Trends in neuroprosthetic technologies to restore sensations

Abstract

The loss of sensory function significantly reduces the ability to perform daily activities and lowers quality of life. Accordingly, the artificial replacement or support of impaired sensory function has become one of the main objectives of neural engineering and neuroprosthetics. This review systematically organizes auditory, visual, and somatosensory restoration technologies by target organ—from peripheral sensory organs to the central nervous system—and provides a comprehensive overview of the achievements and limitations of each approach. Auditory restoration includes cochlear, brainstem, and auditory cortex approaches; visual restoration includes retinal, optic nerve, and visual cortex approaches; and somatosensory restoration includes somatosensory cortex and peripheral nerve approaches. Although neuroprosthetics-based sensory restoration technologies have achieved considerable progress, the sensations currently restored remain far from natural perception. This review discusses future directions to address these limitations, focusing on higher-density electrodes, device and system optimization, decoding-based biomimetic stimulation, and personalized stimulation strategies. Ultimately, these advances may enable more natural sensation and improve the quality of life of patients with sensory impairments.

Introduction

Humans rely on various sensory systems to perceive the outside world and the internal state of the body. Sensation is generally classified into the five senses—vision, hearing, touch, smell, and taste—and each sensory system detects different types of stimuli while playing essential roles in survival, behavior, environmental awareness, and social interaction [1, 2]. The loss of sensory function therefore goes beyond simple inconvenience, significantly reducing a person’s ability to carry out daily life independently and lowering quality of life [3, 4]. Accordingly, technologies designed to artificially replace or support impaired sensory function have long been regarded as an important goal in the fields of neural engineering and neuroprosthetics.

Among the five senses, hearing, vision, and somatosensation have been the primary focus of sensory restoration [5]. This is because these senses play major roles in daily life and functional behavior, affect large patient groups, and offer anatomical and physiological pathways that can be targeted relatively directly through electrical stimulation via neural interfaces. Here, we review the development of technologies for auditory, visual, and somatosensory restoration, with current limitations and challenges. We further discuss the technical and physiological challenges that must be addressed to achieve more natural and precise sensory restoration, and suggest future directions for the field.

Technologies for auditory restoration

The restoration of auditory sensation through neural implants represents one of the most successful applications of neuroprosthetic technology. Among sensory systems, the auditory pathway has proven particularly amenable to electrical stimulation due to its well-defined tonotopic organization and relatively direct mapping between peripheral input and perceptual experience (Fig. 1a–b) [6]. Over the past several decades, cochlear implants (CIs) have emerged as widely adopted clinical solution for restoring hearing in patients with hearing loss [7]. However, CIs are not applicable to patients with damaged cochlear nerve by neuromas or resection. Auditory brainstem implants (ABIs) were developed to restore hearing by stimulating auditory nucleus in the brainstem. More recently, efforts have expanded toward direct stimulation of central auditory structures, particularly the auditory cortex, as a potential next-generation strategy for more complete and flexible sensory restoration (Table 1) [7–14].

Fig. 1.

Approaches for auditory restoration. a Schematic illustration of a cochlear implant system and surgical implantation procedure. Modified from multiple sources including [185–187]. b Tonotopic organization of cochlea [6]. Adapted from Kern et al. (2008), licensed under CC BY 4.0. c Illustration of auditory brainstem implant [30]. Adapted from Slama (2015). d Tonotopic organization of human auditory cortex. For sound stimuli spanning multiple frequencies and intensities, activation maps were generated to illustrate the response level [35]. Adapted from Langers et al. (2012), licensed under CC BY-NC 3.0. e Working model of human auditory cortex [37]. The color-coding system for the isofrequency gradients is directly aligned with the colorscale used in the tonotopy maps. Adapted from Moerel et al. (2014), licensed under CC BY 4.0. f Photograph of thin-film microelectrode arrays implanted on the cortex surface [41]. Adapted from Lee et al. (2021), licensed under CC BY 4.0

Table 1.

Comparison of cochlear, auditory brainstem, and auditory cortical approaches for auditory restoration

Feature	Cochlear approach	Auditory brainstem approach	Auditory cortex approach
Primary target	Spiral ganglion cells	Cochlear nucleus	Primary auditory cortex (A1)
Spectral resolution / auditory acuity	Formal acuity QSMD 61.1%, SMRT 3.1–3.8 RPO CNC word test 41.1%, AzBio in quiet 55.8%, AzBio in Noise 51.1%	Formal acuity SMRT 0.993 RPO CNC word test 8%	Generating random auditory perceptions. No standardized acuity.
Representative stimulation hardware	Cochlear Ltd.: 22 contacts, “n-of-m” channel selection strategy Advanced Bionics: 16 contacts, current steering strategy MED-EL: 24 contacts, fine structure processing (FSP) strategy	Cochlear limited: 21 contacts MED-EL: 12 contacts Oticon Medical: 15 contacts Generally, ABIs share similar stimulation strategies of the manufacturers.	Shank-type electrode: 16 contacts (NeuroNexus, Blackrock Microsystems) Surface MEAs: 8–64, 529 contacts (Huake Hengsheng Med. Tech., Blackrock Microsystems, Original) Generally studied with universal brain electrodes.
Functional outcome	Speech recognition in words and sentences. Listening to music or even playing an instrument in some cases.	Low level of open-set (sound only) speech recognition Higher recognition with visual hints (e.g. 26% CNC word test with lip reading)	Auditory hallucinations (water dripping”, “hammer and nail”, music, or human voices) Distal tones
Long-term usability	Device survival rates for 5, 10, and 15 years: 98.2%, 97.7%, 94.9% (excluding the recalled devices)	62.5% of patients maintained ABIs for 10 years.	Only acute experiment at clinic level. Implanted for several months in animals (pig, rat).
Key limitations	Not applicable for some patients with auditory nerve damage. The functional outcome is highly variable (0-100% in speech recognition).	Low level of spectral resolution and speech perception The functional outcome is highly variable.	Complexity of cortical encoding High variability of perceptual outcomes across individuals Unpredictability in long-term usage due to plasticity
Surgical access / invasiveness	Insertion through round window (RW), extended round window (Ext RW), cochleostomy Comparably safe (5.6% implantation complication)	Craniotomy and brainstem implantation Major and minor complications occurred in 14% and 29.8% of patients, respectively.	Craniotomy and cortical implantation Highest invasiveness and neurological risk
References	[18, 20–25, 160, 188–194]	[11, 31, 32, 195–198]	[38, 42, 199–205]

ABI auditory brainstem implant, A1 primary auditory cortex, CNC consonant-nucleus-consonant word test, FSP fine structure processing, QSM quantitative speech measures, RPO recognition percentage of open-set words, and SMRT spectral-modulation ripple test

Cochlear approaches for auditory restoration

The conceptual foundation of cochlear implants begins from early demonstrations that electrical stimulation of the auditory nerve could evoke auditory sensations in humans [15]. Initial systems in the 1960s employed single-electrode designs, which enabled detection of sound presence but provided limited spectral resolution and poor speech understanding [16]. A major breakthrough occurred with the development of multichannel CIs in the 1970s and 1980s. These systems exploited the tonotopic organization of the cochlea by placing multiple electrodes along its length, allowing frequency information to be encoded spatially [17]. This innovation significantly improved speech perception and ultimately led to the first clinically viable devices and regulatory approvals in the 1980s [7]. According to an investigation in 2021, CIs significantly improved 84.8% of patients in word tasks, 87.6% in quiet sentence tasks, and 78.6% in noisy sentence tasks [18]. Although CIs generally enhance speech recognition, outcomes remain highly variable among individual patients, with improvements ranging widely from 0% to 100% [18]. The variability underscores that improvements are remaining in current technology and clinical strategies.

Modern CIs consist of an external processor and an implanted electrode array inserted into the scala tympani of the cochlea (Fig. 1a). Acoustic signals are transformed into electrical stimulation patterns that activate the auditory nerve [7]. Currently, a few major manufacturers provide commercial CIs with the same principle, but there are some differences in technical details. One of the differences is the number of electrode contacts, which corresponds to the number of frequency bands of the sound information that the CI can express [19]. Commercial CIs typically employ between 16 and 24 intracochlear electrode contacts depending on the manufacturer. CIs from Cochlear Limited providing up to 22 electrodes, Advanced Bionics 16, and MED-EL 24 electrodes [20,21,22]. The manufacturers have developed distinct stimulation strategies to improve auditory restoration besides the number of contacts. For example, Cochlear Limited’s systems utilize “n-of-m” channel selection strategies to maximize spectral efficiency while minimizing channel interaction, whereas Advanced Bionics employs current steering techniques to generate virtual channels and enhance spectral resolution [23,24]. Lastly, MED-EL’s approach emphasizes temporal fine structure processing (FSP) coding to preserve low-frequency timing cues which are critical for pitch perception and speech understanding in noise [25]. Spatial hearing was once considered a major limitation of unilateral implantation; however, it has been partially addressed by bilateral CIs. Bilateral CIs enable access to binaural cues such as interaural level differences (ILDs) and, to a more limited extent, interaural time differences (ITDs) [8,9].

Despite these advances, CIs are still limited by factors such as current spread within the cochlea, restricted electrode count, and variability in neural survival. Consequently, while CI users often achieve good speech perception in quiet environments, challenges remain in complex listening scenarios and music perception. The dependence of CIs on the peripheral auditory system imposes fundamental constraints. Successful CI outcomes require a functional auditory nerve, limiting applicability in cases of auditory neuropathy or severe neural degeneration [26]. Furthermore, the simplified encoding strategies used in current devices do not fully replicate the rich spatiotemporal dynamics of natural auditory signals, resulting in perceptual experiences that are functional but not entirely natural [7].

These limitations have motivated exploration of more central targets within the auditory pathway. Direct stimulation of the auditory cortex offers the potential to bypass damaged peripheral structures and access higher-level representations of sound.

Auditory brainstem approaches for auditory restoration

Auditory brainstem implants (ABIs) were first introduced in 1979, with the initial surgical implantation performed in 1997 [27]. ABIs stimulate cochlear nucleus (CN) located in the brainstem to generate auditory sensation. Early applications of ABIs were primarily focused on patients with acoustic neuromas associated with Neurofibromatosis type 2 (NF2), who experienced sensorineural hearing loss due to tumor progression or surgical resection [28,10]. In these patients, ABIs generally provided limited auditory benefit, most notably in enabling awareness of environmental sounds. However, unlike conventional CIs, ABIs rarely restore functional hearing, and only a small proportion of NF2 patients achieve meaningful conversational speech perception [27]. Over time, the clinical indications for ABIs have expanded to include non-NF2 patients who are not suitable candidates for cochlear implantation [29,11]. Ironically, even if the ABIs were firstly developed for tumor patients, nontumor patients demonstrated generally higher speech recognition performance after ABI implantation [10]. A systematic review comparing ABI outcomes revealed a substantial performance gap between tumor and nontumor etiologies. For the tumor cohort, the weighted mean speech recognition scores were 39.2% (range: 19.6%–83.3%) for closed-set words, 23.4% (range: 17.2%–37.5%) for open-set words, and 21.5% (range: 2.7%–48.4%) for open-set sentences, alongside a mean Category of Auditory Performance (CAP) score of 3.1 (range: 1.0–3.2). Conversely, nontumor patients demonstrated markedly superior hearing rehabilitation, achieving a weighted mean of 79.8% (range: 31.7%–84.4%) for closed-set words and 53.0% (range: 14.6%–72.5%) for open-set sentences, with a mean CAP score of 2.30 (range: 2.0–4.7) [10].

The first ABI approval by United States Food and Drug Administration (FDA) was in 2000 for NF2 patients [12]. Currently, commercially available ABI manufacturers are Cochlear Limited, Med-EL, and Oticon Medical [29]. Cochlear Limited’s latest model, ABI541, has 21 electrode contacts on the 8.5 × 3.0 mm sized paddle [11], which is the largest and has the most contacts (Fig. 1c) [30]. In contrast, MED-EL’s latest ABI, Concerto ABI implant system, features 12 contacts with smaller-sized paddle. MED-EL designed the electrode surface to be the same size and shape as the dorsal cochlear for better stability and adhesion [31]. Oticon Medical’s latest ABI, Digisonic SP ABI brainstem implant, has 15 contacts with 3 × 7.8 mm sized paddle [32]. About the stimulation strategy, ABIs share a common concept with CIs [13]. Since all the manufacturers also have developed CIs, the same external devices including speech processors and receiver stimulators of the CIs are used for ABIs [12]. However, since tonotopy of cochlear nucleus is still poorly understood compared to the cochlea [33], it is very challenging to optimize stimulation strategy for ABIs.

Although ABIs generally demonstrate poorer auditory performance and more limited speech perception compared to CIs, they offer a distinct clinical advantage in patients who are not candidates for CIs. In particular, for individuals with NF2 or those who have undergone surgical resection resulting in damage to the cochlear nerve, ABIs can partially restore auditory sensation by directly stimulating the brainstem. Despite their limitations, ABIs provide valuable alternatives for patients who are not applicable to CIs.

Cortical approaches for auditory restoration

The auditory cortex is the place for processing complex acoustic features, including frequency, temporal patterns, and higher-order perceptual attributes [34]. Unlike the cochlea, which primarily encodes spectral information, the cortex integrates auditory inputs into meaningful perceptual representations. The auditory cortex is organized into multiple hierarchical regions, each contributing to different aspects of sound processing. The primary auditory cortex (A1), located in Heschl’s gyrus, is characterized by a well-defined tonotopic organization and is primarily responsible for encoding basic acoustic features such as frequency and intensity [35] (Fig. 1d–e). Surrounding A1 are the belt areas (secondary auditory cortex, A2), which process more complex spectral and temporal features of sound, including bandwidth and amplitude modulation [36, 37]. Beyond these regions, the parabelt (tertiary auditory cortex, A3), located in the superior temporal gyrus, is involved in higher-order auditory processing such as speech perception and the integration of auditory information with other sensory modalities [36].

Experimental studies have demonstrated that electrical stimulation of the auditory cortex can evoke sound-like percepts and modulate auditory perception. In animal models, intracortical micro stimulation (ICMS)

has been shown to elicit frequency-specific behavioral responses aligned with cortical tonotopy [38]. Additionally, animals can learn to interpret patterned cortical stimulation as distinct auditory cues, suggesting that such stimulation can serve as an alternative sensory input [39]. In human studies, intracranial electrical stimulation of auditory cortical regions has been reported to induce auditory sensations such as tones or noise-like percepts [40]. However, these percepts are typically simple and lack the richness and fidelity of natural hearing. Moreover, the mapping between stimulation parameters and perceptual outcomes remains poorly understood.

Despite its promise, auditory cortex stimulation faces several critical challenges. One major limitation is the complexity of cortical encoding, which involves distributed and context-dependent neural representations that are difficult to reproduce through electrical stimulation [34]. Achieving sufficient spatial and temporal resolution requires high-density electrode arrays, which are currently limited by technological and biological constraints. Cortical microelectrode arrays have been studied to stimulate cortical areas including auditory areas in different numbers and orientation of electrode contacts [41] (Fig. 1f). An electrode array achieved 529 channels in sub-millimeter scales which is one of the greatest number of channels among studies of auditory cortex stimulation [42]. However, neural stimulation using surface electrodes inevitably suffers from the degradation of spatial resolution caused by current spread [43]. Another challenge lies in the variability of perceptual outcomes across individuals, as well as the role of neural plasticity in adapting to artificial stimulation. While plasticity may facilitate learning, it also introduces unpredictability in long-term performance.

Technologies for visual restoration

Light-evoked visual information is initially detected in the retina. The retina consists of three principal neuronal layers: photoreceptors, bipolar cells (BCs), and retinal ganglion cells (RGCs). Among these, photoreceptors convert light stimuli into electrical signals. The generated signals are sequentially transmitted through BCs and RGCs and conveyed via the optic nerve to the lateral geniculate nucleus (LGN). From there, the information is further processed in the primary visual cortex, ultimately leading to visual perception [44].

Abnormalities occurring along this visual information transmission pathway can result in reduced visual function or complete vision loss. Ophthalmic diseases such as retinitis pigmentosa, age-related macular degeneration, and diabetic retinopathy are major causes of visual impairment [45–47]. These diseases damage photoreceptors and retinal microstructures, leading to the loss of visual function. In addition, glaucoma causes optic nerve damage due to increased intraocular pressure, resulting in progressive visual field defects that can eventually lead to central vision loss in advanced stages [48].

Although the pattern of vision loss varies depending on the site of damage and the underlying pathophysiology, these diseases commonly lead to irreversible visual dysfunction. Globally, approximately 43 million people are blind and about 550 million people suffer from visual impairment [49]. To address this growing clinical burden, various visual restoration technologies have been investigated. These approaches can be broadly categorized into retina-, optic nerve-, and cortical-based strategies, depending on the location along the visual information pathway that is targeted (Table 2) [44, 50–52].

Table 2.

Comparison of retinal, optic nerve, and cortical approaches for visual restoration

Feature	Retinal approach	Optic-nerve approach	Cortical approach
Primary target	Inner retina (bipolar-cell and retinal ganglion-cell layers)	Optic nerve axons	Visual cortex, mainly V1 (occasionally V2)
Spatial resolution / visual acuity	Formal acuity PRIMA (100-µm pixels), 1.32 ± 0.16, logMAR Alpha IMS/AMS, ~ 1.4 logMAR Argus II, up to 1.8 logMAR	Phosphene dots 10–40 arcmin Typical envelopes ~ 1–10 deg² No standardized acuity	400-µm-spaced electrode pairs were perceived as two distinct phosphenes in 90% of trials.
Representative stimulation hardware	Epiretinal: 60 electrodes (Argus II) Subretinal: 1,500–1,600 pixels (Alpha IMS/AMS) Subretinal: 378 photovoltaic pixels (PRIMA)	4-contact self-sizing spiral cuff electrode 12-site Optic SELINE intraneural array	Surface arrays: tens of contacts (e.g., 60-contact Orion) Intracortical: 96-channel UEA in humans Intracortical: 1,024-channel systems in nonhuman primates
Functional outcome	Letter/word reading Object identification Localization Mobility tasks	Trained simple-pattern recognition: 63% across 45 patterns Line-orientation discrimination: 100%	Rapid letter-form recognition (up to 86 forms/min) Letter identification and object-boundary recognition
Long-term usability	Argus II: up to 5 years in a prospective multicenter trial.	Several years of functional cuff stimulation in a single implanted participant No modern longitudinal clinical series or standardized long-term acuity dataset	96-channel UEA: 6 months in a human study
Key limitations	Requires preserved downstream retinal circuitry and optic nerve Limited acuity and visual-field coverage Phosphene distortion from electrode–tissue geometry and axonal activation	Requires an intact optic nerve. Irregular phosphene geometry and limited independent channels Sparse human evidence; intraneural selectivity remains preclinical.	Complex simultaneous patterns can merge. Implanted-array longevity and large-scale stimulation remain unresolved.
Surgical access / invasiveness	Intraocular implantation Generally lower risk of serious adverse effects	Direct optic-nerve exposure Technically challenging, with risk of nerve injury.	Craniotomy and cortical implantation Highest invasiveness and neurological risk
References	[57, 58, 61, 64, 65, 68, 71, 72, 78]	[89–91]	[93–97]

V1 primary visual cortex, logMAR logarithm of the minimum angle of resolution

Retinal approaches for visual restoration

Attempts to artificially generate vision by directly stimulating the retina began in the mid-1990s [53–55]. In 2003, the first clinical results of a retinal implant in a patient were reported [56]. Subsequently, devices were developed in several configurations depending on the implantation site, including epi-retinal [57], sub-retinal [58–61], and suprachoroidal [62, 63] approaches (Fig. 2a). Based on these technologies, companies such as Second Sight and Retina Implant AG obtained European CE certification and U.S. FDA approval in the early 2010s [60, 64, 65], leading to clinical trials (Fig. 2b-c). Research then focused on increasing the number of electrodes and photodiodes to improve the resolution of retinal stimulation, and multiple clinical trial results were reported [58, 64, 66–70]. However, the visual resolution actually experienced by patients did not reach the theoretical level predicted from device specifications [71, 72]. In addition, these devices continue to face technical challenges associated with increasing electrode density, maintaining biocompatibility, and ensuring long-term durability in vivo.

Fig. 2.

Retinal and cortical approaches for vision restoration. a Schematic of retinal structure and retinal implant. b Clinical testing of a retinal prosthesis in a blind patient, demonstrating functional evaluation of restored visual perception [65]. Adapted from da Cruz et al. (2013), licensed under CC BY 3.0. c Clinically implanted retinal prostheses: epiretinal device (Argus II, left) [65] and subretinal device (Alpha AMS, right) [68]. Left, adapted from da Cruz et al. (2013), licensed under CC BY 3.0. Right, adapted from Edwards et al. (2018), licensed under CC BY 4.0. d Photovoltaic retinal prostheses: epiretinal device (left) [74] and subretinal device (right) [75]. Left, adapted from Vagni et al. (2022), licensed under CC BY 4.0. Right, adapted from Bhuckory et al. (2025), licensed under CC BY 4.0. e Photovoltaic nanowire retinal device [76]. Adapted from Tang et al. (2018), licensed under CC BY 4.0. f–g Intracortical microelectrode array implanted in the visual cortex of a blind volunteer, with predicted retinotopic map organization (V1–V3) overlaid on a 3D brain reconstruction (f), and phosphene perception evoked by electrical microstimulation, showing charge-dependent perceptual probability, corresponding neural population response from neighboring electrodes, and correlation between neural and behavioral thresholds (g) [97]. Adapted from Grani et al. (2025), licensed under CC BY-NC 4.0

In the mid-2010s, studies reporting material-based devices utilizing photovoltaic effects began to appear, aiming to overcome the limitations of conventional retinal implants that relied on separate electrodes, photodiodes, and power transmission modules. Photovoltaic materials can generate electrical current in response to light, allowing such previously separated components to be integrated into a single element. This integration enables higher stimulation density and compact structure compared to conventional approaches. In 2015, a photovoltaic-based retinal implant was developed and successfully induced visual evoked potentials in rats [73]. In 2022, a photovoltaic epi-retinal device was reported to maintain partial visual responses in minipigs for approximately one month [74]. More recently, in 2025, another sub-retinal photovoltaic device was shown to maintain limited visual responses for about six weeks in a rat model [75] (Fig. 2d). In a related approach, a sub-retinal device coated with photovoltaic nanowires was reported to induce partial visual responses in a mouse model [76] (Fig. 2e). Subsequently, visual responses to light stimulation were confirmed in a primate model in 2024, further demonstrating the potential of nanowire-coated photovoltaic devices for visual restoration [77]. Importantly, the clinical feasibility of this approach has also been demonstrated in patients with geographic atrophy due to age-related macular degeneration. In a first-in-human feasibility study, a wireless subretinal photovoltaic implant elicited prosthetic visual perception in the former scotoma and enabled simultaneous use of prosthetic central vision and residual peripheral natural vision over 18–24 months of follow-up [78].

In parallel, alternative approaches beyond conventional electrical stimulation paradigms have also been explored. For example, a noninvasive ultrasound system capable of stimulating the retina through the cornea has been proposed [79]. Studies published in 2006 and 2012 suggested the possibility of restoring visual function through transplantation of rod photoreceptor precursor cells into the retina [80, 81]. In 2020, partial recovery of visual function was reported following transplantation of chemically reprogrammed photoreceptor-like cells into the retina [82]. Approaches using optogenetics to selectively stimulate ON and OFF retinal pathways in response to light have also been explored [83–86]. More recently, optogenetic therapies have entered the clinical stage, and partial restoration of vision sufficient for object recognition and localization has been reported [87]. Nevertheless, these approaches still face several technical challenges, including phototoxicity, low viral transduction efficiency, and the long-term stability of opsin expression, which currently limit their clinical applicability [52, 88].

Optic nerve approaches for visual restoration

The optic nerve serves as the pathway transmitting visual information from the retina to the brain, which consists of bundles of axons from retinal ganglion cells. Electrical stimulation of the optic nerve using cuff electrodes was first reported in 1998, although subsequent studies remained limited [89, 90]. More recently, in 2019, a study reported the feasibility of inducing visual responses through 12-channel intraneural stimulation of the rabbit optic nerve [91].

Cortical approaches for visual restoration

As early as the 1950s, electrical stimulation of the visual cortex was reported to induce phosphenes [92]. In 1996, ICMS applied to the visual cortex of a blind patient successfully induced two distinct phosphenes [89]. Although studies on visual restoration through cortical stimulation continued thereafter, this approach received relatively less attention compared with artificial retina technologies. However, recent studies reported in the 2020s have demonstrated several notable advances. For example, a Utah array–based 1024-channel implant to stimulate the visual cortex enabled monkeys with induced visual impairment to exhibit behavioral responses distinguishing complex movements and shapes [93]. In addition, a 24-channel surface electrodes enabled a blind patient to recognize simple letters [94]. In 2021, a visual restoration system implanted in a patient was reported to maintain visual function at a level sufficient to discriminate movements and shapes for six months [95]. In 2022, another study demonstrated that dot-like phosphenes spanning a few degrees of visual angle could be used to support motion recognition [96]. These findings suggest that cortical stimulation–based approaches have thus far demonstrated the highest potential for functional visual recovery among currently investigated visual restoration technologies. More recently, a study implanted a 100-electrode intracortical array into the visual cortex of two blind volunteers and demonstrated stable simultaneous stimulation and recording over a 6-month period. Neural activity recorded from neighboring electrodes significantly predicted phosphene threshold, perceived brightness, and the interstimulus interval required to distinguish separate phosphenes, further supporting the feasibility of closed-loop cortical visual prostheses [97] (Fig. 2f-g).

Technologies for somatosensory restoration

Somatosensation is a fundamental sense for perceiving the external environment and the internal state of the body. It is generated by a variety of receptors distributed throughout the body, including the skin, muscles, and joints, each of which detects specific sensory information. The detected somatosensory information is transmitted through afferent fibers in the peripheral nerves, passes through the dorsal roots of the spinal cord, and is ultimately conveyed to the somatosensory cortex of the brain, where it is processed [98]. Damage to any component involved in somatosensation can impair or even lead to the loss of this sensation. However, because somatosensory dysfunction is usually restricted to localized regions of the body, surrounding sensory inputs can compensate for the lost sensation. As a result, the impact on daily life and the severity of this impairment are generally lower than visual or auditory impairments [99]. Accordingly, most current somatosensory restoration technologies have mainly focused on patients with severe functional loss, such as amputees and individuals with paralysis.

However, somatosensory restoration is important for such patients not only for restoring lost sensation, but also for recovering lost motor functions [100]. For these patients, the restoration of lost motor functions is of higher priority, which derived the development of various prosthetic hands and legs. However, simply providing motor function only, without sensory feedback, has major limitations in generating natural and precise movements. One of the main reasons is the lack of sensory information associated with movement [101]. Somatosensation provides diverse information related to movement, including joint and muscle position, vibration generated during movement, and tactile input, all of which serve as precise feedback for motor control. Accordingly, somatosensory restoration technologies have been largely focused on restoring natural and precise movement by restoring the sensation that is evoked by the motor function provided to these patients [100, 101].

Approximately 58 million amputees and over 20 million individuals with spinal cord injury are estimated worldwide [102, 103]. The need for somatosensory restoration continues to grow to support their functional recovery and return them to daily life. Accordingly, various somatosensory restoration technologies have been developed, which can be broadly divided into brain approaches and peripheral nerve approaches (Table 3) [104–109].

Table 3.

Comparison of cortical and peripheral nerve approaches for somatosensory restoration

Feature	Cortical approach	Peripheral nerve approach
Primary target and clinical reach	Hand/arm representation in primary somatosensory cortex (S1)	Residual peripheral nerves Upper limb: median and ulnar nerves Lower limb: sciatic, tibial and/or peroneal nerves
Spatial specificity / projected field	Percept extent: median single-electrode projected-field (PF) area, 2.5 cm² (central 90% range, 0.3–11.3 cm²)	Upper limb: Cuff/FINE interfaces generally evoke broader hand-level percepts, whereas TIME/USEA can provide digit-level or multi-location percepts. Lower limb: Intraneural and multi-contact cuff interfaces can evoke separable toe, plantar, heel, and calf-level percepts. Spatial specificity remains interface and anatomy dependent.
Representative stimulation hardware	Typically, two Utah arrays (60–64 electrodes per array) are implanted in S1	Typically, 1–2 peripheral nerves are targeted using one or more interfaces. Cuff/FINE interfaces provide ~ 4–16 contacts per cuff. TIME interfaces provide 14 active sites per implant and are commonly deployed as 2–4 implants (~ 28–56 sites). USEAs provide ~ 96–100 microelectrodes per array and are typically implanted as 1–2 arrays (~ 96–200 sites).
Percept quality and naturalness	More similar to mechanical touch in perceived quality and naturalness. Potentially advantageous for naturalistic haptic exploration and conveying richer object properties.	Produced stronger, more salient, and more reliable percepts with lower trial-to-trial intensity variability in the same direct comparison. Biomimetic encoding can further improve naturalness.
Long-term usability	Projected-field mapping remained stable over approximately 2–7 years in the cited human cohort	Cuff/FINE: stable sensory percepts for up to 16–24 months lower-limb C-FINE studies included chronic testing beyond 1 year TIME: 6 months USEA: 4–5 weeks in the cited human study
Key limitations	Limited usable intensity range under per-electrode charge-safety constraints Trial-to-trial variability in percept detection and intensity, including occasional missed percepts Projected-field extent or location can change with stimulation magnitude Percept naturalness remains imperfect and varies with intensity, quality, and cortical-state-dependent responsiveness.	Spatial selectivity depends strongly on nerve anatomy, fascicular organization, and implantation level; proximal or mixed-fascicle recruitment can produce broad or diffuse percepts. Conventional stimulation can evoke abrupt, buzzing, vibrating, electrical, or joint-movement sensations. Perceived naturalness depends on intensity, quality, encoding parameters, and user-specific interpretation rather than stimulation modality alone. Chronic stability of percept location, intensity, and usable stimulation sites remains variable across interfaces.
Surgical access / invasiveness	Surgery to place electrodes in S1 through the skull More invasive and carries a higher risk of serious complications.	Surgery to place electrodes around or into a peripheral nerve Less invasive overall, but may cause local nerve damage or irritation.
References	[111, 120, 171]	[129, 132, 136, 139, 141, 171, 206, 207]

S1 primary somatosensory cortex, PF projected field

Cortical approaches for somatosensory restoration

Early studies demonstrated the feasibility of artificial proprioceptive feedback through micro-stimulation of the proprioceptive cortex in non-human primates, highlighting the necessity of sensory feedback in the field of brain-machine interface (BMI) or brain-computer interface (BCI) [110]. As efforts were put into human clinical trials, it was demonstrated that ICMS of the primary somatosensory cortex (S1) in individuals with spinal cord injury could evoke naturalistic pressure sensations in the fingers, which remained stable over several months [111].

Since the early 2020s, the focus has shifted toward improving precision and usability, specifically through the optimization of stimulation parameters and performance improvements through biomimetic encoding. The lower limits of detectable stimulation frequency for human subjects were identified, enabling the specification of optimal parameter ranges for practical BCI systems [112]. Long-term implants lasting over two years enabled the development of detailed sensory maps of the fingertips, significantly improving spatial resolution [113]. Multi-channel ICMS has been reported to enhance performance of stimulation by improving reaction times and reducing detection thresholds [114]. Studies have indicated that the stimulation on somatosensory cortex evokes task-dependent responses of motor cortex neurons, with both the magnitude and direction of the responses differed by the tasks (Fig. 3-a) [115], and that dynamic amplitude modulation can improve sensory detection performance by reducing neural adaptation [116].

Fig. 3.

Cortical approaches for somatosensory restoration. a Evoked task-dependent responses of motor cortex [115]. Adapted from Shelchkova et al. (2023), licensed under CC BY 4.0. b Advanced biomimetic ICMS patterns [117]. Adapted from Hobbs et al. (2025), licensed under CC BY 4.0. c Stimulation parameters to find optimal sensations for virtual objects [119]. Adapted from Verbaarschot et al. (2025), licensed under CC BY 4.0. d Single- and multi-channel ICMS patterns of different charges and reported description of evoked sensation [121]. Adapted from Bjånes et al. (2025), licensed under CC BY 4.0

Recent studies have focused on the qualitative enhancement of sensation and long-term safety and stability of ICMS. The naturalness of artificial touch has been maximized through advanced biomimetic stimulation patterns, improving both the precision and utility of sensory restoration systems (Fig. 3-b) [117]. Additionally, subthreshold stimulation was reported to enhance tactile sensitivity in the physical skin [118], and a customized paradigm was introduced where users directly adjust stimulation parameters to find optimal sensations for virtual objects (Fig. 3-c) [119]. Optimal stimulation strategies have been established by confirming the long-term stability of projected fields during multi-channel stimulation [120] and deriving safe charge density guidelines (Fig. 3-d) [121]. The long-term safety and efficacy of ICMS have been validated through a comprehensive synthesis of human clinical data spanning the past decade [122].

Peripheral nerve approaches for somatosensory restoration

Somatosensory signals from mechanoreceptors are transmitted through peripheral nerves [98]. Unlike tactile information that has been integrated and processed in the brain, these peripheral nerve signals represent more primitive sensory inputs that directly reflect the individual components underlying somatosensation [123]. For this reason, the peripheral nerve approach has the advantage of being able to address individual sensory components more clearly and distinctly. Moreover, peripheral nerves are easier to stimulate selectively at specific sites than the brain, and patients tend to experience less psychological resistance to this approach [123]. Building on these advantages, peripheral nerve-based somatosensory restoration has steadily advanced since early studies emerged in the 1970s [124–126].

Modern sensory feedback research began with the flat interface nerve electrode (FINE), an extraneural electrode, in 2002 [127], and with the longitudinal intrafascicular electrode (LIFE), an intraneural electrode, in 2005 [128]. As this field progressed, two landmark studies using FINE and the transverse intrafascicular multichannel electrode (TIME) were reported in 2014 [129, 130]. Although different electrodes were used, both studies integrated tactile sensors into a myoelectric prosthetic hand and delivered sensory feedback through electrodes, enabling upper limb amputees to evoke multiple sensations, distinguish object shape and stiffness, and achieve more precise prosthetic control (Fig. 4a-c). Thereafter, studies have focused on optimizing stimulation parameters, evaluating the functional utility of restored sensations, and verifying long-term stability [131–135]. In addition, sensory feedback research has expanded beyond the upper limb to lower limb control applications (Fig. 4d-e). Since 2019, studies with lower limb amputees have reported improvements in walking speed, gait and postural stability, reduced metabolic cost, and significant alleviation of phantom limb pain [136–140].

Fig. 4.

Peripheral nerve approaches for somatosensory restoration. a–c Closed-loop somatosensory feedback system for the upper limb using a myoelectric prosthetic hand and FINEs, locations of sensations evoked by upper limb sensory feedback, and representative improvement in object grasping performance. In c, sensory feedback is turned off on the left and turned on on the right [129]. Adapted with permission from Tan et al. (2014). d–f Closed-loop somatosensory feedback system for the lower limb using a prosthetic leg and TIMEs, together with the locations and types of sensations evoked by lower limb sensory feedback [140]. Adapted with permission from Petrini et al. (2019). g Stimulation parameters used in conventional electrical sensory feedback [129]. Adapted with permission from Tan et al. (2014). h Biomimetic stimulation patterns used for sensory feedback [143]. Adapted from Valle et al. (2024), licensed under CC BY 4.0.

However, these technologies still showed limitations in providing natural sensation (Fig. 4f). To address this, researchers using the intraneural approach introduced biomimetic stimulation strategies that mimic the firing patterns of tactile receptors (Fig. 4g-h). In the upper limb, biomimetic stimulation strategies yielded promising results using TIME electrodes in 2018 [141], and using Utah slanted electrode array in 2019 [142]. Biomimetic stimulation technology was also extended to lower limb amputees, with reported improvements in motor function through more natural sensory feedback [143].

In parallel, regenerative approaches have also been explored. Targeted muscle reinnervation (TMR) [144], which redirects residual nerves to residual muscles in amputees, and regenerative peripheral nerve interface (RPNI) [145], which inserts residual nerve endings into free muscle grafts to induce reinnervation, were originally developed to amplify neural signals into muscle signals and improve the precision of prosthetic limb control. However, during the application of these techniques, an unintended partial restoration of sensation was observed [144, 146, 147]. To maximize this sensory restoration effect, techniques such as targeted sensory reinnervation (TSR), which redirects residual sensory nerves to the skin to induce phantom limb sensation [146, 148, 149], and dermal sensory RPNI (DS-RPNI), which inserts sensory nerves into dermal grafts to induce reinnervation of mechanoreceptors, have been reported [150]. In addition, the agonist-antagonist myoneural interface (AMI) technique has recently seen growing clinical applications [151, 152]. AMI focuses on restoring proprioception rather than tactile sensation, and is designed by mechanically linking agonist and antagonist muscles so that contraction of the agonist induces stretching of the antagonist, and vice versa. Through this mechanism, movement of the prosthetic limb actually induces muscle contraction and stretching, allowing proprioceptive sensation to naturally arise [152, 153]. AMI has shown meaningful improvements in the control of prosthetic legs [154, 155]. These regenerative approaches can be applied clinically without any additional electronic equipment, and have contributed to improving patients’ quality of life shortly after their introduction [156–158]. On the other hand, since the outcome of sensory restoration depends on the pattern of nerve regeneration, precise control is difficult, achieving complete sensation remains challenging, and as with AMI, these approaches are limited to targeting specific sensory modalities [107, 151, 159].

Future directions in sensory restoration technologies

Cochlear approaches

Although cochlear implants provide significant auditory restoration for individuals with severe hearing loss, their limitations highlight the need for next generation approaches capable of delivering more natural and comprehensive auditory experiences. Music perception is one of the biggest challenges to CIs because of the complex nature of musical elements such as pitch, melody, rhythm and timbre [8]. Studies investigated how CI users perceive music and have shown that CIs are yet capable of delivering musical experiences intact. CIs altered most fundamental musical features, including pitch, timbre, melody perception, complex rhythm, and duration [160]. Advances in the number of channels and stimulation algorithms will be necessary to express complex auditory environments. However, the performance of implants may not be the only factor for better auditory sensation. It has been reported that rehabilitation and timing of implantation are crucial factors for successful hearing restoration [161, 162]. Speech assessment scores improved across all age groups after cochlear implantation, but the group of earlier implantations showed superior outcomes. Furthermore, hearing and speech performance showed progressive improvement with increasing duration of post-implant rehabilitation [162]. These studies imply that the nervous system’s adaptation and learning stimuli from the CIs are important for auditory restoration. Advancements in rehabilitation and stimulation protocols should be followed by enhancements in device performance. Personalized stimulation or rehabilitation programs using artificial intelligence (AI) can be one of the solutions [163].

The most promising next generation CIs are totally implantable cochlear implants (TICI). TICIs represent an emerging solution for individuals with severe-to-profound hearing loss, characterized by fully internalized components that eliminate external hardware and improve aesthetic discretion. While this approach offers clear advantages in terms of comfort and invisibility, several challenges remain, including surgical risks, high costs, patient selection limitations, variable adaptation periods, potential constraints in sound quality, device reliability issues, battery replacement requirements, and the irreversible nature of the procedure [14]. Despite the difficulties, TICI has reached early clinical validation but has not yet achieved commercial availability due to technical and regulatory challenges [164].

Retina approaches

Just as CIs replace damaged auditory receptor cells, retinal prostheses can replace photoreceptors whose function has been lost under pathological conditions. If they can provide stimulation that sufficiently and precisely mimics photoreceptors, they have the theoretical potential to achieve the most natural form of visual restoration. For this reason, retinal prostheses have long been studied as a representative strategy for visual restoration.

However, current retinal prostheses remain largely limited to basic functions such as phosphene perception, object localization, and simple shape discrimination [63, 68, 69, 71]. Artificial vision elicited by retinal stimulation is experienced not as continuous and natural scene perception, but rather as a combination of spatially localized phosphenes whose brightness, size, shape, and location vary depending on the stimulation site and intensity. Clinically, the epiretinal Argus II system enabled some patients to detect and localize high-contrast objects, recognize letters, and read words [65], whereas the subretinal Alpha AMS study reported improvements in light perception, temporal resolution, object localization, or grating discrimination in five of six participants [68]. However, these outcomes represent the restoration of low-resolution visual function under constrained conditions and remain substantially different from the complex scene analysis, general object recognition, and continuous spatial perception supported by natural vision.

This limited functional outcome reflects the distinct spatial organization and processing complexity of the retina. Unlike the cochlea where receptor cells are organized along a relatively orderly tonotopic axis, photoreceptors in the retina are distributed across a complex two-dimensional mosaic, making precise spatially-patterned stimulation much more important [156, 157]. In addition, visual information is initially processed within the retina by multiple cell types such as amacrine cells, ganglion cells, and other retinal cells, making the overall processing much more complex [158].

To overcome these limitations and achieve more natural retina-based artificial vision, much higher-density electrode technology is required. Photovoltaic retinal prostheses can simplify the system by directly generating current from external light, and on this basis, they have shown promise for high-density integration. However, as photovoltaic retinal prostheses become more densely integrated, the current generated at each pixel may become insufficient to effectively stimulate the retina, thereby imposing an inherent limit on further increases in density [168]. Therefore, photovoltaic retinal prostheses must also advance in a direction that can improve both current efficiency and pixel density at the same time to enable higher-resolution artificial vision.

In addition, whether photovoltaic or conventional, subretinal approaches need to advance toward thinner and more biocompatible forms. This approach aims to directly interface with bipolar cells in place of photoreceptors and is positioned between retinal layers. Therefore, the implanted structure must be sufficiently thin to minimize physical damage to surrounding retinal cells, while also ensuring long-term biocompatibility. It must also be able to establish a stable interface through closer contact with bipolar cells.

In contrast, epiretinal approaches aim to stimulate ganglion cells, which represent the final output stage of the retina. The signals targeted here already reflect visual information after primary processing within the retina. Therefore, epiretinal stimulation cannot rely only on matching the location of incoming light, but must instead be designed to reflect how the retina actually processes visual information and converts it into ganglion cell activity. This requires prior understanding of the principles underlying primary visual processing in the retina, and more refined stimulation strategies must be developed on the basis of that understanding [169]. For this, devices capable of long-term recording of ganglion cell activity in vivo must first be developed. Once this foundation is established, epiretinal approaches could advance into more sophisticated forms that more precisely reflect the visual processing principles of the retina.

Furthermore, implementing such refined stimulation strategies requires image encoding algorithms that go beyond mapping camera-acquired images onto electrodes by simple brightness values, instead preserving the spatial features critical to retinal information processing within a limited number of electrodes. One approach combines a convolutional neural network–based forward model that predicts retinal ganglion cell responses with an actor network that learns image compression; in silico and in ex vivo mouse retina, it preserves natural retinal responses better than conventional pixel averaging when high-resolution images are converted to a limited resolution [169]. Another strategy uses explainable AI to preferentially extract image regions important for object recognition and to selectively activate the corresponding electrodes, computationally reducing unnecessary electrode activation and the expected power consumption without substantially compromising recognition performance, compared with conventional grayscale-based activation [170]. Together, these studies show that improving the quality of artificial vision in retinal prostheses requires not only increased electrode density but also AI-based encoding strategies that reflect the information-processing characteristics of the retina and the behaviorally relevant image information. Such encoding strategies are not limited to retinal prostheses; they are equally applicable to cortical approaches, where camera-acquired images must likewise be encoded into stimulation through a limited number of electrodes. However, these results so far rest on ex vivo retinal responses to projected images and on phosphene simulation with deep-learning-based recognition evaluation; future validation under actual electrical stimulation conditions and in implanted devices will therefore be needed [169, 170].

Cortical approaches

As it directly targets the final organ responsible for sensory processing, brain-based sensory restoration offers broader clinical applicability than approaches targeting individual sensory organs. For example, in vision, this approach may be applicable even to patients with optic nerve damage, such as those with glaucoma. In somatosensation, this approach may be applicable to patients with paralysis caused by spinal cord injury, a condition that cannot be addressed by peripheral nerve–based approaches.

In somatosensory cortical restoration, the translation of tactile sensor signals into stimulation patterns has so far relied not on large-scale AI models but on relatively simple linear encoding or on biomimetic encoding that reflects physiological response characteristics. Early approaches mainly used linear encoding, in which the contact location detected by sensors on a prosthetic or robotic hand is mapped to the cortical electrodes that evoke sensation in the corresponding part of the hand, and the stimulation amplitude or frequency is increased in proportion to the measured contact force [120]. This is a relatively direct mapping in which stronger contact is converted into higher-intensity stimulation and weaker contact into lower-intensity stimulation. Biomimetic encoding goes a step further, embedding into the stimulation pattern the way the somatosensory response evolves over time during contact. In natural touch, a strong transient response arises at the onset and offset of contact, whereas a lower and more stable sensation persists while contact is maintained; biomimetic encoding reproduces precisely this property [116]. That is, by using not only the absolute force measured by the sensor but also its temporal variation, it delivers brief, high-amplitude and densely packed stimulation at the onset and offset of contact, while maintaining low-level stimulation during the sustained hold phase. Through this onset–hold–offset design, both the transient and sustained components of natural contact are reproduced. Some studies have further incorporated this onset–hold–offset structure into changes in stimulation frequency and in the combination of active electrodes, as well as amplitude, seeking to realize spatiotemporal stimulation patterns closer to natural contact than simple linear stimulation [117].

Despite these efforts, the level of sensation achieved by current brain-based somatosensory restoration remains limited, and it has not yet progressed far beyond the induction of unnatural or localized sensations [108]. ICMS can evoke relatively localized tactile percepts in specific regions of the fingers or palm. In early human studies, these sensations were most commonly described as pressure applied to a particular region of the hand, while some stimulation sites evoked sensations resembling light touch or warmth [111]. In contrast, tingling and electrical sensations were also elicited at some electrodes; these sensations were described as being more similar to mechanical vibration than to electrical stimulation applied to the skin surface [111]. In addition, increasing stimulation amplitude was reported to increase perceived intensity, and stimulation at different sites could evoke stable sensations in distinct regions of the fingers and palm [111, 113, 120]. More recently, biomimetic stimulation patterns designed to reflect the temporal activity observed in the somatosensory cortex during natural contact have been explored to improve the naturalness of artificial touch. Such biomimetic patterns were rated as more similar to mechanical pressure stimulation than conventional linear stimulation at some individual electrodes or multi-electrode combinations [117]. In addition, a study in which users directly adjusted stimulation amplitude, frequency, and spatiotemporal stimulation patterns reported that participants could construct stimulation patterns corresponding to different virtual objects and discriminate some objects above chance level even in the absence of visual cues [119]. However, these approaches remain limited to conveying selected tactile features, such as pressure, warmth, or surface properties, through a restricted number of electrodes and pre-explored stimulation parameters. They still cannot fully reproduce the temporally evolving contact information, spatial distribution of force, fine texture, and integrated sensory submodalities that characterize natural touch during real hand manipulation. A recent single-participant direct-comparison study further reported that ICMS evoked more focal percepts and sensations closer to natural touch than peripheral nerve stimulation, suggesting that cortical stimulation may offer a qualitative advantage in approximating natural tactile experience [171].

A comparable limitation is observed in visual cortical stimulation, in which complex visual perception has yet to be restored and current systems primarily convey visual information through basic phosphene percepts [93, 95]. A study using a 96-channel intracortical array reported that phosphene-based stimulation enabled the identification of some letters and object boundaries [95]. However, stimulating multiple electrodes simultaneously does not necessarily cause individual phosphenes to automatically integrate into a coherent visual image in the same manner as pixels on a display. Instead, concurrent stimulation can make individual phosphenes difficult to distinguish or evoke larger amorphous percepts. Therefore, generating meaningful visual forms requires not only increasing the number of electrodes but also precisely controlling stimulation locations and temporal sequences [95]. As an alternative stimulation paradigm to overcome these limitations, a strategy combining dynamic sequential stimulation and current steering has been proposed to move the stimulation location over time [94]. Under this approach, a blind participant reported perceiving a sensation resembling “a line being drawn” and was able to reproduce or distinguish letter-like patterns corresponding to different stimulation sequences. In particular, five randomly presented patterns were verbally identified with 93% accuracy (14/15 trials). However, these outcomes were obtained using pre-mapped phosphenes and a limited set of simple geometric patterns, and therefore should not be interpreted as restoration of the complex spatiotemporal structure of natural scenes or general object recognition.

To overcome these limitations and restore more natural sensations, advances in electrodes are essential. Electrodes technologies have been advanced to achieve high-density integration [172–174], long-term stability [175–177], and wireless operability [178, 179]. However, the problem is that much of this progress has been concentrated on electrodes that position multiple channels at a single localized site in the brain, such as shank-type electrodes [172–178]. These electrodes offer major advantages for recording brain activity, but they also have inherent structural limitations for neuroprosthetic applications such as sensory restoration and neurorehabilitation. It is because shank-type electrodes interface with the brain only at a specific site, making it difficult to represent the distributed neural activity that underlies brain function. In the cerebral cortex, sensory information is not processed at a single location. Instead, different cortical columns within the same cortical region process different types of information, and their integration gives rise to the specific sensations we perceive [180–182]. For example, in the visual cortex, there are orientation columns, each tuned to a different direction [182]. Therefore, shank-type electrodes, which can interface with only a single column at a single location, are not well suited for achieving fine sensory perception.

Accordingly, sensory restoration studies have relied on large-area electrodes that can access multiple cortical columns simultaneously. However, large-area electrodes for sensory restoration still largely rely on electrode platforms developed more than two decades ago, such as low-density ECoG electrodes and the Utah array. In particular, intracortical electrodes for high-resolution sensory restoration require three-dimensional fabrication, which makes the fabrication process highly challenging, and as a result, no clear next-generation alternative to the Utah array has yet been developed. As a result, the large-area electrodes used for sensory restoration still face a range of technical limitations associated with electrodes, including limited channel count and spatial resolution, foreign body responses, and reduced long-term stability. Therefore, further advances in sensory restoration will require the development of next-generation large-area electrodes that can overcome the limitations of current approaches.

In addition, effective sensory restoration requires a deeper understanding of how the target sensation is processed in the brain. One of the key limitations repeatedly emphasized by leading researchers in the field of sensory feedback is the lack of understanding the neural language underlying sensory perception [101, 109]. In the past, sensation was mainly evoked through simple stimulation [110]. More recently, however, the focus has been shifting toward biomimetic stimulation aimed at reproducing more natural sensory experiences [119, 120]. As a result, the need for comprehensive understanding of sensory processing principles has become even greater. Ultimately, future sensory restoration technologies should advance toward decoding-based sensory feedback that decode the real-time formation of individual sensations and implement precise stimulation based on this information. Such progress will deepen our understanding of how sensations are generated in the brain and enable the restoration of more natural sensations in patients with sensory impairment.

Peripheral nerve approaches

Through peripheral nerves, we can access raw signals of the sensory receptors involved in somatosensation [123]. Compared with brain signals, peripheral nerve signals are more intuitive and easier to decode the location of their origins and firing patterns. Therefore, peripheral nerves may be suitable targets for naturalistic somatosensory restoration.

As in cortical approaches, peripheral nerve–based somatosensory restoration has so far relied not on large-scale AI models but on biomimetic encoding that reflects the physiological responses of sensory receptors and afferent fibers. Early methods directly mapped the pressure or force signals from tactile sensors onto stimulation amplitude or frequency, whereas more recent approaches first convert the pressure distribution and its temporal changes measured by the sensor into the expected firing patterns of mechanoreceptor afferents, and then reproduce these patterns in the temporal structure of the peripheral nerve stimulation train [141–143, 183]. FootSim, for example, is a model that predicts the firing activity of different mechanoreceptor afferents from the pressure distribution and skin deformation of the sole [184]. Such predicted firing patterns can be converted into stimulation patterns that reflect the sensory information changing during contact, weight loading, and lift-off in gait, and they have been used to provide more natural sensory feedback in lower-limb amputees [143].

Such peripheral nerve stimulation–based sensory feedback has already provided tangible functional benefits during prosthetic limb use, including the detection of object contact, force modulation, discrimination of object shape and stiffness, and improved gait stability [129–140]. However, these functional gains do not necessarily amount to a faithful reproduction of natural touch. Conventional peripheral nerve stimulation can evoke sensations such as tingling, buzzing, vibration, electrical sensations, or paresthesia alongside percepts of contact or pressure. Furthermore, the perceived location and spatial extent of the evoked sensation are determined by the recruited nerve fibers and stimulation parameters, which can result in a broad projected field. A recent head-to-head comparison likewise found that conventional peripheral nerve stimulation elicited a larger sensory area and lower naturalness than cortical stimulation, suggesting that current peripheral nerve stimulation does not fully reproduce the spatial and qualitative characteristics of natural touch [171]. Even the biomimetic strategies described above, which incorporate the firing characteristics of tactile afferents, have improved the naturalness of evoked sensations, tactile sensitivity, and manual dexterity in peripheral nerve interfaces [141–143], yet they remain fundamentally model-based: current biomimetic stimulation patterns are largely constructed from computational models of receptor firing, external sensor signals, or predefined stimulation rules, and do not reflect, in real time, the sensory signals actually arising in an individual patient’s peripheral nerves [143, 183, 184].

To further improve the naturalness and selectivity of evoked sensations, therefore, it is necessary to go beyond merely mimicking receptor firing patterns and to directly measure and decode how each sensory input is encoded in the peripheral nerve itself. This would make it possible to understand how each sensory receptor contributes to sensory formation, and such understanding would provide the basis for more refined biomimetic stimulation. Furthermore, to maximize the potential of peripheral nerves in somatosensory restoration, peripheral nerve sensory signals must be decoded at the level of individual axons, and precise axon-level stimulation must be implemented accordingly. Because this approach more directly mimics the language of sensation, it could ultimately lead to sensory restoration that is closer to real sensation.

However, a major technical challenge here is the difficulty in recording axon-level signals in vivo from peripheral nerves. Peripheral nerves are highly susceptible to physical and physiological noise, their surrounding environment is highly dynamic, and the implantation is challenging due to the anatomy [100]. As a result, electrodes capable of recording in vivo axon-level signals from peripheral nerves are very limited. Because of these limitations, current studies on peripheral nerve-based sensory feedback often mimic the firing patterns of sensory receptors using simulation instead of actual recording [143, 183, 184]. However, simulation alone cannot sufficiently reflect the actual firing patterns of sensory receptors. Accordingly, the need for high-resolution peripheral nerve signal recording has been emphasized [100].

Therefore, for better somatosensory feedback in the future, electrodes capable of long-term axon-level signal recording in peripheral nerves must first be developed. These electrodes should then be advanced to perform recording and stimulation simultaneously. Ultimately, decoding-based stimulation technologies need to be developed to decode individual somatosensory signals and deliver stimulation based on that information. Natural somatosensory restoration in patients with complete loss of somatosensory perception will become possible only when this technological foundation is established.

System-level challenges for fully implantable sensory restoration

In fully implantable sensory restoration systems, not only electrode performance and stimulation strategies, but also sensory input processing, generation of stimulation commands, wireless communication, and power supply must be integrated within a single system. In particular, in high-channel neural interfaces, the scale of signal processing and data transmission increases with the number of electrodes; therefore, it is important to efficiently integrate electrodes, signal processing circuits, data telemetry, and wireless power supply [179]. In sensory restoration systems, multiple tactile or visual inputs must be processed in real time and then compressed and converted into a limited number of stimulation commands. Thus, information bandwidth and low-power processing remain important technical challenges [109]. A reliable energy supply is also a key limiting factor in fully implantable devices. For example, in the development of totally implantable cochlear implants, sufficient power delivery and the implementation of implantable microphones have been identified as major technical challenges [14]. Although recent totally implantable cochlear implants have demonstrated safety and hearing performance close to that of conventional devices in early clinical studies, further technologies are required to more efficiently integrate sensory input, signal processing, stimulation, and power supply in order to extend such systems to a broader range of sensory prostheses [109, 164].

Declarations

Conflict of interest

The authors declare that they have no conflict of interest.