Bypass neural interfaces for paralysis: clinical translation, challenges and future directions

Abstract

Advances in neural recording, real-time decoding and bioelectronic stimulation have enabled a new class of systems that re-establish functional communication across disrupted neural pathways. In the context of paralysis, these neural bypass interfaces link upstream neural intent to effector activation downstream of the lesion, effectively circumventing sites of injury within the nervous system to restore volitional movement. In this review, we define neural bypass interfaces as an emerging category of bioelectronic medicine, distinct from conventional brain–computer interfaces and neuromodulation technologies. We first outline key neurophysiological and systems-level considerations underlying bypass design, before tracing their evolution from bench to bedside. We then focus on the clinical translation challenges that govern real-world deployment, specifically signal stability and fidelity, stimulation performance, decoding robustness, closed-loop integration and long-term implant viability. Importantly, this review highlights two emerging directions that may shape the next generation of neural bypasses: the use of the spinal cord itself as a source of neural intent, and the development of bidirectional bypasses integrating sensory feedback to enable more adaptive, physiologically aligned control. Ultimately, neural bypasses may go beyond simply restoring movement to drive biological recovery, redefining neurorestorative therapies for paralysis.

Key points

• Neural bypasses reroute neural intent around lesions to restore volitional movement, with most current systems taking brain-to-muscle or brain-to-cord-(to-muscle) pathways.

• Recording and decoding challenges include unstable signal fidelity, biological signal drift, and trade-offs between signal complexity and latency, ultimately hindering real-time control. Using phase-aware strategies may help preserve synchrony.

• Stimulation and implementation barriers include limited target selectivity, parameter optimisation and poor long-term usability across diverse patients.

• Key future directions include bidirectional sensory-integrated bypasses for adaptive closed-loop control, more precise recreation of sensory percepts, and spinal cord–machine interfaces.

Introduction

Neural interfaces as an emerging neurorestorative strategy

Neural interfaces have recently emerged as a potential strategy for restoring neurological function in humans. By circumventing disrupted neural circuity, their applications are far-ranging: including the restoration of sensory percepts (body-to-brain) [1–4], vision [5, 6], olfaction [7, 8], autonomic functions such as bladder control [9, 10], as well as sensorimotor and speech restoration [11–13]. Among these applications, restoration of motor function through intrinsic limb reanimation or extrinsic prostheses control remains the most clinically mature use case, thus forming the primary focus of this review. In these systems, neural activity from intact upstream structures may be recorded, and decoded in real time to control external actuators (e.g., robotic arms) or stimulate the patient’s own nerves, spinal circuits or muscles [11, 14–16]. Other classes of neural interfaces operate exclusively through direct stimulation without recording or real-time signal decoding [17–20]. Of these modalities, neural bypasses for paralysis have demonstrated some of the most rapid translational progress from proof-of-concept systems to early human clinical deployment in the past decade.

Paralysis and the need for neural bypasses

Paralysis is a debilitating condition marked by the loss of voluntary motor functions, and may be accompanied by sensory and autonomic impairments depending on the site of injury. Aetiologies include cerebral accidents (e.g., stroke and brain injuries), spinal cord injuries and peripheral nerve trauma. In North America alone, 1 in 50 individuals suffer from some degree of paralysis [21]. Not only do these patients lose independence in executing activities of daily living, but the lifetime economic burden imposed ranges from $1.04 million for incomplete paraplegia to $2.08 million for tetraplegia [22].

Traditionally, neurorehabilitation has been the mainstay of treatment to reduce motor impairment and prevent complications like spasticity and pain. However, efficacy is limited in severely paralysed limbs due to restricted volitional engagement [23], which creates a need for novel therapies. Emerging restorative approaches show promise, such as virtual reality-based and robot-assisted rehabilitation shown to promote neuroplasticity [24], alongside 3D bioprinting or stem cell therapy [25, 26]. In particular, attention has increasingly shifted toward neural bypasses for patients with chronic tetraplegia due to severe or clinically complete spinal cord injury (SCI). Unlike patients with mild or incomplete injuries, who may retain sufficient residual connectivity to benefit from existing neuromodulatory interventions like spinal cord stimulation (SCS) [27, 28], patients with complete SCI are in greater need of direct restoration of communication across disrupted neural pathways [16, 29–31]. For these patients, restoring large proximal limb movements is currently more achievable than recreating high-degree-of-freedom functions such as fine hand movements, at least until the neurobiology of movement control is better understood and stimulation precision further improves.

Defining neural bypasses

Neural bypasses represent a specialised class of neural interfaces designed to restore signal transmission between disconnected segments of the nervous system. In the context of paralysis, unlike traditional neuroprosthetic systems, where decoded neural activity controls external effectors like exoskeletons or robotic arms, neural bypasses reroute signals from the cortex, spinal cord or peripheral nerves back into the patient’s own nervous system or musculoskeletal apparatus to stimulate intact native pathways [32]. In doing so, neural bypasses have the potential to produce more naturalistic motor output with multiple degrees of freedom as compared to external prostheses that tend to produce coarse movements instead [32]. Moreover, they can mitigate muscle disuse atrophy and promote activity-dependent neuroplasticity for long-term functional recovery [33]. While many bypasses still rely partially on external hardware for signal acquisition, decoding or power delivery, their use of functional neural tissue at both the recording and stimulation sites fundamentally distinguishes them from traditional neuroprosthetic approaches, where either site is replaced with an artificial organ or limb (Fig. 2).

Fig. 2.

Illustration of stimulation targets and related modes of stimulation in neural bypasses. Created with biorender.com. SCS: Spinal Cord Stimulation; PNS: Peripheral Nerve Stimulation; FES: Functional Electrical Stimulation

That is why many body-to-brain applications outside of paralysis are more appropriately classified as neuroprostheses rather than true neural bypasses, because the primary pathology lies in loss of the sensory apparatus itself rather than interruption of signal transmission. This is seen in visual (subretinal and intracortical vision prostheses) [34, 35], speech and auditory prostheses [13, 36], which substitute for degenerated recording structures using artificial encoding strategies. In contrast, sensorimotor bypasses target disorders where transmission through native pathways is impaired, while upstream information and downstream targets are preserved. These systems may not merely function as a conduit to directly stimulate targets, and can also function in neuromodulatory capacities by amplifying residual signals. Thus, while neuroprosthetics broadly replace or augment neurological function, neural bypasses specifically use functional neural or muscular components upstream and downstream of the lesion to restore communication.

Neural bypasses typically comprise three core components: a sensor, signal decoder, and stimulator [37]. For cortical recording interfaces, intracortical microelectrode arrays (MEAs) are particularly well-suited for neural bypasses because they capture single-unit spiking activity with high spatial and temporal resolution [38], as opposed to less invasive approaches (e.g. electroencephalography (EEG) or electrocorticography (ECoG)) that use coarser signals like event-related potentials [39]. In paralysis applications, this enables finer motor control with greater degrees of freedom, as individual neurons encode specific movement parameters like direction, velocity and force. Moreover, higher-density intracortical recordings improve decoding fidelity and speed [40–42]. While these advantages come at the cost of substantially greater invasiveness and lower spatial coverage, they are particularly crucial for real-time volitional movement. However, emerging high-density micro-ECoG (µECoG) arrays may provide a compromise by improving spatial resolution beyond conventional ECoG while avoiding cortical penetration [43, 44]. For spinal cord applications, high-density flexible surface arrays offer the strongest potential for capturing stable, high-fidelity recordings, though no single interface has emerged as definitively superior [45–47]. Intraspinal penetrating arrays have yet to be approved for use in humans [48]. Recently, functional ultrasound imaging (fUSI) has emerged as a alternative by accounting for hemodynamic activity to decode spinal network states, potentially enabling closed-loop optimisation of spinal stimulation in future neural bypasses [49]. In the peripheral nervous system (PNS), intraneural electrodes provide superior selectivity and spatial resolution whereas extra-neural configurations afford greater chronic stability – making electrode choice highly context-dependent [50]. Figure 1 illustrates these recording targets and associated modalities that differ in invasiveness, signal fidelity and informational complexity. Meanwhile, Fig. 2 depicts representative examples of stimulation targets and methods, to highlight how downstream activation can occur at the level of muscles, peripheral nerves, or spinal cord. Previous work by Zuccaroli et al. have gone into greater detail summarising recording, signal processing and stimulation techniques alongside the pathways bridged by neural bypasses [51]. Building on this, our review further examines the neurophysiological basis of bypasses, alongside the technical and translational challenges limiting widespread adoption.

Fig. 1.

Illustration of recording sites and modalities in neural bypasses. Created using biorender.com. EEG: Electroencephalography; ECoG: Electrocorticography; MEA: Microelectrode Array

Neurophysiological considerations for neural bypasses

Although not a defining feature, neural bypasses often demand greater precision than conventional neuroprosthetic systems. In reanimating paralysed limbs, the main challenge lies in replicating natural kinematics with sufficient strength and accuracy, as even simple movements require coordination of multiple motor pools. In contrast, robotic exoskeletons or powered orthoses rotate a single joint or set of joints to a predefined angle or torque, generating movement through deterministic actuator control (e.g. proportional-integral-derivative loops) while bypassing native spinal circuits and muscle synergies that underlie natural movement [52]. As a result, movements often exhibit lower degrees of freedom and far less biological variability.

Neural bypasses, instead, rely on direct electrical stimulation of peripheral nerves, muscles or cord, contending with the complexity of the human musculoskeletal system inherently designed for fluid, adaptable motion. Most joints are controlled by multiple muscles with distinct and synergistic force vectors – a human hand exhibits over 20 mechanically significant degrees of freedom. Even seemingly simple actions, such as wrist flexion, depend on the orchestrated efforts of several forearm flexors. Replicating such movements therefore demands a high level of temporal precision and graded stimulation of specific motor units, which in turn often requires synchronisation of stimulation with defined oscillatory phases or movement states to maintain coordination.

Consequently, neural bypasses require more sophisticated control algorithms accounting for heterogeneity in motor unit size, force output, fatigue resistance, and nonlinear musculoskeletal dynamics. Movement reconstruction is further complicated by the spinal cord, which functions not merely as a passive conduit for electrical signals, but also contains baseline reflexes, and acts as an active sensorimotor processing centre capable of generating central pattern generators and complex movements even without direct supraspinal input [53, 54]. Following any degree of SCI, residual and reorganised spinal networks may therefore interact unpredictably with stimulation commands delivered by the neural bypass, producing unintended or spastic contractions. Therefore, future bypasses may consider employing neuromodulatory strategies that utilise or compensate for these spinal circuits (e.g. targeted activity-dependent stimulation of the cord) rather than attempting to override it entirely [55].

Changes following neural injuries and choice of landing sites

Understanding post-injury neural reorganisation is essential, as bypasses must adapt to, and in some cases can leverage, changes in circuitry after damage to remobilise limbs. Two key concepts in designing bypasses are Wallerian degeneration and the secondary alterations in muscle physiology after motor neuron loss.

In Wallerian degeneration, there is breakdown of axonal segments distal to the site of axonal disruption, through axonal swelling, cytoskeletal collapse and myelin degradation. In Central Nervous System (CNS) lesions, such degradation persists [56], as regeneration remains largely ineffective due to the inhibitory cues from astrocytes, oligodendrocytes and myelin-associated factors like Nogo proteins [57]. In SCI, because descending corticospinal axons degenerate antegrade to the injury site, they fail to synapse with ventral horn alpha motoneurons, necessitating bypass recording sites positioned upstream of the cord lesion like cortical areas where signal output is still viable. The chronicity of Wallerian degeneration also dictates the optimal window for bypass implantation, as prolonged degeneration and glial scarring may impair signal fidelity and interface stability. Meanwhile, in lower motor neuron injuries, denervation leads to rapid muscle atrophy, fibrosis, and loss of excitability due to the absence of trophic input [58]. For neural bypasses used in PNS lesions, this markedly narrows the number of viable downstream targets and window of time for effective intervention before irreversible muscle degeneration limits functional recovery.

Even when suitable recording and stimulation sites are identified, neural bypasses do not necessarily restore coordinated activity as neuronal populations above and below the lesion are decoupled, and may reorganise independently. The cortical and subcortical motor systems undergo both topographic and functional reorganisation by leveraging intrinsic neuroplasticity, expanding representations of spared muscles rostral to the lesion and recruiting parietal, premotor and cerebellar networks to compensate for lost outputs downstream of the lesion [59–61]. Recording sites must therefore account for altered cortical encoding instead of relying on pre-injury motor maps. At the spinal level, loss of supraspinal projections induces disordered axonal resprouting and synaptogenesis below the injury [62]. Any residual reticulospinal and propriospinal projections may also form maladaptive collaterals onto motor neuron pools [63, 64]. Appropriate neuromodulation by bypasses, such as through epidural spinal stimulation, should direct adaptive neuroplasticity to restore appropriate connectivity [65].

Information content, complexity and distribution

Informational complexity of neural signals increases in a hierarchical fashion from peripheral nerves to the spinal cord and subsequently the cortex, influencing the choice of recording and stimulation sites. In motor pathways, the primary motor cortex encodes higher-order movement plans that integrate sensory, contextual and cognitive inputs through distributed laminar–columnar microcircuit activity [66, 67]. These signals are subsequently compressed into lower-dimensional representations within the corticospinal tract, and are ultimately relayed as segmental motor commands through spinal roots to peripheral effectors [68]. Cortical-peripheral neural bypasses can leverage this natural hierarchy by decoding high-dimensional cortical action plans, and transforming them into lower-dimensional kinematic outputs based on the desired level of granularity for more precise stimulation of effectors. Cortical-spinal bypasses may further utilise intact spinal interneuronal networks to refine finer muscle synergies [69]. However, contextual inputs that inform natural signal transformation, such as from secondary descending pathways (e.g., reticulospinal, propriospinal) or spinal interneurons are often inaccessible to bypasses. This may call for complementary strategies like shared control with external sensors, or predictive algorithms that use probabilistic modelling based on biomechanical principles.

In parallel, restoration of ascending sensory pathways is a substantially more complex challenge in neural bypass design. Sensory feedback, particularly tactile and proprioceptive input, is critical for dexterous object manipulation, force modulation and embodiment during closed-loop motor control. However, unlike motor output, somatosensation comprises multiple modalities transmitted from peripheral receptors (e.g. mechanoreceptors in the skin, proprioceptors in muscle spindles, joints) through specialised ascending pathways including the dorsal column–medial lemniscal, spinothalamic and spinocerebellar tracts—before integration within the thalamus and distributed somatosensory cortical networks. As such, localised stimulation of S1 may evoke heterogeneous percepts due to overlapping cortical representations, making precise sensory reconstruction difficult.

One potential solution is to stimulate individual pathways (e.g. those specifically responsible for proprioception) in the spinal cord before they get integrated in the brain. Nevertheless, recent studies involving intracortical microstimulation (ICMS) have demonstrated reliable generation of tactile percepts such as squeezing, tapping and vibration through stimulation of S1, improving grasp modulation and object discrimination [1, 2, 4, 31, 70–72]. Force sensors can be mapped onto somatotopically corresponding S1 electrodes, with contact force encoded through stimulation amplitude. Such tactile feedback reduces reliance on visual feedback and mitigates under- or overestimation of grasp force during manipulation tasks. However, proprioceptive percepts remain more difficult to reproduce consistently, likely due to the distributed and multimodal nature of proprioceptive encoding within sensorimotor cortex [73, 74]. Higher stimulation amplitudes have been shown to preferentially evoke proprioceptive sensations, potentially through recruitment of deeper cortical regions associated with positional processing [74]. Many aspects of proprioception-specific, and more broadly modality-specific stimulation, are poorly understood and could be a future area of research.

Distributed sensorimotor processing, error correction and implications for future bypasses

These pathways form a bidirectional sensorimotor network involving cortical, thalamic, cerebellar, spinal and peripheral circuits that continuously transform and integrate motor and sensory information (Fig. 3). Existing neural bypasses have primarily recreated selective components of this network, including decoding movement intent from M1 and PMC, and restoring modality-specific sensory percepts from S1. However, current M1 recording interfaces incompletely capture the laminar organisation of sensorimotor processing. Among the 6 cortical layers, Intracortical MEAs often record from a single layer (typically layer V responsible for corticospinal projections) while ECoG recordings are dominated by deep-layer activity, both largely missing superficial layers involved in cortico-cortical integration of premotor and sensory inputs [67, 75–78]. Future neural bypasses may thus require multi-layer recording approaches (e.g. neuropixel probes) to better capture the full hierarchy of motor planning, sensory integration and feedback processing [79, 80].

Fig. 3.

Sensorimotor pathways relevant to neural bypass interfaces

Furthermore, because intermediate pathways and structures are bypassed, several important modulatory and error-processing functions remain incompletely represented. Current systems do not adequately account for cerebellar, reticulospinal and vestibulospinal contributions to posture, balance and muscle tone, nor sensory prediction error computations executed by the cerebellum in which predicted sensory consequences of movement are continuously compared against incoming proprioceptive and visual feedback to calibrate motor output [81, 82]. Current bypasses similarly do not account for basal ganglia functions involved in suppressing unwanted movements and reinforcing successful motor actions through dopaminergic reward prediction error signalling [81, 83]. Importantly, while some of these computations may be partially embedded within M1 output through existing cortico-cerebellar and cortico-subcortical connectivity, many modulatory pathways operate through parallel circuits that are not captured by conventional bypass architectures. As the aforementioned error signals further shape subsequent motor adaptation, their disruption can impair movement stability and progressively degrade decoding accuracy over time [84, 85].

Sensory gating is similarly underexplored in current neural bypasses. From peripheral receptors to S1, sensory information undergoes extensive filtering and transformation which changes depending on movement states [86]. Within the spinal cord, GABAergic interneurons mediate presynaptic inhibition, selectively suppressing or permitting sensory transmission according to movement state, proprioception and descending motor commands [87]. Disruption of these circuits can produce severe oscillatory reaching movements [88]. The thalamus further regulates sensory inflow through graded thalamocortical inhibition and integration of convergent cortical and subcortical inputs [89, 90]. Beyond relaying information to S1, thalamic pathways shape sensory inflow to M1 through indirect and direct routes. Sensory information from S1 can be conveyed to M1 through transthalamic pathways involving higher-order thalamic nuclei [91], alongside direct corticocortical S1–M1 projections. In parallel, proprioceptive information reaches M1 independently of S1 through much faster cerebello-thalamo-cortical pathways, including transcortical long-latency stretch reflex circuits [92, 93]. Preserving these sensory gating pathways, particularly for proprioception, are important because proprioceptive prediction error computations that occur within deep layers of M1 aids in rapid movement correction, and even long-term stability [94–96].

In short, existing gaps in neural bypasses arise from not just a lack of sensory input, but disruption of multiple interacting sensory and error-processing loops that modulate motor output. Although the brain can still generate error-related cortical potentials when a bypass fails, gated sensory feedback remains necessary to restore physiological error computation and adaptive control. Future neural bypasses may therefore need to restore multiple feedback channels simultaneously: combining ICMS, error-related potential detection and closed-loop external sensors to approximate the error-correction architecture of the native nervous system. Current bidirectional state-of-the-art systems represent important progress toward adaptive closed-loop control [30, 31], but still rely predominantly on visual feedback rather than restoring the fast proprioceptive feedback loops that normally operate at millisecond timescales. Dynamic motor control in BCIs and neural bypasses remains poorly understood, and represents an important area for future research.

Dissimilarity in natural and artificial information structure

Although neural bypasses aim to replicate physiological signal transmission pathways, they can only approximate the native flow of neural information. Despite single-neuron recording capability [97], current interfaces suffer from incomplete signal sampling, as they fail to capture the full diversity and specificity of synaptic connectivity including the complex micro- and macro-connectome features that underlie normal brain function [98]. Moreover, current materials and architectures cannot reproduce the biophysical dynamics that govern synaptic plasticity, constrained by inherent trade-offs between ionic mobility, retention stability, energy efficiency, and scalability [99, 100]. Reconstructed pathways thus often diverge from native information routes. Minassian et al. demonstrated in an epidural stimulation study that sustained lumbar stimulation intended to suppress spasticity unexpectedly activated intrinsic locomotor networks by transiently organising spinal interneurons into functional units, indicating that continuous afferent input can be reinterpreted by the spinal cord as a central motor command [19]. As such, to improve precision, decoding algorithms may need to compensate for functions normally performed by interposed structures like the cerebellum and spinal cord to accurately transform brain-intended motor output. This may require more advanced decoding strategies, or bidirectional bypasses with additional external sensors to provide sensory feedback.

From bench to bedside

Preclinical Neural Bypasses

The translational progression from neuroprostheses to the first neural bypass in humans is summarised in Fig. 4. Early bioelectronic experiments showed that cortical activity could directly control external actuators in rodents and monkeys to execute simple movements [101, 102]. While this work laid the groundwork for neural bypass development, it also exposed inherent limitations of prosthetics such as limited sensory feedback, catalysing a shift toward true neural bypasses by the late 2000s. Against this backdrop, Moritz et al. demonstrated the first cortical-muscular neural bridge in Macaca nemestrina to generate bidirectional wrist torques [103]. Building on this concept, successive brain-computer interfaces (BCIs) progressed from binary activation to graded control of paralysed limbs in rodents and primates [104, 105]. This could be achieved via multiple strategies: spike threshold decoding to discriminate motor intention [106], phase-locked stimulation [107], scaling stimulation amplitudes proportional to cortical ensemble activity [108], and reproducing natural spatiotemporal activation patterns of motoneuron pools during locomotion [109]. Phase-dependent stimulation, in particular, plays a distinct role by aligning neuronal activation with specific locomotor phases (e.g. swing and stance of the gait cycle) or with cortical beta-band desynchronisation signalling movement intent, thereby facilitating re-learning of physiological spinal circuit activation patterns. These approaches led to functionally meaningful outcomes across species, as exemplified by brain-computer-spinal interfaces (BCSIs) in rodents and primates where proportional lumbar epidural or subdural stimulation restored coordinated limb movements and naturalistic gait [109, 110]. Neural bypass efficacy in non-human primates (NHPs) was especially significant, as their motor cortex and movement repertoire closely resemble those of humans, allowing more accurate modelling of human motor function. Collectively, these preclinical findings established the feasibility of neural bypasses for restoring movement after SCI, but also underscored ongoing challenges like interface biocompatibility and signal decoding inconsistencies [111, 112].

Fig. 4.

Timeline of key milestones in the evolution of neural bypasses with representative studies, depicting progression of a neural recording capabilities b stimulation modalities, and c externally controlled neuroprostheses to closed-loop neural bypass systems. Created using biorender.com. MEA: Microelectrode Array; BSI: Brain-Spine Interface; SCI: Spinal Cord Injury; DBS: Deep Brain Stimulation; FES: Functional Electrical Stimulation; NMES: Neuromuscular Electrical Stimulation; SCS: Spinal Cord Stimulation

Clinical applications of bypass neural interfaces

Neural bypasses can be classified by the sites of reconnection relative to the lesion. In humans, cortical-muscular bypasses are the most established, particularly for stroke and SCI [51, 113], where cortical motor signals are recorded, decoded and relayed to specific muscle groups or lower motor neurons innervating the target muscle groups.

Building on preclinical proof-of-concept studies, clinical trials have validated the human cortex as a robust source of motor signals for bypasses. Whether employing surface or intramuscular electrodes to activate selected muscles, or extra-neural nerve cuffs to engage distinct fascicles, bypasses using cortical recording interfaces consistently demonstrate that high-dimensional cortical activity can be decoded in real-time sufficient accuracy to drive motor output [14, 16, 114]. Beyond peripheral stimulation, spine-targeted bypasses offer greater efficiency by engaging the dense neuronal networks upstream of branching for peripheral muscular innervation, though such greater complexity relative to peripheral structures poses more technical challenges. The wireless BCSI developed by Lorach et al. represents the most advanced implementation of this approach to date, employing dual ECoG implants, a multilinear algorithm decoding movement amplitude and direction, and a lumbosacral epidural stimulator to restore coordinated overground walking in a tetraplegic patient [30]. Recently, researchers from Fudan University further replicated Lorach et al.’s success in four patients with SCI, achieving partial lower limb recovery within a month using two wireless subcentimeter chips implanted in the brain and cord [115]. These fully wireless devices improved signal fidelity by eliminating cable-related impedance, minimising signal noise [116].

Building on preclinical demonstrations of graded control of muscle contractions, the most advanced bypasses enable increasingly complex, coordinated movements as opposed to simple binary modes of activation [117]. This can be largely attributed to novel non-linear decoding algorithms that have mitigated interference between stimulation-evoked and residual motor activity. These principles have previously been applied to synchronise intact and paralysed limbs [117], and to reproduce rhythmic, fluid movements such as thumb wiggling by emulating spinal central pattern generator (CPG) networks [118]. Nonetheless, implementing graded control remains challenging as it relies heavily on support vector regression models to capture non-linear dynamics, improving precision but at the cost of signal interpretability and computational latency.

Beyond cortical-muscular and cortical-spinal bypasses, bypass modalities continue to expand in range, from inter-cortical to inter-peripheral and autonomic interfaces. Table 1 summarises the major categories of neural bypasses and representative studies, highlighting how bypass architectures differ according to lesion location and target pathway, along with key results. Table 2 further compares the technical design of the systems in these studies, detailing the recording interfaces, decoding strategies and stimulation modalities that collectively determine bypass performance and translational feasibility. Studies included were identified through targeted PubMed searches using key neural bypass–related search terms, such as ‘neural interface’, ‘brain-computer interface’ or ‘BCI’ and ‘brain-machine interface’ or ‘BMI’. Studies were selected through discussions between the co-authors, based on both historical significance and translational relevance. As such, seminal studies establishing each bypass modality in humans were prioritised where applicable. Otherwise, more recent works were included to reflect current technological advances.

Table 1.

Overview of neural bypass approaches for paralysis: study paradigms, functional outcomes, and future directions

Bypass pathway	Study	Type of study	Patient Demographic	Experimental Paradigm	Duration of therapy	Key findings
Brain to muscle	Bouton et al. 2016 [14, 117–120]	Clinical trial (n = 1)	24 year old male with chronic C5/6 SCI (ASIA class A with partial preservation).	A 96-channel MEA was implanted in the left primary motor cortex. A high-resolution NMES sleeve stimulated forearm muscles.	15 months. Stimulation up to 3 times/ week.	Continuous cortical control of six distinct wrist/ hand movements with ~ 70% overall accuracy was achieved, with functional gains to the C7–T1 level. Multi-step tasks were completed successfully.
Brain to peripheral nerve to muscle	Herring et al. 2023 [16]	Clinical trial (n = 1)	27 year old male with chronic C3/4 AIS B tetraplegia.	A 64-channel MEA was implanted in the left sensorimotor and associative cortices. 16-channel extra-neural cuffs were wrapped around major right upper limb nerves and triceps. S1 arrays delivered intracortical microstimulation (ICMS).	1 year.	Decoded cortical signals generated functional reaching and grasping via multi-nerve stimulation. Motor signals were mapped to intended movements with 79% accuracy. ICMS of S1 successfully evoked localised fingertip sensations, while nerve stimulation produced spatially organised, muscle-specific activation.
Brain to cord to nerve to muscle	Lorach et al. 2023 [30]	Clinical trial (n = 1)	38 year old male with chronic C5-C6 ASIA C tetraplegia.	64-channel WIMAGINE® ECoG implants inserted over the sensorimotor cortex bilaterally. Real time decoding modulated a 16-contact epidural paddle over the lumbosacral dorsal root entry zones targeting hip, knee, and ankle motor pools.	15 weeks, 40 sessions each 1–3 h. Currently on 3-year follow-up for home BSI use.	Motor intentions were decoded via a Markov-switching multilinear algorithm with > 90% accuracy to proportionally modulate epidural stimulation. The patient regained natural, continuous cortical control over standing, stepping, and overground walking, including adaptation to complex terrains. Performance remained stable for > 1 year. Prolonged training, produced lasting neurological recovery, as voluntary walking with crutches for > 10 m was possible even in StimOFF conditions.
Cord to nerve/ muscle	Burke et al. 2021 [121]	Proof of concept clinical trial (n = 10)	10 patients with chronic pain undergoing a spinal cord stimulator (SCS) trial.	Temporary 8/16-contact epidural electrodes over the cervical or thoracic cord enabled direct spinal electrogram (SEG) recordings in awake participants. EMG was recorded while participants performed elbow and knee flexion–extension, overground walking, stepping in place and sensation of light touch.	7–10 days.	SEGs were safely recorded in awake participants, showing reliable movement-related spectral modulation. Most channels were usable after artifact screening, confirming feasibility of direct human spinal recordings as a source for future cord-effector bypasses.
Nerve to nerve	Sliow et al. 2019 [122]	Preclinical rodent trial (n = 30)	30 female Wistar rats weighing 287 ± 16 g, with the left sciatic nerve manually transected.	A bioadhesive wireless gold-chitosan graft-antenna was used to repair the transected nerve without sutures at the proximal anastomosis. The graft-antenna converted transcranial magnetic stimulation (TMS) pulses into induced currents to evoke muscle responses and axonal regeneration.	8–12 weeks, 1 h per week.	Compared to adhesive-only controls, treated rats exhibited marked electrophysiological and functional recovery – higher muscle and nerve action potentials, faster conduction velocity, muscle reinnervation and gait recovery. Histological evidence confirmed axonal regrowth and remyelination without inflammation or neuroma formation. Graft-antenna systems may be a viable minimally invasive nerve-to-nerve bypass.
Cord to cord	Woodington et al. 2024 [123]	Preclinical rodent trial (n = 16)	Sixteen adult female Sprague-Dawley rats weighing 200–250 g.	A flexible 32-channel circumferential epidural array was implanted around the spinal cord at the T9-L1 level. The proximal array recorded compound action potentials at T9–T10, which then triggered ventrolateral stimulation at T12–L1 across a complete T11 transection to re-engage hindlimb muscles.	6 weeks.	The array restored coordinated hindlimb flexor–extensor activity, maintaining stable evoked compound action potential transmission and stimulation effects for 6 weeks without evidence of tissue necrosis or inflammatory encapsulation. Signal integrity remained stable. Submillisecond conduction latency enabled precise activation across the lesion. Human cadaver studies further confirmed the array’s conformability and surgical feasibility for translation to cord-to-cord bypass applications.
Brain to brain, between individuals	Rao et al. 2014 [124]	Clinical trial (n = 6)	3 pairs of senders and receivers. Age 21–38, 100% Male.	EEG-based motor imagery from senders was decoded in real time and transmitted over the internet to receivers 1 mile away. Transcranial magnetic stimulation (TMS) in receivers caused a quick jerk in the receiver’s hand, triggering a “click” event on the touchpad.	Pair 1: 10 trials. Pairs 2 & 3: 16 trials. Total 60 min.	83.3%, 25.0% and 37.5% correct successful cursor control by receivers and an average latency of 650 ms. ~4–13 bits transferred per experimental block.

Table 2.

Overview of recording, decoding and stimulation details for each study in Table 1

Study	Recording		Method of decoding	Stimulation
	Site	Mode		Site	Mode	Parameters
Bouton et al. 2016 [14, 117–120]	Left primary motor cortex (hand area, fMRI-guided)	96-channel intracortical Utah Microelectrode Array (Blackrock Microsystems, USA), sampled at 30,000 samples per second.	Multiunit spiking activity was transformed into mean wavelet power features (updated every 100 ms), which were classified by SVMs trained on six hand/wrist movements. The highest-likelihood output triggered the corresponding stimulation pattern in real time.	Right forearm muscles: primarily wrist flexors, wrist extensors, finger flexors, finger extensors, and thumb-related forearm muscles.	High-density neuromuscular electrical stimulation via custom-built 130-electrode flexible sleeve.	Current-controlled rectangular pulses were delivered. 50 Hz frequency; 500 µs pulse width; 0–20 mA amplitude. Updated every 100 ms.
Herring et al. 2023 [16]	Left M1 (hand/arm areas), left S1 (fingertip maps; 2 arrays), left anterior intraparietal area (AIP), left inferior frontal gyrus (IFG).	Six 64-channel intracortical microelectrode arrays, routed to percutaneous NeuroPort pedestals/headstages (Blackrock Neurotech).	A linear discriminant analysis (LDA) classifier used multiunit spiking activity from IFG, motor and premotor areas to decode up to eight different hand or wrist movements with ~ 79% accuracy. The decoded outputs modulated multi-contact peripheral nerve–cuff stimulation patterns to produce graded hand and wrist movements.	Radial, median, ulnar, musculocutaneous, axillary, suprascapular, lateral pectoral, medial pectoral, and long thoracic nerves. Also triceps.	Peripheral nerve FES delivered via nine 16-channel composite flat interface nerve electrodes (C-FINEs). Simultaneous multi-contact stimulation across nerves produced coordinated movements. S1 sensations delivered via ICMS (Cerestim R96, Blackrock).	Charge-balanced biphasic pulses were delivered. S1 ICMS (for touch) delivered as cathode-first biphasic trains at 100 Hz frequency, 200 µs pulse width, 80 µA amplitude, (Cerestim R96)
Lorach et al. 2023 [30]	Bilateral sensorimotor cortex (ECoG grids over M1/S1).	64-channel WIMAGINE® ECoG arrays (custom, Wyss Center/ ONWARD Medical). 32 contacts per implant, sampled at 586 Hz.	High-gamma and beta-band ECoG features were decoded using a recursive exponentially weighted Markov-switching multilinear model (REW-MSLM), where a hidden-state gate selected context-specific linear decoders (e.g., walking vs. joint control). Decoder parameters were updated online every 15s. In parallel, a recurrent neural network was trained offline to predict continuous lower-limb kinematics from ECoG features, validating the REW-MSLM performance.	Lumbosacral dorsal root entry zones, T12-S1 (motor pools for hip, knee, ankle)	Epidural electrical stimulation delivered by an implantable pulse generator (Medtronic Activa RC) connected to a 16-contact percutaneous epidural paddle lead (Medtronic Specify 5-6-5).	Charge-balanced rectangular pulses (0.1–500 Hz frequency depending on the task, 60–450 µs pulse width, up to 25.5 mA amplitude) were delivered. Operational ranges remained stable over ~ 1 year of use.
Burke et al. 2021 [121]	Dorsal epidural surface of the cervical (C4-C6) and thoracic (T8-T10) cord.	8- or 16-contact epidural spinal cord stimulation (SCS) leads were used (cylindrical or paddle). 56 cervical electrodes, 44 thoracic electrodes across the 10 participants.	Spinal local field potentials were analysed offline using time–frequency decomposition with Morlet wavelets to quantify movement-related spectral power changes between movement and rest periods.	N/A
Sliow et al. 2019 [122]	N/A			Graft-antenna was laser bonded circumferentially around the transected sciatic nerve ends, bridging the 10-mm nerve gap.	Transcranial magnetic stimulator (TMS) was coupled to a loop antenna encircling the nerve, inducing microampere currents that generated tangential electric fields evoking action potentials.	Single TMS pulses (350 µs pulse width, 0.72 T magnetic field magnitude) were applied for 1 h weekly over 8–12 weeks, inducing currents of 13 µA in the gold loop and stable CMAPs of 1.43 ± 0.11 mV.
Woodington et al. 2024 [123]	Dorsal epidural surface of thoracic spinal cord at T9–T10, rostral to the transection.	32-channel circumferential epidural parylene-C/PEDOT: PSS array (Neurosoft i360, Russia), wrapped around the thoracolumbar cord. Sampled at 30 kHz for spinal eCAPs.	N/A	Implanted caudal to the transection at T12–L1 for stimulation below the lesion site. In the epidural plane, between exiting nerve roots.	Epidural spinal cord stimulation via custom 32-channel circumferential, flexible, thin-film epidural array (i360, Ripple Neuro). PEDOT: PSS-coated gold electrodes enabled direct current-controlled tonic stimulation.	Charge-balanced biphasic pulses (50–100 Hz frequency, 300 µs pulse width, 10–200 µA amplitude) in 10–20-pulse trains.
Rao et al. 2014 [124]	Sender’s scalp wearing EEG cap.	64-channel Ag/ AgCl electrode cap (actiCAP, Brain Products GmBH, Germany), sampled at 512 Hz	Motor imagery was decoded by quantifying mu-band (8–12 Hz) power suppression from the electrode with strongest correlation with right-hand motor imagery during training. Reduced mu power was mapped to upward cursor movement for command generation	Part of M1 controlling the extensor carpi radialis muscle	Computer-controlled TMS. Pulses delivered using a 90 mm MagStim circle coil connected to a Super Rapid2 magnetic stimulator (MagStim, UK	TMS delivered in single pulses, win 20 s intervals. Maximum intensity of electric field: 530 V/m. Maximum intensity of the induced magnetic field is 2.0T.

Challenges in clinical implementation of neural bypasses

Although neural bypasses have demonstrated compelling functional recovery in early clinical settings, their translation towards widespread, routine use remains limited, constrained by challenges spanning neural recording, signal decoding, stimulation and user adoption (Fig. 5).

Fig. 5.

Summary of challenges faced in clinical translation of neural bypasses. Created using biorender.com

Challenges in recording

Neural bypasses inherit several recording challenges common to neural interfaces, with clinical translation still limited by signal fidelity, biocompatibility, and communication architecture. As such, maintaining clean, stable signals while minimising information loss remains a persistent obstacle. Non-invasive modalities like EEG and MEG often miss physiologically relevant activity from deeper sources due to coarse spatial resolution, with further loss in fidelity intrdocued by distortion, artifact rejection, and narrow band-pass filtering [125]. Invasive arrays, in turn, struggle with signal localisation in densely packed structures like the cord, where volume conduction and tight contact spacing impede the isolation of activity from specific neurons or tracts [47]. High-density or poorly insulated electrode arrays can further exacerbate cross-talk in these rergions, compounding spike sorting errors during decoding [126]. In chronic settings, electrode degradation and foreign-body responses further undermine both recording selectivity and signal fidelity, as extensively discussed in reviews like Carnicer-Lombarte et al. and Zeng and Huang [112, 127]. Beyond issues in signal capture, neural bypasses also face data transmission constraints: tethered systems restrict mobility and add mechanical impedance [128], whereas wireless designs remain limited by power consumption and transmission range [129]. These power restrictions become particularly evident when processing dense neural recordings where ultra-efficient signal amplification and decoding is demanded, often exceeding the capacity of implantable batteries operating within safety and thermal limits. However, recent advances in ultra-low-power circuitry and battery-less platforms may mitigate these constraints [130].

Issues in signal transformation in neural bypasses

Decoding remains a major computational bottleneck in the implementation of neural bypasses. Among the wide range of challenges spanning different architectures and modalities, as discussed in-depth by Tam et al. and Livezey and Glaser [131, 132], we highlight two that are particularly limiting for neural bypasses. Firstly, real-time transformation of neural signals into kinematic output is constrained by a fundamental trade-off between decoding complexity and processing speed. More sophisticated models are needed to capture the non-linear dynamics underlying multi-joint movements, but these gains in accuracy often introduce computational latency that disrupts the temporal precision needed for real-time decoding to produce smooth movements [133]. Adopting phase-aware or latency-tolerant decoding strategies could meaningfully mitigate temporal misalignment and better preserve the synchrony required for coordinated movement. Signal variability imposes another major limitation. Even in healthy individuals, mappings of neural activity to specific actions drift continuously due to biological stochasticity, sensory context and changes in internal state (e.g., arousal, fatigue) [134–136]. These fluctuations are exacerbated in patients with SCI due to injury-induced remodelling of cortical and spinal circuits, which can continue to evolve with repeated bypass usage through Hebbian and homeostatic plasticity mechanisms [137]. As a result, static decoders deteriorate rapidly or require frequent recalibration, pointing toward adaptive, closed-loop approaches as a necessary next step for ensuring long-term decoder stability.

Stimulation parameters, selectivity and safety

While decoding determines how neural intent is interpreted, stimulation governs how that intent is executed. Stimulation selectivity depends largely on the target site and degree of invasiveness. Transcutaneous FES of muscles is often imprecise, limited by the indiscriminate spread of surface current that may co-activate antagonistic muscle groups in close proximity, causing spasticity. Even slight shifts in surface electrodes during contraction can further undermine stimulation accuracy. Implantable approaches like intramuscular and peripheral nerve stimulation offer greater precision by directly engaging effectors to induce specific muscle activation [16]. However, proximal stimulation along mixed nerve trunks reduces selectivity, as densely intertwined fascicles hinder isolation of specific muscle targets. Muscle selectivity becomes even more nebulous in spinal cord or nerve-root stimulation. The dense cytoarchitecture and overlapping fibre tracts preclude precise targeting of axons responsible for evoking a specific movement. While gross movements involving multiple muscle groups such as locomotion can still be elicited, tasks demanding fine precision like movement of individual phalanges may remain difficult to achieve. As such, stimulation of posterior rootlets is sometimes preferred over dorsal column fibres, as the latter produces non-selective activation and paraesthesia through longitudinal current spread, recruiting multiple dermatomes and myotomes [138]. When residual supraspinal pathways are preserved, targeted stimulation of dorsal rootlets corresponding to specific motor hotspots is even more essential to recruit sub-functional locomotor circuits and selectively excite the relevant spinal segments for voluntary motor control [28, 139]. Equally important is optimising stimulation parameters like frequency, amplitude and pulse width. Minassian et al. demonstrated that low frequency lumbosacral epidural stimulation at 5–15 hz induces tonic extension, whereas higher frequencies at 25–50 Hz evoke rhythmic, stepping-like movements instead [19]. In fact, excessively high amplitudes or frequencies can paradoxically dampen muscle tone, trigger spasms and induce autonomic dysreflexia [140, 141]. Currently, no optimal set of parameters has been agreed upon as parameters seem to be titrated on a case-by-case basis.

Finally, while biocompatibility concerns parallel those of recording interfaces, the long-term safety of chronic stimulation remains insufficiently characterised. Sustained or supraphysiological activation lasting up to 16 h can induce over-excitation of nerves or muscles, leading to neuronal hyperactivity, excitotoxcitity and eventual axonal and synaptic degeneration [142, 143]. In parallel, muscle fatigue limits the sustainability of FES as contractile efficiency declines with repeated activation. Closed-loop systems that adjust stimulation parameters based on real-time changes in muscle properties may mitigate fatigue, but their efficacy is constrained by the lack of reliable implantable muscle torque sensors and progressive changes in neuromuscular tissue such as atrophy, that alter muscle response over time [144]..

Challenges in user adoption and acceptance

The adoption of neural bypasses remains limited by practical, ethical and generalisability barriers. As with most surgical procedures, implantation carries risks of infection, immune rejection and neurovascular injury [145]. Postoperatively, patients often have to endure lengthy training regimens and frequent recalibration for bypasses without adaptive algorithms, contributing to user fatigue and frustration. Poor device ergonomics further impair long-term, independent usage in the community setting. In addition, ethical and financial concerns impede uptake – patients may fear loss of privacy or autonomy from algorithmic control especially in bidirectional systems [146], while high costs of bypass implantation and maintenance restrict accessibility [147]. However, a more fundamental challenge lies in limited generalisability: current systems are highly individualised, with responses to stimulation varying across patients due to differences in neural architecture, muscle properties, paralysis aetiology, lifestyle factors and comorbidities [128]. Consequently, device parameters optimised for one user seldom translate effectively to another, underscoring the need for scalable, adaptive control frameworks that can accommodate inter-patient variability.

Vision for the future

Addressing the aforementioned translational constraints will be essential for realising the next generation of neural bypass systems, motivating emerging efforts toward adaptive and biologically integrated architectures.

Sensory and bidirectional bypass neural interfaces

As neural bypass technologies advance, future progress hinges on achieving adaptive control in bidirectional systems through integration of sensory feedback. Closed-loop systems incorporating feedback can continuously refine stimulation parameters in real-time, enabling more natural movement via improved predictive and reactive control [148, 149], alongside reduced dependence on repeated recalibration. Beyond ICMS of S1, several peripheral sensory feedback strategies—including regenerative peripheral nerve interfaces, peripheral nerve stimulation (e.g. tibial nerve) and transcutaneous electrical stimulation—have demonstrated restoration of tactile and proprioceptive percepts [150–153]. These have translated into improved object discrimination, prosthetic agility and reduced cognitive burden during movement. The functional value of bidirectionality is already more established in conventional neuroprosthetic systems, where sensory feedback has consistently improved grasp and locomotor control [11, 12]. Recently, these principles have begun to translate into neural bypass systems. Developed by Chandrasekaran et al., the first “double neural bypass” integrated M1 decoding of movement intent, S1 ICMS driven by force sensors for tactile feedback, and transcutaneous cervical dorsal root SCS for motor facilitation [31]. Using a nested LSTM and deep reinforcement learning architecture for grasp regulation, the participant regained the ability to self-feed, grasp delicate objects, and hold and feel another person’s hand, with these improvements persisting over the long-term.

The spinal cord as a source of information

Future neural bypass architectures may shift toward spinal cord–machine interfaces, leveraging the spinal cord not only as a stimulation target but also as a bidirectional source of motor and sensory information. The viability of the spinal cord as a stimulation interface is already well established. Multiple clinical trials have demonstrated the use of epidural and transcutaneous SCS for successful restoration of voluntary movement and functional recovery [20, 30, 31, 154–157]. More recently, SCS has also demonstrated the ability to restore somatosensation. Calvert et al. demonstrated the first human spinal interface capable of simultaneously restoring lower limb movement and somatosensory feedback (lower limb sensory percepts and proprioception) using perilesional eSCS [158]. Similarly, Chandrasekaran et al. showed that eSCS and tSCS could evoke stable somatotopically organised sensory percepts by leveraging preserved dorsal column organisation [159, 160], while the Up-LIFT trial demonstrated that tSCS improved light touch and pinprick sensation in over half of participants with previously absent upper limb sensation due to chronic cervical SCI [161].

An emerging question is whether the spinal cord can also serve as a recording source for future bypasses. Preclinical studies have demonstrated that the cord may provide motor information independent of cortical recordings, correlating strongly with limb kinematics for accurate movement reconstruction [162–165]. Although these approaches largely relied on intraparenchymal electrodes and have yet to be translated clinically, Burke et al. was among the first to demonstrate the feasibility of recording viable spinal electrograms from the epidural space in humans during motor tasks, providing a potential framework for mapping spinal motor neuron activation [121]. More recent work by Woodington et al. has further demonstrated the feasibility of a unidirectional spinal cord bypass framework in rats [123]. Apart from motor signals, the spinal cord is also a rich source of sensory information. Lo et al. provided one of the most comprehensive demonstrations of a bidirectional spinal cord–machine interface, showing that MEAs implanted within the fasciculus gracilis could record compound action potentials with high spatial precision corresponding to dermatomal somatotopy. Distinct tactile stimulation sites produced spatially separable signals, while proprioceptive changes during hindlimb abduction modulated dorsal column discharge frequency [166]. Beyond the dorsal columns, the dorsal root ganglion represents another promising sensory recording target, with studies demonstrating recording of both population-level and single-unit sensory activity, including decoding of joint position and movement timing [167–169]. Importantly, emerging studies have also demonstrated the feasibility of non-invasive spinal cord recordings in humans, suggesting that electrospinographic approaches may eventually provide a scalable alternative to highly invasive spinal interfaces for future spinal cord–machine systems [170, 171]. Finally, spinal interfaces may also extend beyond sensorimotor control toward autonomic functions, including bladder control, where fUSI has been used to visualise spinal activation during bladder activity alongside targeted stimulation of descending spinal pathways [9, 10, 49].

Nevertheless, spinal cord recordings face greater challenges than cortical interfaces, particularly regarding signal quality and long-term stability. Continuous cord motion within the vertebral column generates substantial recording artefacts and mechanical strain on implanted electrodes [172], while foreign body responses are amplified in the cord [173], further degrading signal fidelity over time. As such, chronic spinal cord recordings remain difficult to maintain and likely face more stringent durability requirements than intracortical interfaces [172]. Moreover, while spinal signals may encode movement execution, higher-order intention and planning remain more prominently represented within cortical networks [45].

Effect of bypass neural interfaces on biological recovery

Perhaps the most transformative aspect of neural bypasses is their potential to not just temporarily restore movement, but also to promote long-term biological recovery by leveraging the brain and cord’s intrinsic neuroplasticity. Through sustained usage coupled with sensory feedback, bypasses can drive synaptic remodeling and re-engagement of dormant corticospinal and propriospinal circuits to fine-tune motor coordination.[55]. This could ultimately enable restoration of volitional control independent of the device – a phenomenon demonstrated in BCIs for stroke rehabilitation and more notably, in the cortical-spinal bypass developed by Lorach et al. where locomotor function persisted even after bypass deactivation [30, 174]. Ultimately, neural bypasses may serve not as permanent prosthetic aids but as transient catalysts for self-sustaining motor recovery. To elucidate these restorative mechanisms, longitudinal studies linking functional gains with structural and connectivity changes in the brain and spinal cord will be essential.

Conclusion

In conclusion, neural bypasses represent a major leap in neurorestoration, enabling paralysed individuals to regain volitional control over once-immobile limbs. Within just over a decade, the field has advanced from preclinical models to human trials, with recent innovations in adaptive decoding algorithms that enhance robustness against signal variability across patients and time. The next frontier lies in developing more fully bidirectional systems that integrate sensory gating and error signals to restore naturalistic movement. Moving forward, larger trials with more participants will be essential to establish generalisability across decoding algorithms, training regimens, and stimulation parameters, and to address persistent challenges in signal fidelity, stability, and chronic usability before widespread clinical translation.

Author contributions

M.J.R.L. and A.K.S.T. drafted the manuscript and contributed to critical revision of the article. Y.T.L. and L.J. contributed to conceptual development of the review and critically revised the manuscript. C.L. contributed to critical revision of the manuscript. All authors reviewed and approved the final manuscript.

Funding

No external funding was used for this study.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors contributed to this article and approved the submitted version.

Competing interests

The authors declare no competing interests.