Review of Recent Advances in Implantable Brain-Computer Interfaces for the Restoration of Motor Function in Patients With Paralysis

Introduction

Motor impairment – particularly paralysis – represents one of the most severe and burdensome consequences of neurological disease or injury, including stroke, spinal cord injury, and traumatic brain injury [1]. At the global level, recent burden-of-disease analyses report about 12.2 million new stroke cases and 101 million stroke survivors, as well as roughly 0.9 million incident and 20.6 million prevalent spinal cord injury cases in 2019, indicating that tens of millions of individuals worldwide live with chronic motor impairment secondary to these conditions [2,3]. Damage to the brain, spinal cord, or the connecting pathways disrupts the transmission of motor commands to peripheral effectors, resulting in loss of volitional movement and profound reductions in quality of life and social participation. Conventional rehabilitation approaches, such as physical therapy, task-oriented training, and robot-assisted therapy, can improve strength, coordination, and motor control to some extent; however, their efficacy is often limited in individuals with severe or complete paralysis, where true motor restoration and re-establishment of functional neural pathways remain challenging [4,5]. Against this backdrop, implantable brain–computer interfaces (BCIs) have emerged as a compelling neurotechnology for improving motor function in severe motor disability. Unlike noninvasive BCI modalities (eg, electroencephalography [EEG] and near-infrared spectroscopy), implantable BCIs place high-spatial-resolution electrodes on the cortical surface, within the cortex, or in deep structures, to capture high-bandwidth, high–signal-to-noise neural activity, enabling extraction of more precise and stable motor intent signals [6]. In this review, motor intent refers to the cortical representation of an intended movement, generated even in the absence of overt motor execution, which can be decoded from neuronal activity to control external effectors. This direct brain–machine linkage creates the possibility of establishing a “neural substitute pathway” or “digital bridge” for motor control. Recent advances across electrode materials and micro/nanofabrication, signal processing and machine learning, and low-power wireless communication and hermetic packaging have collectively accelerated progress in both the basic research and early clinical explorations of implantable BCIs [7]. Recent developments in electrode materials and micro/nanofabrication have expanded the design space for implantable neural interfaces. Figure 1 highlights representative strategies and considerations relevant to these advances. Over the past decade, implantable BCIs have progressed through several major clinical milestones. The earliest demonstrations in individuals with tetraplegia showed reliable cortical control of robotic arms, including reach-and-grasp movements [8,9]. Endovascular implantation using the Stentrode device enabled minimally invasive cortical signal acquisition without craniotomy [10], with subsequent clinical studies demonstrating sustained daily communication in patients with amyotrophic lateral sclerosis [11]. A landmark proof-of-concept demonstration using brain-controlled functional electrical stimulation restored reaching and grasping in a person with chronic tetraplegia [12]. Most recently, closed-loop brain–spine interfaces have enabled overground walking in individuals with spinal cord injury, marking a new phase in BCI-driven neurorehabilitation [13]. The central premise in motor restoration is to decode cortical motor intent in real time and translate it into control commands for external effectors, such as robotic manipulators, exoskeletons, functional electrical stimulation systems, or spinal cord stimulation electrodes, to drive limb movement or re-engage peripheral and central pathways [14]. This pipeline, illustrated in Figure 2, encompasses accurate intent recognition (acquisition to processing to decoding), coupled with feedback, closed-loop control, and targeted induction of neuroplasticity. Neuroplasticity, the brain’s capacity to reorganize synaptic connections and functional networks, is central to BCI-enabled motor restoration. Unlike passive rehabilitation, implantable BCIs can directly couple cortical activity to peripheral or prosthetic effectors, actively engaging plasticity mechanisms to promote functional recovery. In principle, prolonged neurofeedback and combined training may facilitate network reorganization, axonal sprouting, or functional re-mapping, thereby promoting genuine motor recovery beyond device operation. Evidence from animal models has substantiated the feasibility of implantable BCIs in eliciting and guiding motor recovery [15]. For example, in spinal cord injury models, recording cortical intent and driving spinal or muscular stimulation can partially restore lower-limb function [16]. Notably, a recent study implemented a BSI to re-establish communication between cortex and spinal cord in a person with chronic tetraplegia, achieving naturalistic overground walking and showing signs of carry-over improvements even when the interface was not active [13]. Such “digital-bridge” strategies, leveraging plasticity mechanisms under closed-loop feedback, are widely regarded as among the most promising directions for motor restoration with implantable BCIs. Early clinical translation has likewise yielded encouraging signals. Pilot trials using cortical or endovascular brain–computer interface approaches have demonstrated controllable movements, prosthetic manipulation, and enhanced electromyography-driven motor output in individuals with paralysis, including the landmark demonstration of brain-controlled muscle stimulation for reaching and grasping in a person with tetraplegia [12]. While large-scale, multicenter studies with long-term follow-up are still lacking, these early experiences provide important feasibility data and practical insights for applying implantable BCIs to motor recovery. Accordingly, the review synthesizes recent advances in implantable BCIs for motor restoration in paralysis, spanning mechanistic underpinnings, experimental models, key technological innovations, and pathways and challenges for clinical translation. Our goal is to offer a clear perspective for researchers, clinicians, engineers, and policymakers, thereby supporting the continued progression of implantable BCI systems from laboratory demonstrations to bedside application. Several reviews [17] have previously summarized BCI technologies for motor rehabilitation, focusing on noninvasive EEG-based systems, decoding algorithms in isolation, or specific etiologies, such as stroke or spinal cord injury. However, most of these articles either treat implantable BCIs only as a subset of broader BCI taxonomies or emphasize technical aspects without systematically linking mechanistic insights to clinical translation. In contrast, the present review concentrates specifically on implantable BCIs for motor recovery after paralysis, integrating multiscale neural mechanisms, electrode and materials innovations, advanced decoding strategies, and current clinical trial experience. This article aims to review recent advances in implantable BCIs for the restoration of motor function in patients with paralysis.

Fundamental Mechanisms: Principles of Neural Signal Acquisition for BCIs

NEURAL SIGNAL SOURCES AND INFORMATION SCALES:

Implantable BCIs use voltage fluctuations from neuronal membranes, including spikes and local field potentials (LFPs) [13]. Spikes primarily serve to provide high-resolution, single-neuron–level information, reflecting whether individual neurons participate in motor encoding at precise time points. This enables extremely fine-grained motor intent decoding, including continuous control of high–degree-of-freedom movements such as hand and finger actions. By contrast, LFPs primarily serve to characterize the synchronized activity of local neuronal populations, capturing network states, rhythmic modulations, and task-related dynamics. Owing to their robustness and long-term stability, LFPs are well suited for motor state monitoring, intent inference, and serving as a reliable feature source in multimodal BCI systems [18]. These scales are complementary: single-unit signals offer fine granularity but are more vulnerable to instability, whereas LFPs are comparatively robust yet less specific. Combining both can materially enhance decoding performance. To access these neural phenomena, multiple electrode technologies have been developed [19]. Each class entails distinct trade-offs in spatial resolution, signal-to-noise ratio, invasiveness, and long-term stability, and these properties ultimately bound the attainable performance for motor restoration.

ELECTRODE-TISSUE INTERFACE AND ACQUISITION MODALITIES:

Transduction of neural activity into measurable electrical signals requires reliable coupling at the electrode–tissue interface. Interface quality and stability directly influence signal amplitude, signal-to-noise characteristics, and chronic reliability. Penetrating microelectrode arrays can approach or traverse the perisomatic space of individual neurons, enabling high-amplitude, low-noise single-unit recordings. They primarily serve to record the activity of individual neurons or small neuronal clusters with high spatial and temporal resolution, enabling fine-grained decoding of motor intent. Penetrating microelectrode arrays are especially useful for high-precision control tasks, such as multi–degree-of-freedom hand and finger movements, and therefore often constitute the core recording platform in high-performance, invasive BCI systems. However, they can incur local tissue injury and inflammatory responses that precipitate impedance increases and signal degradation over time [20]. Electrocorticogram (ECoG) electrodes mainly function to capture stable population-level cortical activity from the brain surface (eg, β, γ, and high-γ band power), offering a favorable balance among spatial resolution, invasiveness, and long-term stability. They are well suited for multi–degree-of-freedom motor control, cortical mapping (eg, seizure focus localization), and the clinical translation of long-term implantable BCI systems. By comparison, surface ECoG electrodes do not breach the cortex, offering reduced surgical burden and generally favorable safety profiles, but they measure population activity with lower spatial specificity. Substantial materials and structural innovations have been sought to mitigate chronic failure modes: flexible polymer substrates, nanowire electrodes, and ultrathin metallic conductors reduce mechanical mismatch and micromotion-induced trauma, thereby limiting drift in signal quality [21]. Conformal surface modifications, such as carbon nanotubes, graphene, platinum black, and conductive polymers, can decrease impedance, improve signal-to-noise ratio, and enhance biocompatibility [22]. Of particular note, endovascular electrodes (eg, the Stentrode platform) position recording contacts along the cerebrovascular wall, achieving cortical signal capture without craniotomy; resultant signal quality approximates that of conventional surface electrodes and shows promising translational potential [23]. Each modality offers characteristic benefits: single-unit activity affords maximal precision but depends critically on interface stability; multi-unit activity aggregates several nearby neurons to balance fidelity and robustness; LFPs are durable yet less granular; ECoG and depth electrodes survey wider territories; and endovascular approaches seek an intermediate point between invasiveness and signal quality [24]. Contemporary systems increasingly adopt multimodal, multichannel strategies, pairing high-density arrays with parallel acquisition of diverse signal types, to jointly optimize fine control and resilience in motor-restoration applications.

LONG-TERM STABILITY AND PREPROCESSING CONSIDERATIONS:

Neural signals are microvolt-scale and noisy, requiring low-noise amplification, filtering, and ADC for decoding. A representative front end consists of high-pass filtering (approximately 0.1–1 Hz) to suppress baseline drift, band-pass filtering to attenuate out-of-band noise, and high-rate digitization (kHz to tens of kHz) to preserve temporal fidelity. Across this chain, wide dynamic range and low input-referred noise must be balanced against stringent power budgets to mitigate heat generation and avoid thermal injury – an overarching constraint for chronically implanted systems. Despite robust short-term performance, chronic implantation introduces several degradation pathways. Foreign body response, including inflammatory cascades, astroglial proliferation, and encapsulation, elevates impedance and diminishes signal amplitude. Peri-electrode neuronal loss further weakens recording yield; brain–electrode micromotion introduces artifacts and slow drift; and a subset of channels may fail progressively over time. Multiple engineering strategies have been advanced to address these limitations [25,26]. (1) In regards to materials and mechanics, flexible substrates and stretch-tolerant architectures reduce mechanical mismatch and micromotion-induced trauma; surface modifications (eg, nanotexturing, conductive polymer or carbon-based coatings) lower impedance and improve biocompatibility. (2) In array design, optimized geometries and spatial layouts enhance capture probability while distributing mechanical load; redundancy coupled with online impedance monitoring enables self-calibration and fault-tolerant channel switching. (3) In electronics and algorithms, ultra-low-power integrated circuits and adaptive filtering/suppression schemes maintain signal quality within thermal safety limits and counteract nonstationarities. As illustrated in Figure 1, transparent electrodes enable co-localized electro-optical acquisition (Figure 1B) but introduce practical complications such as optical shadowing and photoelectric artifacts (Figure 1C). These realities heighten the importance of the present section’s measures: low-noise amplification, appropriate band-pass filtering, synchronized sampling, and adaptive denoising. Leveraging intrinsically transparent materials with mesh/porous architectures (Figure 1D) can simultaneously mitigate foreign body response and impedance drift, thereby laying a device-level foundation for stable long-term recordings and reliable closed-loop decoding. Together, these measures ensure that the neural signals acquired through implanted electrodes remain sufficiently stable and interpretable for downstream computational analysis. Building on this foundation, the following section focuses on how these recorded signals are transformed into task-relevant representations and decoded into meaningful motor outputs.

Neural Decoding and Pattern Recognition

FEATURE CONSTRUCTION AND TIME-FREQUENCY REPRESENTATIONS:

Prior to decoding, digitized neural time series must be transformed into task-relevant feature representations. As outlined in Figure 2, this process follows a standardized pipeline consisting of preprocessing, spatial selection (eg, common spatial patterns), feature extraction, and decoding/classification, with nonstationarity handling at the train–test interface; multimodal fusion is typically implemented between the feature extraction and classifier stages to yield the final intent output. For spike-based signals, practical features generally include windowed spike counts or firing-rate estimates, often smoothed to improve decoder stability [27]. These features directly enter the extraction module, as shown in Figure 2, and sort-less multi-unit approaches are frequently used to reduce the impact of electrode drift or unit turnover during long-term use [28]. For LFP and ECoG signals, decoding typically relies on band-limited spectral power or compact time-frequency features. β-band desynchronization (approximately 13–30 Hz) and γ/high-γ activity (>60–200 Hz) remain the most reliable markers for motor-related tasks, particularly for upper-limb and grasp decoding in ECoG/μECoG recordings [29,30]. These features likewise enter the feature extraction module and are subsequently routed to the classifier. To obtain stable representations, the pre-processing stage (Figure 2) typically applies band-pass/notch filtering, common average reference, and artifact suppression. During spatial selection (Figure 2), common spatial patterns or related supervised spatial projections, together with channel selection, enhance task-relevant signal-to-noise ratio; cross-channel regularization further attenuates common-mode noise and volume-conduction effects, thereby improving generalization [31,32]. For continuous decoding, short-time Fourier transform, Morlet wavelets, or multi-resolution filter banks can generate time-frequency maps, often combined with log-power and baseline normalization (event-related desynchronization/event-related synchronization) to increase discriminability. More recently, contrastive/self-supervised representations – for example, enforcing temporal-contrastive consistency across channels or views at the same time point – have been explored to construct more stable latent features across heterogeneous electrode layouts and subjects, facilitating cross-task and cross-cohort transfer [33,34]. Within the pipeline shown in Figure 2, these methods primarily target the feature–classifier interface and the train–test handoff, alleviating distributional shifts due to nonstationarity. In summary, the spike- and time-frequency features, pre-processing and spatial-selection procedures, and nonstationarity-aware adaptive/self-supervised strategies described above map one-to-one onto the modular structure of Figure 2, which provides the ordered implementation scaffold, while this section specifies concrete feature choices and robustness measures for each module.

TAXONOMY OF DECODING MODELS:

Recent work indicates that transformer-based architectures outperform conventional models in intracranial tasks, offering a unified framework for surface and depth recordings. A salient paradigm emerging in parallel is neural-manifold/latent-dynamics–based decoding [35,36], in which high-dimensional population activity is first mapped to a low-dimensional, smooth latent trajectory using generative or factor models, such as latent factor analysis via dynamical systems, variational recurrent/autoencoding models, or neural ordinary differential equation formulations, and a linear or nonlinear readout is then trained in this latent space to predict kinematics or discrete intent [37]. By relocating variability to a dynamical latent representation, these methods reduce channel-level fragility at the decoder interface and can support cross-session alignment. Classical linear decoders remain the clinical baseline due to their simplicity, interpretability, and low latency. Representative approaches include Wiener/Wiener-cascade regression, which maps recent firing rates to instantaneous position/velocity via causal filters; Kalman filtering/extended Kalman filtering, which models hand kinematics as a linear-Gaussian state space with neural observations to enable recursive predict-update cycles; and generalized linear models or point-process formulations, which treat spikes as an inhomogeneous Poisson process governed by conditional intensity functions that capture spike history and exogenous covariates [38–40]. The deep sequential models RNN/LSTM/GRU, temporal convolutional networks, and, increasingly, Transformers leverage attention or gated recurrence to handle long-range dependencies, benefiting discrete tasks (speech, gestures) and continuous multi-joint dynamics [41]. Manifold-centric approaches further buffer moderate nonstationarity and enable cross-day/session alignment by decoding within a task-invariant latent coordinate system [35,42]. In addition to head-to-head evaluations against linear baselines in motor cortex, recent work has proposed near real-time nonlinear dynamical decoders (eg, MINT, BRAID) designed to meet closed-loop constraints [43,44].

NONSTATIONARITY ISSUES AND CROSS-DAY CONSISTENCY:

A central challenge for implantable BCI is nonstationarity: across hours to days, firing-rate drift, changes in electrode–neuron coupling, strategy shifts, and fatigue induce train–test distribution shift degrade static decoders within days [45,46]. Contemporary advances address robustness at 3 levels. (1) Online adaptation and co-adaptation: within the closed loop, low-dose supervised or unsupervised criteria (eg, entropy minimization, pseudo-label consistency, alignment regularizers) continuously adjust the decoder, while users learn decodable control strategies under feedback; this reciprocity improves steady-state accuracy and learning rate [47]. (2) Latent alignment: nonlinear manifold alignment with dynamical consistency (eg, NoMAD) maps high-dimensional activity from a new session onto a “Day-0” latent dynamical coordinate system, thereby maintaining downstream readouts without retraining [48]. (3) Cross-subject/task transfer: multi-subject pretraining followed by rapid subject-specific finetuning establishes a population prior, reducing annotation/calibration burden for new users or tasks [49]. Empirically, these strategies deliver statistically significant gains over static linear baselines in multi-day stability, zero-calibration start-up, and relearning speed; contrastive learning and domain adaptation have also shown promise for integrating heterogeneous electrodes (ECoG/stereo-EEG) and across-subject pooling [50,51]. Real-time viability, however, hinges on low-latency implementations (closed-loop budgets typically <100 ms), motivating lightweight inference, edge deployment, and incremental updates.

MULTIMODAL SIGNAL FUSION AND ERROR/REWARD SIGNALS:

In motor restoration, a single channel or modality rarely spans all control dimensions. Early (feature-level) or late (decision-level) fusion of spikes/multi-unit activity with LFP/ECoG high-frequency power – after temporal alignment – often yields complementary benefits of precise motor commands plus robust network-state tracking [52]. When the system also incorporates electromyography or peripheral sensors, or the state of stimulators (eg, functional electrical stimulation or spinal stimulation), joint modeling can substantially mitigate failure cascades when any single modality degrades [53]. Under closed-loop operation, detecting error-related potentials or using reinforcement learning–style intrinsic rewards can drive adaptive updates and policy refinement toward individualized optimal control mappings. For multistage actions such as grasp, phase-specific β/high-γ modulations provide temporal anchors for phase-aware decoding; for speech/intent tasks, Transformer-based sequence-to-sequence models have demonstrated robustness across subjects and electrode configurations [54]. Taken together, a clinically oriented decode–feedback–retune loop is moving from algorithmic prototypes toward deployable systems.

Applications and Recent Progress

Evidence from Experimental Models

EVIDENCE FROM EXPERIMENTAL MODELS:

A substantial body of animal and primate evidence underpins the clinical feasibility, mechanistic rationale, and systems optimization of implantable BCIs for motor recovery in paralysis. Below we synthesize representative paradigms and recent advances, together with mechanistic insights and limitations.

BSI PARADIGMS: In spinal cord injury models, the BSI has emerged as a leading approach. As early as 2016, Capogrosso et al mapped motor cortical activity to epidural lumbar stimulation in nonhuman primates, restoring weight-bearing locomotion on treadmill and overground tasks within 6 days after implantation [55]. The platform combined wireless telemetry, real-time decoding, and spatially selective stimulation with an implant architecture designed for human translation. Subsequent iterations refined stimulation parameters and timing coordination to improve gait symmetry, reduce load imbalance, and enable smoother step transitions [55]. In 2023, Lorach et al reported a “naturalistic walking” BSI in monkeys that further approximated physiological gait, reinforcing the potential of BSI in ecological locomotor contexts [56].

UPPER-LIMB RESTORATION VIA CORTICAL DECODING AND STIMULATION: Analogous BCI-stimulation strategies for upper-limb recovery have also shown promise in primates. In spinal cord injury models, decoding cortical signals to drive muscular or spinal stimulation has produced controlled wrist and finger movements, with several studies noting residual improvements even after the interface was turned off, which is consistent with activity-dependent plasticity [57,58]. Earlier, Nishimura et al demonstrated a cortico–muscular bypass in monkeys, creating an artificial conduit that circumvents the injured spinal pathways to restore upper-limb function [58]. Follow-up investigations have examined stimulation modes, feedback strategies, and training regimens in relation to long-term stability, muscle fatigue, and decoder adaptability [59,60].

SMALL-ANIMAL MODELS AND NEUROMODULATORY COUPLING: In rodents, investigators frequently use cortical electrodes paired with functional electrical stimulation or stimulation of spinal segments to probe motor recovery [61,62]. Multiple studies show that motor cortical activity can drive hindlimb muscle stimulation, yielding partial improvements in gait or toe movement in rats with spinal cord injury [61,63]. Reviews by Alam et al highlight the potential of BCI–spinal stimulation paradigms to meaningfully promote functional recovery in rodent models [64,65]. These preparations also permit joint recording–stimulation protocols that map plastic feedback across cortex–spinal cord–muscle loops, offering experimental leverage to interrogate mechanisms of reorganization.

SYSTEMS-LEVEL READOUTS WITH IMAGING:

A subset of studies use imaging-based endpoints to track network-level effects following BCI implantation and intervention, including positron emission tomography and functional magnetic resonance imaging [66,67]. For example, comparative assessments of regional metabolism, perfusion, or enzymatic activity before and after BCI-driven stimulation have begun to relate stimulation locus, intensity, and timing to pathway reconnection and compensatory circuit recruitment along the motor axis [68,69]. Although still limited in number, such systems-level readouts provide a crucial bridge between mechanism and intervention design.

SYNTHESIS AND IMPLICATIONS: Collectively, these models demonstrate that implantable BCI strategies coupled to neuromodulation, especially BSI, can restore elements of motor function in animal and primate preparations. Persistent gains observed after interface deactivation support the hypothesis that BCIs may not only serve as control devices but also induce plasticity. Multi–degree-of-freedom control of complex tasks in nonhuman primates underscores the practical value of high-density electrode arrays and advanced decoders. Emerging imaging and metabolic mapping begin to link stimulation to reconfigurable distal networks, informing electrode targeting and stimulation policies [70,71]. Caveats and translational constraints. Important limitations warrant caution. Species differences in cortical architecture, spinal tract organization, and inflammatory responses limit direct translatability to humans. Many studies remain short-term; public data on months-long stability, channel attrition, and interface responses are scarce. Experimental conditions often involve constrained or highly structured tasks that diverge from everyday human motor behavior with balance demands, visual feedback, and environmental perturbations. Finally, balancing implant safety, patient tolerability, and channel failure rates remains a central challenge in advancing from laboratory proof-of-concept demonstrations to clinical deployment.

Comparison of Invasive and Noninvasive Modalities

Noninvasive EEG is widely used due to its safety, accessibility, and low cost. However, its spatial resolution is limited due to volume conduction and skull attenuation, which results in difficulty capturing high frequency motor-related activity, thus limiting its precision for fine motor control and decoding [72]. Compared with EEG, ECoG provides higher spatial resolution by recording directly from the cortical surface, offering improved signal-to-noise ratio and the ability to capture high-γ rhythms, which are important for motor control and task-related dynamics [73]. Furthermore, intracortical microelectrode arrays, while highly invasive, provide the highest resolution, allowing for single-unit and multi-unit spiking activity recordings that enable the most precise motor intent decoding and support high degree-of-freedom control. However, they come with increased surgical risks and long-term stability challenges [74].

Thus, the comparison of EEG, ECoG, and microelectrodes highlights the trade-off between invasiveness and information content. EEG offers broad accessibility but limited spatial and temporal resolution, ECoG provides higher resolution and better signal fidelity, and microelectrodes offer the highest precision but with greater invasiveness. This comparison establishes a functional bridge between signal acquisition mechanisms and decoding strategies, which is discussed in the following sections.

Key Technological Innovations

Recent technological advances are steadily propelling implantable BCIs from laboratory demonstrations toward clinical deployment. Foremost among these are innovations in electrode materials and structural design. The introduction of carbon fibers, graphene, carbon nanotubes, and conductive polymers has markedly improved the electrode–tissue interface by lowering impedance, increasing signal-to-noise characteristics, and mitigating inflammation secondary to mechanical stiffness mismatch [75-77]. Notably, ultrafine carbon-fiber microelectrodes have achieved stable recordings over several months in animal models, demonstrating superior chronic performance, compared with conventional silicon-based arrays [78]. In parallel, flexible substrates and devices with mechanically adaptive properties, such as rigidity during insertion and flexibility thereafter, better accommodate the brain’s micromotions, reducing tissue injury and prolonging recording longevity [79]. Beyond material science, endovascular approaches, such as the Stentrode platform, have emerged as minimally invasive alternatives. Early clinical studies indicate favorable safety and practicality, enabling cursor control and text entry without craniotomy [80]. Such strategies offer a lower-risk pathway for broader clinical adoption. System architecture has also undergone important shifts. As channel counts and sampling rates escalate, centralized processing struggles to meet power and latency constraints. Distributed acceleration and near-electrode edge processing using application-specific integrated circuits and field-programmable gate arrays have therefore been proposed, offloading early feature extraction to the periphery and substantially reducing bandwidth and power demands [81,82]. Concurrently, advances in surface engineering and bioactive interface design, such as coatings that elute anti-inflammatory agents or neurotrophic factors, help suppress glial responses, delay channel attrition, and extend implant lifetime [83]. Importantly, system-level closed-loop integration is gaining traction: sense–decode–stimulate frameworks not only decode motor intent but also deliver targeted feedback stimulation to promote functional recovery, thus shifting BCIs from mere motor substitution toward catalysts of motor reconstruction [84].Taken together, cross-disciplinary innovation spanning materials science, microelectronics, neurobiology, and artificial intelligence is converging to deliver higher resolution, greater stability, and improved clinical suitability for implantable BCIs, thereby laying the groundwork for wider application in individuals with paralysis [85].

Clinical Translation

OVERVIEW OF CLINICAL TRIALS:

Implantable BCIs have achieved notable progress in clinical testing, particularly in motor restoration and neurorehabilitation. Table 1 summarizes recent representative trials, spanning diverse implantable BCI platforms, patient cohorts, study objectives, and preliminary outcomes. Populations include individuals with tetraplegia, stroke, and spinal cord injury, with aims ranging from safety and feasibility assessments to exploratory evaluation of neuroplastic mechanisms. Early results suggest that implantable BCI systems can improve motor function, quality of life, and markers consistent with activity-dependent plasticity [89]. Notwithstanding these advances, important challenges remain. Chronic stability and biocompatibility constraints, inter-individual variability in neural signals and learning rate, and ethical and privacy considerations continue to limit scalability and generalizability. Moreover, many studies are characterized by small sample sizes and limited follow-up, underscoring the need for larger, multicenter trials with long-term surveillance to validate efficacy and safety. Looking ahead, broader clinical adoption will likely depend on (1) standardized protocols and outcome measures (including functional endpoints relevant to activities of daily living), (2) optimization of device design and longevity, (3) improved recruitment and retention with rigorous follow-up, and (4) proactive attention to ethical and legal frameworks. With continued maturation across these fronts, implantable BCIs are poised to play a more substantive role in neurorehabilitation and motor function recovery.

Patient Outcome Assessment

FUNCTIONAL INDEPENDENCE:

Because the core therapeutic aim is to decode neural signals and translate them into executable commands that restore motor capacity [88], outcome batteries must quantify gains in day-to-day autonomy, including eating, dressing, and ambulation. These endpoints capture real-world value beyond laboratory metrics and align with the rehabilitation goals that are meaningful to patients and caregivers.

QUALITY OF LIFE:

The clinical utility of BCIs extends to psychological well-being, social participation, and overall life satisfaction, particularly over prolonged courses of therapy [89]. Measures of quality of life contextualize functional gains, indicating whether restored capabilities translate into broader reintegration at home, work, and in the community.

PATIENT-REPORTED OUTCOMES:

Provide-reported outcomes, including perceived benefit, usability, acceptability, and comfort, directly inform adherence and long-term use [90]. Because user experience influences uptake and sustained engagement, future device design should prioritize ergonomics, ease of donning/doffing (when relevant), and cognitive load associated with operation and training.

PHYSIOLOGICAL AND NEUROBIOLOGICAL BIOMARKERS:

Objective indices provide convergent evidence of neurorestoration. Biomarkers such as brain-derived neurotrophic factor and motor-evoked potentials can reflect plasticity and corticospinal integrity [91]. For example, in a randomized controlled trial in subacute stroke, a BCI-controlled robotic training arm outperformed controls on primary functional scales, the Fugl-Meyer Assessment and Loewenstein Occupational Therapy Cognitive Assessment, and exhibited higher serum brain-derived neurotrophic factor levels, consistent with enhanced neuroplasticity [92]. Likewise, in chronic stroke, combined BCI–functional electrical stimulation interventions improved gait speed, endurance, and range of motion, with significant gains on the 10-meter walk test relative to conventional therapy [93]. These findings support the premise that BCI-based approaches, particularly when paired with adjunctive neuromodulation or robotics, can augment motor performance and may potentiate plasticity mechanisms.

Current Limitations and Paths Forward

Outcome heterogeneity remains a major barrier, since variability in scales and timing undermines cross-study synthesis, contributes to inconsistent judgments about efficacy, and impedes payer and regulatory decision-making. Small sample sizes and short follow-up intervals further limit inference about durability and safety [94]. Priorities for the field include (1) development and validation of standardized, BCI-appropriate outcome toolkits with harmonized timing and reporting; (2) routine inclusion of functional endpoints tied to activities of daily living, quality-of-life measures, patient-reported outcomes, and mechanistic biomarkers; and (3) adoption across multicenter, longitudinal studies to establish generalizable evidence. Standardization will provide the evidentiary substrate necessary for broader clinical implementation.

Discussion and Outlook

TECHNOLOGICAL ADVANCES AND OUTSTANDING CHALLENGES:

Progress in BCIs has been most evident in signal acquisition, decoding, and feedback control. On the acquisition front, advances in electrode materials and micro/nanofabrication have markedly improved recording quality and stability. The incorporation of carbon-based conductors (eg, carbon nanotubes and graphene) and related conductive polymers reduces electrode–tissue impedance and dampens inflammatory responses, thereby enhancing signal-to-noise characteristics [76,95]. Flexible and stretchable electrode platforms further accommodate brain micromotion during chronic implantation, mitigating mechanical mismatch and limiting progressive signal decline. Nevertheless, chronic stability and biocompatibility remain central hurdles, as foreign body responses, impedance drift at the tissue–electrode interface, and material aging can culminate in attenuation of signals and eventual channel failure [96]. Decoding has likewise advanced. With the adoption of deep learning, BCI systems have achieved improved accuracy, latency, and robustness. Models based on convolutional neural networks and recurrent neural networks, including LSTM/GRU variants, can extract complex motor-intent patterns from large-scale neural datasets and translate them into control commands [97,98]. However, nonstationarity and electrode drift remain persistent threats to long-term performance. Although adaptive algorithms and transfer-learning strategies have been proposed, ensuring stable decoder function over extended time scales continues to be a problem.

Clinical Translation and Current Trial Landscape

Clinically, implantable BCI systems in early-phase trials show promising motor restoration gains for paralysis [99,100]. Studies have focused on populations with spinal cord injury, stroke, and related neurological conditions, pairing cortical recordings with external effectors, such as robotic arms and exoskeletons, to achieve volitional control. Between 2020 and 2025, several trials have reported preliminary benefits. For example, combined BCI–functional electrical stimulation paradigms in chronic spinal cord injury have yielded improvements in gait, with residual benefits persisting even after the interface was deactivated, which is consistent with induced plasticity. In parallel, multiple investigations suggest meaningful improvements in quality of life and motor independence, particularly in limb function and volitional control [101,102]. Widespread adoption, however, is constrained by several factors. Most trials remain small and short in duration, limiting inference regarding durability and safety. Individual variability, device–patient matching, and ethical considerations, including privacy and informed consent for long-term implantation, further complicate translation. Addressing these gaps will require larger, multicenter, longitudinal studies with harmonized endpoints and rigorous post-implant surveillance.

Artificial Intelligence Models for Neural Decoding

Artificial intelligence has become an essential component of modern implantable BCI systems, complementing advances in electrode design and neural signal acquisition. Deep neural networks – including convolutional, recurrent, and Transformer-based architectures – can extract richer spatiotemporal representations from intracortical signals than classical linear decoders, improving accuracy and robustness in both non-human primate and human studies [103]. Latent-dynamics approaches, such as latent factor analysis via dynamical systems and related manifold models, map population activity into smooth low-dimensional trajectories that support stable decoding and cross-session alignment [104]. Artificial intelligence techniques are also used to mitigate nonstationarity, a central challenge for implantable BCIs: domain adaptation and manifold-alignment frameworks such as NoMAD stabilize decoder performance across days without extensive retraining [48]. Collectively, these approaches provide a computational foundation that enhances decoding performance and supports long-term clinical viability of implantable BCI systems.

Future Directions

Despite these challenges, the trajectory of implantable BCIs remains strongly favorable. Continued convergence of materials science, microelectronics, and neuroscience is expected to yield devices with higher precision, greater stability, and lower power consumption, potentially overcoming limits inherent to current electrode platforms. From a usability standpoint, future development will likely emphasize system simplification and convenience to enhance acceptability and adherence. Minimally invasive or microinvasive interfaces, such as endovascular electrodes or next-generation flexible implants, may reduce surgical risk and complications, widening the eligible patient pool. Finally, ethical and legal frameworks will require equal attention. Because BCI deployment entails large-scale acquisition and processing of biological data, robust safeguards for privacy and data security are essential. Sustainable clinical translation will depend not only on technical performance but also on thoughtful governance, health-economic evaluation, and societal acceptance. In summary, implantable BCIs have demonstrated substantial potential in motor restoration and neurorehabilitation. With continued technical refinement and cross-disciplinary collaboration, these systems are poised to advance from experimental paradigms to routine clinical tools, to deliver tangible functional benefits and improved quality of life for individuals with neurological injury.

Conclusions

BCI technology shows great promise for motor restoration, but challenges remain, particularly regarding electrode biocompatibility and long-term stability. Despite advancements, signal degradation and tissue response issues still limit performance. Decoders must also address inter-individual variability and nonstationarity in neural signals. While early clinical trials show positive outcomes, further standardization and long-term studies are needed. Future progress hinges on advancements in materials, artificial intelligence, and neurobiology to improve stability and decoding accuracy. Key milestones in the next 5 to 10 years include fully implantable wireless BCIs, multi-center trials, and translation to real-world applications. With continued cross-disciplinary collaboration, BCIs can evolve into catalysts for true motor recovery, advancing neurorehabilitation practices.