in the entire population of recruited fibres 99 : it is impossible with the current technologies to separately stimulate different afferent types with the same active site.
Recent studies have presented more complex approaches 53,65,66 , which stimulate the implanted fibres following the close-to-natural ensemble activation of tactile afferents (called biomimetic 99 ; Fig.  6b). The results showed that biomimetic stimulation elicited more natural sensations than the previously presented strategies, with consequent benefits in terms of control of the prosthesis and its embodiment. Similar results were also confirmed with a participant implanted with a USEA with biomimetic sensory encoding strategies adopted 66 .
Computational modelling for optimization. Electrodes used for stimulation can have different geometries, numbers of stimulating contacts, placement and stimulation protocols. This high-dimensional problem cannot be solved by empiric brute-force approaches, but requires exact computational models, exploiting accumulated knowledge. These hybrid models combine the electromagnetic solution of the field induced in the nerves by the electrical stimulation, with the consequent, nonlinear response from the axons 104 (Fig. 6b), and are implemented in in silico nerves. Different electrode materials and configurations (that is, electrode placement, shape and dimensions) can be tested in the model, thus obtaining the optimal geometrical selectivity (for example, stimulation of the sole fascicles innervating the forefoot with respect to the heel). The model computation of the axon activation of target fascicle(s) with respect to all fascicles could be used to evaluate the specificity and therefore the efficiency of the proposed designs. In parallel, different stimulation strategies could be tested obtaining the optimal fibre recruitment, expressed as a lower threshold and smaller curve steepness of activated fibres. Models can therefore give direct indications to manufacturers both for the design of the materials and their shape, used in electrodes, and also for the optimal algorithms to be ported onto the system controllers and implantable units.
Towards widespread use. The main goal of these technologies is to pass from proof-of-concept studies to clinical reality. To achieve this, several technological and regulatory steps are required 105 . The electrodes should be easy to implant with minimally invasive surgery. Due to all the problems with percutaneous wires (including possible infections), the development of fully implantable systems communicating with outside sensors (Figs. 1, 2 and 6b) is a mandatory next step. Commercial neurostimulators have pre-programmed stimulation protocols, without the need for continuous transcutaneous communication. Contrary, in bionic limbs, a high burden of information has to be transmitted through the skin to enable bidirectional communication. This involves an important constraint to battery capacity and implant size, which represent notable technological challenges.
However, even if the perfect system is achieved, providing some benefits for patients, this will not be sufficient for wide clinical use. Health and quality-of-life benefits should be assessed and documented, helping healthcare systems to save spending that would otherwise be necessary for the treatment of patients. This would increase the likelihood that these neurotechnologies can be reimbursed by insurances and healthcare systems, therefore guaranteeing their widespread to all patients in need.
Future developments could be systems that provide more selective or less invasive interfacing with nerves, such as optogenetics 106 , conductive polymers 107 and ultrasound based 108 , which today are at the stage of animal experimentation. The neurotechnologies presented have enormous potential in terms of social and health impacts, and after the recent discoveries, we are a bionic step closer to their achievement.

Data availability
The data and calculations that support the findings of this Review are available from the corresponding authors upon reasonable request.