1. Introduction

This systematic literature review synthesizes two decades of research to assess the current state and near-term future of brain-computer interfaces. It finds that invasive systems, like intracortical implants, have achieved meaningful clinical restoration of motor and communication functions, while non-invasive methods offer reliable but lower-performance assistive control. Advances in machine learning have significantly improved neural signal decoding across all platforms. However, major engineering bottlenecks persist, including chronic device longevity, wireless bandwidth, and the body’s immune response. A critical insight is that BCI operation is a closed-loop, co-adaptive system, where user learning and decoder adaptation are interdependent, explaining challenges like performance instability. The review identifies a significant translational barrier beyond engineering: the lack of robust health economic evidence and reimbursement pathways, which is as consequential as technical hurdles. Furthermore, the field underutilizes user-centered design, risking device abandonment. For accelerated progress, coordinated investment is needed not only in biocompatibility but also in economic evidence generation, standardized outcomes, participatory co-design, and adaptive regulatory frameworks.

The capacity to establish a direct communicative channel between the human brain and external computational systems represents one of the most consequential frontiers in contemporary neuroscience and biomedical engineering. Brain-computer interfaces — devices and systems that record, interpret, and in some cases modulate neural activity to enable interaction with external hardware — have evolved from theoretical constructs into clinical realities over the course of the past two decades [1, 2]. What began as proof-of-concept demonstrations in research laboratories [3, 4] has matured into a field characterized by competing technological platforms, expanding patient populations, and an increasingly urgent set of questions about where the field is heading and how rapidly it will get there. At this particular moment, the literature surrounding BCIs is distinguished by a confluence of circumstances that makes a systematic synthesis both timely and necessary: hardware miniaturization has accelerated, machine learning methods have dramatically improved signal decoding fidelity [5, 6], high-profile commercial ventures have entered a space historically dominated by academic medicine, and regulatory bodies in multiple jurisdictions are actively grappling with frameworks that have no clear precedent [7, 8]. Compounding these technical and regulatory dynamics, the health economic and reimbursement architecture for implantable neural devices remains strikingly underdeveloped [9, 8] — a gap that threatens to determine deployment timelines as decisively as any unresolved engineering challenge. The field is, in short, at an inflection point, and the literature reflects a corresponding tension between cautious clinical incrementalism and ambitious near-term projection [10, 11].

This systematic review examines the published literature spanning 2000 through 2025, drawing on fifty-one peer-reviewed studies to synthesize the current state of BCI capabilities and to characterize the technical, clinical, economic, and regulatory landscape shaping near-term development trajectories. Four primary research questions orient the review. First, what can current invasive and non-invasive BCIs demonstrably achieve in clinical applications, and how do those achievements compare across patient populations and conditions [12]? Second, what technical barriers — in signal resolution, device longevity, and information bandwidth — continue to constrain system performance [13, 14]? Third, how are the principal technological approaches, including intracortical Utah arrays, next-generation commercial implants, electrocorticography platforms, and non-invasive modalities [15, 16], positioned relative to one another across different use cases and clinical contexts? Fourth, what regulatory pathways, health economic constraints, and surgical considerations shape the realistic timeline for broader deployment of these systems [7, 17]? The scope of the review is deliberately bounded to human applications or translational research with direct human applicability, and it prioritizes studies with clinical data, systems-level analysis, or substantive engineering characterization over purely animal or theoretical work.

The review is organized around five thematic areas, each of which addresses a distinct dimension of the BCI literature while contributing to an integrated account of the field’s current standing. The first theme addresses clinical milestones and performance across invasive interface modalities, examining what intracortical arrays, electrocorticography systems, and emerging endovascular approaches have achieved in restoring communication, motor function, and sensory feedback to individuals with paralysis and neurological injury [18, 19, 20]. The second theme turns to non-invasive approaches — primarily electroencephalography-based systems [21, 22] — and to the hybrid architectures and speech or language decoding interfaces that have emerged as a distinctive and rapidly advancing subfield in their own right [23, 24]. The third theme is cross-cutting in nature, examining the neural signal decoding and computational strategies that translate raw electrophysiological data into actionable output, including the role of deep learning, adaptive decoders, and transfer learning methods [6, 5] that have substantially raised the ceiling of what any given hardware platform can achieve. The fourth theme addresses the hardware engineering and biocompatibility challenges that ultimately determine whether a given device remains viable in the body over clinically meaningful timescales, covering electrode materials, packaging strategies, foreign body response, and the competing design tensions inherent in building for chronic implantation [13, 17]. The fifth and final theme situates the technical literature within the broader clinical translation pathway, examining neurorehabilitation applications [25, 26], trial design, health economic barriers to reimbursement, usability and adoption challenges informed by the assistive technology abandonment literature [27, 28, 29], informed consent challenges, and the evolving ethical and regulatory frameworks that will govern how and at what pace BCIs move from specialized clinical centers into wider use [8, 30].

What makes this moment in the literature distinctive is not simply that BCI systems are improving — they have been improving, steadily, for years [1, 10] — but that the rate of change across multiple dimensions simultaneously has created genuine uncertainty about which approaches will define the field’s near-term trajectory. Invasive and non-invasive approaches are no longer in a straightforwardly hierarchical relationship [15, 31]; decoding algorithms have complicated longstanding assumptions about how much resolution hardware actually needs to deliver [6, 23]; commercial development has introduced timelines and incentive structures that academic literature is only beginning to process; and the absence of established reimbursement pathways and cost-effectiveness evidence threatens to constrain deployment in ways that technical capability alone cannot overcome [9, 7]. Moreover, the broader assistive technology literature — documenting decades of experience with device abandonment, user-centred design, and the gap between laboratory demonstration and sustained real-world adoption [27, 29, 28] — provides a sobering empirical context that the BCI field has been slow to internalize [12, 8]. This review aims to provide a coherent synthesis of where the evidence currently stands.

2. Methodology

The literature underpinning this review was identified and assembled through a systematic search of the OpenAlex database, supported by structured relevance filtering, a citation expansion stage, and targeted supplementary searches addressing health economic, reimbursement, and user-centred design dimensions. The process was designed to balance breadth of coverage with precision, ultimately yielding a final corpus of 51 papers spanning publication years 2000 to 2025.

Search Strategy

The search strategy employed five targeted queries to OpenAlex, each designed to address distinct dimensions of the review’s scope. These queries covered: clinical applications of brain-computer interfaces; comparative performance across invasive and non-invasive paradigms; specific hardware platforms including Utah arrays, Neuralink, and electrocorticography systems; regulatory and surgical translation considerations; and the current state of neural interface bandwidth and technological limitations. The initial retrieval yielded 154 candidate papers. After applying a relevance scoring threshold of 0.6, 30 papers remained from keyword search alone.

To ensure comprehensive coverage, a citation network expansion was initiated to identify additional relevant work through forward and backward citation tracing. While 42 papers examined during this expansion were rejected, one additional paper was incorporated, representing a modest but non-trivial coverage delta of 0.02. The expansion process was terminated once the collection reached its predefined target of 90 candidates, which supported a final corpus of 30 papers.

A supplementary search addressed the health economic, cost-effectiveness, and reimbursement dimensions of BCI deployment. This domain was identified during initial synthesis as critically underrepresented in the primary corpus, despite its direct relevance to near-term deployment trajectories. The supplementary search targeted the intersection of neurotechnology and health economics, encompassing direct economic evaluations of neural implants, institutional and regulatory frameworks governing device reimbursement, and transferable evidence from established neuromodulation therapies including deep brain stimulation [32, 33] and cochlear implants. Eight additional papers meeting relevance and quality criteria were incorporated from this supplementary search, extending the temporal range of the corpus to 2000 to capture foundational health economic analyses that preceded contemporary BCI development.

A further targeted search addressed the usability, user-centred design, and assistive technology adoption dimensions of BCI deployment. This domain was identified as critically relevant to near-term deployment trajectories given the well-documented phenomenon of assistive technology abandonment in disability populations [29]. This search encompassed user-centred design evaluations of BCI systems [27, 30], validated adoption and usability frameworks applied to neural interfaces and related assistive devices, home deployment studies [34, 35], and participatory design methodologies for disability technologies. Eight additional papers meeting relevance and quality criteria were incorporated from this search, further extending the evidence base underpinning the clinical translation analysis.

Selection and Filtering

Papers were evaluated against several quality criteria applied consistently across the candidate pool. For older publications, a minimum citation count of five was required to ensure a baseline of disciplinary uptake — a threshold consistent with filtering practices adopted in BCI-focused systematic reviews [36, 21, 37].

To anchor the review to contemporary developments in a rapidly evolving field [38, 10, 11, 16], a recency window of two years was defined, with at least 35% of the final corpus required to fall within that window. All papers that cleared these thresholds were then assessed against the same relevance score minimum of 0.6, ensuring topical alignment across both established and emerging BCI literature [39, 12].

Two papers that initially met the selection criteria could not be retrieved and were replaced with alternative papers drawn from the remaining candidate pool, ensuring no reduction in the final corpus size.

Processing

All 51 papers in the final corpus underwent full-text analysis; no papers were processed from abstracts or metadata alone, and there were no unresolved retrieval failures following the substitutions described above. This complete full-text coverage supports the analytic depth required for a review engaging with technical performance metrics [12, 8], device architectures [16, 21], clinical translation pathways [8, 20], health economic frameworks, and usability and adoption dynamics [27, 30].

Thematic Organisation

Following processing, the corpus was organised into five thematic clusters, allowing the review’s findings to be presented along coherent conceptual lines rather than as an undifferentiated survey. These clusters reflect the primary axes of inquiry that emerged across the literature: the comparative capabilities of invasive and non-invasive paradigms [15, 40], hardware-specific performance and longevity [41, 42], signal resolution and bandwidth constraints [43, 44], clinical and regulatory translation including health economic barriers and usability challenges [8, 27, 7], and near-term technological trajectories [10, 45]. This structure was derived from the content of the assembled papers rather than imposed a priori, and it shapes the organisation of the synthesis sections that follow.

The corpus’s temporal range — from 2000 through 2025 — reflects the dual necessity of capturing foundational technical, health economic, and user-centred design work [2, 28], much of which established the empirical benchmarks and evaluation precedents against which contemporary systems are still evaluated [21, 12], while foregrounding the developments most relevant to the field’s current state and near-term direction.

3. Clinical Milestones and Performance of Invasive Brain-Computer Interfaces

The empirical record of invasive brain-computer interfaces (BCIs) spans more than two decades of iterative progress, from foundational primate laboratory demonstrations to landmark clinical trials in humans with paralysis. Across this trajectory, researchers have established increasingly precise benchmarks for motor decoding, communication restoration, and neuroprosthetic control, while simultaneously revealing the gap between current system capabilities and the demands of naturalistic human function. A 2025 systematic review and individual patient meta-analysis [46] synthesising evidence from inception to April 2024 characterised this evolution as exponential, reporting pooled performance metrics of 76% accuracy for cursor control, 80% for motor tasks, and 93.27% for communication tasks — figures that substantiate both the field’s progress and the unevenness of what different modalities have achieved.

Primate Foundations and the Translation to Human Trials

The scientific foundation for current clinical invasive brain-computer interfaces (BCIs) was established in the early 2000s through systematic primate research. A landmark 2003 study by [47] demonstrated that rhesus macaques could learn to control a robotic arm for reaching and grasping using neural ensemble activity alone, with EMG recordings confirming the absence of overt arm movements. Crucially, this work revealed that multiple cortical areas — including M1, dorsal premotor cortex (PMd), supplementary motor area (SMA), primary somatosensory cortex (S1), and medial intraparietal area (MIP) — collectively contributed to BMI control. While M1 neurons provided the strongest individual predictions, no single area supplied complete information about all motor parameters. This distributed encoding principle was later elevated to a theoretical cornerstone by [1], who argued that distributed neuronal ensembles rather than individual neurons or brain regions constitute the true functional unit underpinning optimal BMI performance. This principle directly shaped the multi-electrode array designs that subsequently entered human trials.

The first direct translation of these primate paradigms into a human clinical context was achieved by [3], who implanted a 96-microelectrode Utah array into the primary motor cortex of a 25-year-old tetraplegic patient. Remarkably, motor cortex neurons remained active and directionally selective years after spinal cord injury. Using a linear decoding algorithm to translate their ensemble activity into two-dimensional cursor control, the patient achieved 73–95% success in acquiring visual targets and operated external devices — including a prosthetic hand and robotic arm — often while simultaneously engaged in conversation. This proof-of-concept demonstrated that the neural ensemble dynamics characterized in primates generalized robustly to the injured human motor system.

Building directly on these primate paradigms, a 2006 study by [4] pushed performance benchmarks further by demonstrating information transfer rates of up to 6.5 bits per second in monkeys implanted with 96-electrode arrays in dorsal premotor cortex. Peak throughput was achievable from neural recordings as brief as 70 milliseconds. This figure — achieved in a highly constrained eight-target task under idealised laboratory conditions — remains among the highest reported for any BCI system [48]. By contrast, electrode-based prostheses in early human trials typically operated at approximately 1.0–1.5 bits per second for target selection. This performance gap underscores why comparisons between primate and human systems have since been characterized as somewhat unfair to clinical systems operating under real-world constraints.

High-Performance Communication in Human Participants

The BrainGate programme and related intracortical BCI trials constitute the primary empirical record of invasive BCI performance in humans. A 2017 study by [18] reported that three participants with paralysis, each implanted with 96-channel silicon microelectrode arrays in motor cortex, achieved typing rates of 24.4 ± 3.3 correct characters per minute during free-paced question-and-answer sessions. Notably, one participant (T6) reached 31.6 correct characters per minute using the OPTI-II keyboard layout, compared with 23.9 correct characters per minute on a QWERTY layout. This difference illustrates how interface design itself modulates measured performance. Information throughput across the three participants ranged from 1.4 to 3.7 bits per second, with a fourth participant reaching 4.16 bits per second on a denser target grid. While these rates represent a substantial achievement relative to prior clinical BCI demonstrations, they fall meaningfully short of the approximately 10 bits per second estimated for natural motor communication. Consequently, the authors’ framing of results as “high-performance” has prompted debate about the appropriate reference standards for the field [49].

A broader literature review by [46] confirms that throughput varies considerably across studies, with figures ranging from approximately 1.4 to 6.5 bits per second depending on task structure, participant characteristics, and decoding algorithms. This wide spread complicates direct cross-study comparisons. Notably, the upper bound of this range aligns with non-human primate studies using direct end-point decoding strategies, which achieved up to 6.5 bits per second (approximately 15 words per minute) with eight-target configurations using brief 70-millisecond neural integration windows [4]. This alignment underscores that paradigm design is as consequential as hardware improvements. Recent software advances, including the deployment of recurrent neural network decoders noted in [46] and supported by deep learning approaches demonstrating robust hand-kinematics decoding from spiking activity [6], have contributed to incremental throughput gains. Additionally, [50] highlighted that decoding of complex motor skills such as handwriting and speech from cortical activity — including work showing stable speech BCI decoding without recalibration in individuals with ALS [24] — has opened additional communication pathways beyond traditional cursor-and-click paradigms.

Motor and Limb Prosthetic Control

Restoring volitional limb movement in individuals with chronic tetraplegia represents a more demanding challenge than communication. The proof-of-concept demonstration by [51] in 2017 marked a decisive milestone in this pursuit. In that study, neural signals recorded from a motor cortex microelectrode array were decoded in real time to drive functional electrical stimulation (FES) of arm muscles, restoring voluntary reaching and grasping movements in a participant with chronic spinal cord injury. These findings demonstrated that motor cortical representations remain sufficiently informative years after injury to support naturalistic limb control — a conclusion with important implications for the population of potential BCI users [3]. This work effectively bridged two parallel lines of development that [1] had described as emerging but not yet converged: efferent BMI control and peripheral stimulation for functional restoration. Subsequent research has extended this approach beyond FES alone. For example, Benabid et al. [52] demonstrated full-body exoskeleton control driven by an epidural wireless brain-machine interface in a tetraplegic patient, while Samejima et al. [53] developed a brain-computer-spinal interface that routes cortical signals directly to spinal stimulation targets, bypassing the lesion site and thereby broadening the landscape of viable efferent restoration strategies.

Despite these advances, the gap between offline decoding capability and online BCI control remains a persistent concern. [49] explicitly noted that while offline analyses demonstrate multi-degree-of-freedom hand grasp decoding, no BMI user in an online control setting had exceeded one degree of freedom for grasping at the time of that review. Tam et al. [5] similarly identified this offline-online discrepancy as a central unresolved issue in their survey of human motor decoding from neural signals, underscoring its durability as a field-wide limitation. This discrepancy between laboratory decoding demonstrations and real-world controllability represents one of the field’s most consequential unresolved tensions, and subsequent years have not yet produced published evidence of sustained multi-DOF online hand control in humans. As Section 5 develops in greater detail, this offline-online gap has a principled explanation within the co-adaptive learning framework: decoders trained on neural data collected without closed-loop feedback are fitted to a distributional regime that shifts once the user engages in real-time control. This shift occurs because cortical representations adapt to the new feedback environment in ways that systematically diverge from the training distribution [54, 1, 55].

Bidirectional Interfaces: Integrating Sensory Feedback

The incorporation of sensory feedback has become increasingly recognized as an essential dimension of invasive BCI performance for functional restoration. Early BMI systems operated in an open-loop fashion, relying entirely on visual feedback to guide prosthetic movements [47, 3]. A systematic review by [56] analyzed how direct electrical stimulation (DES) via electrocorticographic (ECoG) electrodes can generate artificial somatosensory percepts by stimulating primary somatosensory cortex, thereby enabling closed-loop BCI control with tactile feedback.

DES offers several advantages over non-invasive neurostimulation approaches, including greater spatial precision, deeper targeting, and reduced off-target stimulation effects [57]. Complementing this approach, intracortical microstimulation (ICMS) of somatosensory cortex has been shown to evoke stable, graded tactile percepts over extended implant durations exceeding 1,500 days [41], demonstrating the long-term viability of sensory restoration via direct neural stimulation.

The integration of DES into BCI systems represents both a technical and conceptual shift from unidirectional decoding devices to genuinely bidirectional cortical interfaces. This development was framed by [49] as necessary for naturalistic limb function, since feedforward motor control alone is insufficient for dexterous manipulation. Beyond its functional importance for motor dexterity, sensory feedback may also be theoretically necessary for stable co-adaptive learning between the user and the decoder [2]. This point is increasingly supported by the computational methods literature and will be elaborated in Section 5.

Gaps, Limitations, and the Frontier in 2025

Despite documented milestones in brain-computer interface (BCI) research, several structural limitations constrain the conclusions that can be drawn from the existing evidence base. A 2025 meta-analysis by [46] estimated that only approximately 80 individuals worldwide have received implantable BCIs, severely limiting the generalisability of any performance claims. Furthermore, long-duration stability data — essential for evaluating whether intracortical arrays, ECoG grids, or intravascular stentrodes maintain performance over clinically meaningful timescales — remain sparse across all modalities [50, 46].

Where such stability data do exist, the findings are sobering. Hughes et al. [41] documented signal quality declines of 47–72% across stimulated and non-stimulated microelectrode arrays over 1,500 days in a single participant. This result underscores that even the most carefully monitored implants face substantial degradation trajectories, and that findings from single participants cannot be readily generalised [13].

The emergence of intravascular stentrodes as a less invasive alternative to open-craniotomy implants was noted by [46] as a meaningful hardware advance. However, head-to-head comparisons between stentrodes, ECoG arrays, and intracortical microelectrode arrays under matched task and population conditions are absent from the literature, making modality-level performance comparisons inferential rather than empirical.

Similarly, [1] observed that the field’s reliance on small, highly selected participant populations, combined with non-standardised outcome measures, makes it difficult to assess whether reported performance benchmarks reflect the upper bound of system capability, the average user experience, or simply the best sessions from the most capable participants. A systematic review of implantable BCI outcome measures across 77 studies and 53 unique participants found that engineering metrics dominated reporting — appearing in over 76% of studies — while only 22% included any clinical outcome measure. Those studies that did include clinical measures employed 20 different instruments with substantial heterogeneity, leaving no consensus on endpoints suitable for regulatory evaluation [12].

Collectively, these gaps underscore that the empirical backbone of invasive BCI research — while genuinely impressive in what it has demonstrated — remains structurally fragile as a basis for broad clinical or translational claims.

4. Non-Invasive BCI Approaches, Hybrid Systems, and Speech/Language Interfaces

Non-invasive brain-computer interfaces (BCIs) represent the dominant mode of BCI research and deployment by volume, offering accessibility and safety advantages that invasive approaches cannot match. This section traces the conceptual and technical evolution of non-invasive BCIs—from foundational EEG signal acquisition principles through hybrid architectures, imagined speech paradigms, and intelligent language decoding frameworks—examining how these strands have converged, diverged, and been substantially reoriented by a wave of publications in 2025.

EEG Signal Acquisition: Foundations and Persistent Constraints

The intellectual lineage of non-invasive BCI research extends back over a century. Early investigations into electroencephalography during the nineteenth and early twentieth centuries established that the brain’s electrical activity could be measured from the scalp, and the first BCI demonstrations in the 1970s built directly on this foundation [58]. Throughout this history, a characteristic trade-off has remained constant: EEG offers high temporal resolution, low cost, portability, and complete absence of surgical risk, but is constrained by limited spatial resolution and vulnerability to artifacts from muscle activity, eye movements, and electromagnetic interference [58, 16]. These fundamental constraints are not merely technical inconveniences—they shape the entire architecture of non-invasive BCI research, motivating artifact mitigation pipelines, multi-modal fusion strategies, and the search for robust feature representations that can survive the noise floor of scalp recording [16]. Notably, no universal gold standard for artifact removal has yet emerged; techniques range from simple bandpass filtering to independent component analysis and machine learning-based approaches, with selection depending on signal characteristics, channel count, and real-time processing requirements [16].

A comprehensive survey covering 220 peer-reviewed studies from 2014 to 2024 illustrates how these constraints have been operationalized across healthcare applications. For example, BCI-controlled wheelchair systems have achieved classification accuracies as high as 96.9% using support vector machine algorithms, and emotion recognition from combined EEG and galvanic skin response data has reached 84.5% accuracy [11]. Such figures reflect genuine engineering progress, yet they also obscure a persistent gap between controlled laboratory performance and real-world deployment [16, 59, 21]—a point returned to below.

Hybrid BCI Architectures: EEG-EMG Fusion and Graceful Degradation

One influential response to the limitations of unimodal EEG has been the development of hybrid BCIs, which fuse brain signals with peripheral physiological measures. Foundational work in this area, articulated in a framework paper by Müller-Putz and colleagues [31], established both the conceptual vocabulary and empirical evidence base for hybrid systems. Crucially, their experiments demonstrated that combining EEG with electromyography (EMG) produced gradual rather than abrupt performance degradation as EMG signals were progressively attenuated. This property, described as “graceful degradation,” makes hybrid systems particularly attractive for patients with progressive motor diseases such as amyotrophic lateral sclerosis, where residual muscular activity diminishes over time [31]. The same framework showed that simultaneous detection of motor imagery commands and error-related potentials could reduce online BCI error rates from approximately 32% to 7%, demonstrating that hybrid signal fusion improves not only robustness but also reliability under operational conditions [31].

While the graceful degradation narrative has been influential, it carries an important caveat that the field has not fully resolved. Although hybrid BCI literature claims smooth performance transitions, the achievable performance floor with pure EEG—the endpoint state for a fully paralyzed user—remains substantially lower than hybrid modes [60, 40]. This asymmetry means that hybrid architectures defer rather than solve the fundamental problem of pure EEG-based communication for late-stage motor disease. ALS progression can ultimately leave patients in a completely locked-in state in which even oculomotor control is compromised, rendering eye-gaze alternatives unavailable for 10–17% of advanced patients [20]. This limitation further underscores the inadequacy of any residual-muscle-dependent strategy. Longitudinal evidence demonstrating that patients successfully transition through degradation stages in clinical practice remains sparse [20, 60], and the broader translational literature highlights that limited device availability continues to prevent the large-scale trials needed to validate long-term hybrid BCI use in real-world settings [20].

Non-Invasive Brain Stimulation as a BCI Enhancement Strategy

A parallel development in the 2025 literature extends the hybrid logic beyond signal fusion to incorporate non-invasive brain stimulation (NIBS) as an active component of BCI-adjacent systems. Rudroff [61] proposes a bidirectional framework integrating transcranial direct current stimulation (tDCS) and transcranial alternating current stimulation (tACS) with artificial intelligence to address both speech production and auditory processing deficits simultaneously. In this framework, tDCS enhances cortical excitability in regions supporting speech production, while tACS entrains auditory oscillations to improve speech discrimination in impaired listeners—effectively using stimulation to shape the neural state that the BCI must then decode or support [61].

This coupling of stimulation with decoding reflects a broader closed-loop logic in which real-time neural state information guides when and how stimulation is applied, a design principle that has been explored more generally in non-invasive closed-loop neuroscience [62]. The bidirectional conceptualization treats the brain as both the signal source and a modifiable target, representing a significant departure from classical BCI architectures that treat the neural substrate as fixed [2].

The conceptual framework raises important empirical questions about the performance gains from NIBS: whether these gains are durable, generalize across individuals, and survive repeated stimulation sessions. These challenges apply to closed-loop stimulation paradigms broadly [62], but the framework itself does not resolve them empirically.

Imagined and Inner Speech BCIs: Trainability and Neural Correlates

Among the most scientifically rich and practically ambitious non-invasive BCI paradigms is imagined or inner speech decoding—the attempt to classify mental phonation without overt articulation or motor output. The appeal is self-evident: a BCI that responds to silent speech would offer natural communication without requiring any residual motor function, making it particularly relevant for patients with severe communication disorders [63]. Early work in this space was largely exploratory, demonstrating that imagined speech produces detectable EEG signatures but offering limited insight into whether users could learn to produce more decodable signals through training [23]. Early systems operating in controlled settings achieved binary classification and simple command recognition, with accuracy rates around 80–100% in basic paradigms, yet were constrained by signal quality dependence, poor cross-subject generalization, and high temporal delays [23].

A 2025 study by Bhadra, Giraud, and Marchesotti directly addresses this gap [64]. Fifteen healthy participants trained over five consecutive days to operate a binary EEG-based imagined speech BCI with continuous real-time feedback, and performance improved significantly across the training period. Crucially, the study identified the neural mechanisms accompanying skill acquisition: frontal theta power increased—interpreted as indexing heightened cognitive engagement—while focal low-gamma activity emerged in temporal regions associated with speech-specific neural processing [64]. This dissociation suggests that users are not simply learning to perform the task better in a generic cognitive sense, but are genuinely tuning speech-relevant neural circuits. The finding that imagined speech BCI skills are trainable, with identifiable electrophysiological correlates of learning, substantially advances the theoretical grounding of the field. These user-side learning dynamics constitute one component of the co-adaptive process that characterizes closed-loop BCI operation — a framework developed more fully in Section 5 — in which the user’s neural representations and the decoder’s parameters are coupled through feedback and evolve interdependently. Evidence that passive somatosensory stimulation can prime sensorimotor cortex and enhance event-related desynchronization during subsequent motor imagery [65] further reinforces the view that the neural substrate contributing to BCI control is actively modulable rather than fixed, and that interventions targeting user-side adaptation can meaningfully shape BCI performance. Nevertheless, the study’s binary paradigm, five-day window, and healthy-participant sample leave unanswered whether the same learning dynamics operate over longer periods, in clinical populations, and across vocabularies of naturalistic size. More advanced imagined and silent speech decoding systems have demonstrated scaling toward larger vocabularies—Herff et al.’s “brain-to-text” system achieved word error rates as low as 25% for 100-word vocabularies, and Anumanchipalli et al.’s neural synthesis approach yielded correlation scores of 0.69 against actual speech patterns during silent miming [23]—underscoring how much remains to be bridged between controlled binary paradigms and clinically viable natural communication.

Feature Extraction Debates: Multi-View Graph Fusion for Speech Imagery

A persistent methodological controversy concerns which EEG features best support speech imagery decoding. No consensus currently exists: different studies have favored different frequency bands, spatial configurations, and feature extraction approaches, making cross-study comparison difficult and progress incremental [63]. Feature extraction methods span time-domain, frequency-domain, and time-frequency approaches — each capturing different aspects of the non-stationary EEG signal — and purely temporal or spectral analyses have individually proven insufficient for robust decoding [16]. Empirical work on imagined speech has confirmed that discriminative features are distributed across multiple spectral scales simultaneously: alpha and low-beta activity (8–16 Hz) over left central regions and gamma-band activity over bilateral temporal regions both contribute to syllable discrimination, and the relative weighting of these contributions shifts meaningfully as users gain experience with a BCI system [64]. This fragmentation reflects both the genuine complexity of imagined speech neural dynamics and the absence of standardized benchmarks across research groups.

A 2025 contribution by Zhao and colleagues attempts to move beyond single-domain representations by proposing a multi-view graph fusion (MVGSF) model that integrates features extracted from multiple EEG domains while adaptively weighting their contributions to decoding [63]. This self-weighting mechanism allows the model to identify which feature views are most informative for a given participant or task, rather than committing to a fixed feature set. The approach positions speech imagery as a more natural and flexible BCI control paradigm than alternatives such as steady-state visually evoked potentials (SSVEP) or P300, which impose strong constraints on user attention and environment [63, 16]. While the multi-view approach represents a principled response to the feature consensus problem, the absence of standardized benchmarks means that performance comparisons across systems remain difficult to interpret.

The ICI Framework: From Signal to Intelligent Communication

The most architecturally ambitious 2025 contribution to non-invasive speech BCI is a systematic review proposing a three-stage developmental framework—designated the ICI architecture—for BCI-based language decoding [23]. The first stage, Signal Interpretation, encompasses direct EEG-to-character or word decoding with processing latencies in the range of 3–26 seconds; the second stage, Dynamic Communication, incorporates adaptive language models and contextual inference to accelerate output; the third stage, Intelligent Interaction, envisions multimodal integration and closed-loop AI assistance, reducing latencies to 50–235 milliseconds despite handling substantially more complex linguistic tasks [23]. This compression of latency across stages—by orders of magnitude—is presented as evidence that the bottleneck shifts from signal processing to linguistic modeling as the framework matures.

The ICI architecture is valuable as an organizing schema, but it is explicitly developmental: most non-invasive systems currently operate firmly in Stage 1, and the transition pathways between stages remain largely uncharted. The framework implicitly acknowledges the scalability problem that runs through imagined speech research—small vocabularies, short training periods, and laboratory settings—without resolving it. Contemporary EEG-based imagined speech systems, for instance, commonly operate on binary or very small phoneme sets [63], and performance gains across training sessions are marked by substantial inter-individual variability, with accuracy rising from roughly 55% to 70% over five days of practice even under optimal closed-loop feedback conditions [64]. These empirical constraints on current Stage 1 systems underscore why the ICI framework’s more advanced stages remain aspirational rather than operational.

Gaps, Limitations, and the Real-World Chasm

Across these sub-themes, a cluster of overlapping limitations recurs. Laboratory performance metrics, even impressive ones, do not straightforwardly predict real-world usability [11, 58]. Most imagined speech and hybrid BCI studies are conducted under controlled conditions with trained research staff, calibrated equipment, and compliant participants—conditions that differ substantially from the home environments of potential end users [25, 37]. Signal artifacts from everyday movement, variable ambient noise, and the absence of expert supervision all degrade system performance in ways that controlled studies rarely capture [16]. Standardized benchmarks for comparing non-invasive speech and language BCIs across laboratories are lacking [63, 23], making cumulative scientific progress harder to measure than publication volume might suggest; efforts to establish common reporting standards for neurally controlled systems remain aspirational rather than implemented [66]. And while NIBS-augmented frameworks [61] and multi-view feature fusion [63] represent genuine conceptual advances, the validation evidence for each remains preliminary, typically resting on small participant samples tested in single-session laboratory protocols. The 2025 literature has substantially enriched the theoretical landscape of non-invasive BCI—particularly in imagined speech trainability [64], bidirectional stimulation frameworks, and language decoding architectures—but has not yet closed the distance between laboratory demonstration and naturalistic, at-home communication assistance.

5. Neural Signal Decoding and Computational Methods

Neural signal decoding sits at the intersection of neuroscience, signal processing, and machine learning, representing the computational heart of any brain-machine interface. The central challenge is one of translation: converting the stochastic, high-dimensional activity of neural populations into continuous, discrete, or linguistic outputs with sufficient speed, accuracy, and robustness for practical use. Early work in this domain relied on relatively simple linear estimators and population statistics, but the field has progressively layered greater algorithmic complexity onto richer signal representations, culminating in deep learning architectures capable of capturing temporal dependencies that classical methods cannot. Understanding this trajectory—and the unresolved tensions it has generated—requires tracing both the evolution of decoder design and the parallel debates about which neural signal types are worth decoding in the first place. Critically, however, no account of decoder performance is complete without recognizing that BCI decoding is not a unidirectional estimation problem but one half of a coupled dynamical system in which the user’s own neural adaptation constitutes the other half. This co-adaptive framework is the organizing theoretical lens for the entire section: it is introduced directly following the classical decoding overview, and each subsequent empirical sub-theme—signal-type comparisons, deep learning decoders, and the field’s persistent gaps—is evaluated through it. Foregrounding the co-adaptive perspective at the outset is not a rhetorical choice but a substantive one: many of the field’s most persistent anomalies, from the offline-online performance gap to BCI illiteracy, become structurally coherent when viewed as consequences of coupled adaptive dynamics rather than as isolated technical shortcomings.

Classical Decoding Frameworks and Their Limitations

The foundational era of computational BMI research demonstrated that meaningful motor control could be extracted from multi-electrode recordings using relatively tractable statistical methods. The landmark study by [47] showed that rhesus macaques could operate a robotic arm for reaching and grasping using population vector algorithms and Wiener filters applied to spiking activity recorded across M1, PMd, SMA, S1, and MIP simultaneously. Crucially, this work established two principles that would shape subsequent decoder development: first, that the functional unit relevant to decoding is a distributed neuronal ensemble spanning multiple cortical areas rather than any single region [47]; and second, that each cortical area contributes information about all motor parameters, precluding a clean modular decomposition. This distributed ensemble perspective was later elevated to a theoretical principle by [1], who argued comprehensively that optimal BMI performance necessarily requires sampling large numbers of neurons across multiple brain areas concurrently, and that sustained BMI operation reshapes cortical representations through plasticity, gradually incorporating artificial actuators into the brain’s body schema.

Within this classical paradigm, Kalman filters and Wiener filters became the workhorses of offline and online decoding alike [5]. The Kalman filter’s state-space formulation accommodated the temporal dynamics of movement trajectories, while the Wiener filter offered a computationally lightweight linear regression approach. Unscented Kalman filters extended the framework to handle mild nonlinearities. However, a persistent limitation of these classical approaches is their assumption of stationarity and linear input-output relationships—assumptions that neural population codes routinely violate [6]. Neural representations are subject to session-to-session drift and long-term reorganization, necessitating frequent recalibration that linear decoders are ill-equipped to handle automatically [55]. These methods also struggle to exploit the long-range temporal dependencies present in neural spike trains [6], an architectural constraint that would later become a focal point for deep learning alternatives.

Information-theoretic analyses revealed how constrained classical decoders were in practice. [4] demonstrated that with 96-electrode arrays in dorsal premotor cortex and highly optimized neural population vector decoding, information transfer rates could reach approximately 6.5 bits per second using recordings as brief as 70 milliseconds—a result that stood as a benchmark for nearly two decades. Yet [43] placed this figure in a sobering context: even this relatively high-performing system falls short of the roughly 10 bits per second required for natural human communication, and most invasive BMI systems achieve substantially less than 3 bits per second. The electrode-tissue impedance mismatch further compounds these limitations, as the mechanical rigidity of conventional silicon and metal electrodes triggers neuroinflammatory responses that progressively degrade signal quality over months to years [43, 67, 13]. Chronic implants face a cascade of foreign body responses—including glial scarring and neuronal die-back around electrode shanks—that erode signal-to-noise ratios independently of any decoder improvements [67]. This bandwidth ceiling is not merely an algorithmic problem; it reflects fundamental constraints in signal acquisition that decoder sophistication alone cannot overcome.

BCI Operation as a Coupled Co-Adaptive System

A theoretical reframing that provides mechanistic coherence to several otherwise disparate observations across the BCI literature treats brain-computer interface operation not as a unidirectional decoding problem but as a closed-loop co-adaptive system in which the user’s neural representations and the computational decoder constitute coupled adaptive agents [28]. In this framework, the user is not a static signal source whose patterns the decoder must passively estimate; rather, the user is an active learner whose cortical representations shift in response to feedback, task demands, and the decoder’s behavior, while the decoder—if adaptive—simultaneously adjusts its parameters in response to the user’s changing neural output. The dynamics of BCI control thus emerge from the interaction between these two learning processes, and the stability, performance, and generalizability of any BCI system depend on how well this coupled adaptation is managed. The empirical subsections that follow—on signal-type comparisons, deep learning architectures, and persistent gaps—each carry implications that are most fully understood through this co-adaptive lens.

The empirical foundations for this perspective are now substantial. Longitudinal evidence from the Cybathlon BCI competition demonstrated that two tetraplegic participants, trained over several months with a motor imagery BCI incorporating systematic mutual learning protocols, achieved dramatic improvements in both command accuracy (from 53.8% to 93.8% and 81.9% to 96.8%) and race completion times [54]. Crucially, the improvements reflected changes on both sides of the interface: users developed more discriminable sensorimotor rhythm modulation patterns—a neurophysiological adaptation measurable in EEG—while the decoder and application design were iteratively refined to accommodate the evolving signal characteristics [54]. This mutual adaptation underscores that BCI performance is not a fixed property of either the hardware or the user but an emergent property of their coupled dynamics. The study’s rigorous longitudinal design, conducted under real-world competitive conditions with tetraplegic participants rather than healthy laboratory volunteers, provides unusually strong ecological validity for the co-adaptive framework.

The neurofeedback literature provides theoretical scaffolding for understanding the user-side learning component. Neurofeedback protocols operate through a closed-loop BCI architecture comprising signal acquisition, online preprocessing, feature extraction, feedback signal generation, and the active learner—with operant conditioning as the primary learning mechanism [68]. Critically, the temporal constraints on this loop are stringent: feedback delays exceeding 250–350 milliseconds degrade conditioning efficacy, meaning that the decoder’s latency characteristics directly modulate the user’s capacity to learn [68]. Individualized feature extraction further improves neurofeedback efficacy, reflecting the substantial variability in peak frequency and other neural characteristics across individuals [68]. These findings have important implications for deep learning decoders whose computational requirements may introduce latencies that interfere with the very learning processes they depend upon.

Zrenner and colleagues formalized two distinct closed-loop configurations relevant to co-adaptive BCI: a “brain-state dynamics” loop, in which stimulation or feedback is triggered by instantaneous neural states measured in real time, and a “task dynamics” loop, in which the brain is embedded within a dynamically changing environment that creates goal-directed behavioral coupling [62]. Effective BCI operation, particularly for motor and communication applications, requires both loops to operate concurrently—the user must modulate neural states in response to moment-to-moment feedback (brain-state loop) while pursuing task-level goals that structure the learning trajectory (task loop). This dual-loop architecture explains why BCI performance is often fragile under conditions that disrupt either loop—for instance, when visual feedback is occluded, when task demands shift unexpectedly, or when decoder recalibration changes the mapping between neural patterns and output. He and colleagues similarly emphasized that strategies targeting the “brain side” of the interface—including EEG source imaging to improve signal resolution, tDCS to modulate cortical learning, and mind-body awareness training to improve user skill acquisition—can substantially enhance sensorimotor rhythm-based BCI performance by optimizing the user’s contribution to the co-adaptive loop [69].

The co-adaptive framework illuminates why the offline-online performance gap—documented in Section 3 as a persistent tension in the invasive BCI literature [49]—is not merely a technical artifact but a structural consequence of how decoders are typically trained. Offline decoders are fitted to neural data collected under conditions where the user has no access to closed-loop feedback and therefore no opportunity to engage in the co-adaptive learning process. Once the loop is closed, the user’s cortical representations shift in response to the decoder’s output, producing neural patterns that diverge from the training distribution. This distributional shift is not noise; it is the expected consequence of a learning system responding to a new feedback environment. The severity of this problem is underscored by work showing that neural population activity drifts continuously across days and weeks, requiring ongoing decoder recalibration to maintain stable performance—Wilson and colleagues demonstrated that unsupervised recalibration algorithms applied to cursor BCIs could sustain long-term performance precisely by tracking this non-stationarity without requiring explicit retraining sessions [70]. Degenhart and colleagues further showed that stabilizing a BCI by aligning low-dimensional neural activity subspaces across sessions—rather than fitting a decoder to absolute firing patterns—substantially reduced the performance degradation attributable to neural drift [55]. [1] anticipated this dynamic in arguing that sustained BMI operation reshapes cortical representations through plasticity, gradually incorporating artificial actuators into the brain’s body schema. Decoders that cannot accommodate this shift will underperform their offline benchmarks, not because they are poorly designed, but because they were trained on data from a fundamentally different dynamical regime.

This perspective also connects directly to the user training findings discussed in Section 4. The demonstration that imagined speech BCI skills are trainable, with frontal theta increases indexing cognitive engagement and temporal low-gamma emergence reflecting speech-specific processing [64], represents the user-side adaptation component of the co-adaptive system. Sensorimotor priming—such as the passive somatosensory stimulation shown by Kusano and colleagues to enhance event-related desynchronization in both alpha and beta bands over sensorimotor cortex [65]—further demonstrates that the user’s neural readiness for BCI control can be actively shaped through feedback and stimulation, reinforcing the view that the neural substrate is not a fixed signal source but a plastic, modulable component of the interface.

The co-adaptive framework carries a further important implication for bidirectional interfaces. If stable BCI performance depends on coupled learning between user and decoder, then the quality and informativeness of the feedback channel constraining the user’s adaptation becomes a critical system parameter. Open-loop BCIs, which rely entirely on visual feedback, provide the user with information about decoder output but not about the decoder’s internal confidence or about how specific neural pattern changes map to output corrections. Somatosensory feedback delivered through direct electrical stimulation [56] or peripheral stimulation [65] provides a richer, lower-latency feedback channel that may support more efficient co-adaptive learning by enabling the user to form tighter associations between neural modulation and its perceptual consequences. This argument provides a theoretical justification for bidirectional interfaces that extends beyond the functional requirement for tactile sensation in dexterous manipulation: sensory feedback may be necessary for the user-side learning that sustains stable long-term BCI performance.

Not all users, however, are equally equipped for co-adaptive learning. BCI illiteracy—the inability of 15–30% of individuals to generate sufficiently robust brain signals for effective BCI operation—represents a fundamental constraint on the universality of the co-adaptive framework [10]. Neurobiological predictors of BCI proficiency, including gray matter volume in sensorimotor areas, spectral entropy, power spectral density from resting-state EEG, and corticospinal excitability, suggest that individual differences in cortical plasticity and baseline neural dynamics determine the user’s capacity to participate in the co-adaptive loop [10]. Psychological factors including attention, fatigue, and personality traits further modulate this capacity [10, 71], indicating that user-side adaptation is not purely a function of neural hardware but depends on cognitive and motivational states that vary both across and within individuals. These individual differences imply that co-adaptive BCI systems will need to accommodate not only decoder-side variability but also heterogeneous learning dynamics on the user side—an optimization problem substantially more complex than fitting a decoder to a stationary signal source.

Signal Type Comparisons: Challenging the Spike-Sorting Orthodoxy

With the co-adaptive framework established, a central question follows: which neural signal representations are best suited to support stable co-adaptive learning over time? This question reframes what might otherwise appear to be a purely technical debate about signal preprocessing choices. Conventional wisdom long privileged single-unit activity (SUA) as the gold standard, on the grounds that isolated single neurons provide the highest-fidelity representation of motor intent [1]. This view implicitly treated spike sorting—the computationally expensive process of isolating individual neuronal waveforms—as a prerequisite for high-performance decoding [48]. The review by [5] reflected this orthodoxy while also highlighting the complementary role of ECoG signals, noting that high-gamma band activity (≥75 Hz) and local motor potentials offer rich information for hand movement and finger trajectory decoding without requiring penetrating electrodes.

This consensus was challenged directly by [6], whose systematic comparison of all combinations of four neural signal types—SUA, multi-unit activity (MUA), local field potentials (LFP), and entire spiking activity (ESA)—across multiple decoder architectures revealed that ESA, which aggregates threshold-crossing events without waveform discrimination, consistently outperformed all other signal types across all decoders, sessions, and electrode configurations. This result is practically significant: ESA requires minimal preprocessing, is robust to electrode drift, and avoids the computational overhead and potential errors of spike sorting. The implication is that the information content relevant to hand kinematics decoding is not lost—and may even be enriched—by retaining the population-level aggregate rather than attempting to resolve individual units. Viewed through the co-adaptive lens, ESA’s robustness to electrode drift is particularly significant: chronic electrode recordings are known to suffer progressive signal degradation and unit loss over time [41, 7], and because the user’s neural patterns shift during closed-loop operation, a signal representation that is inherently tolerant of population-level changes may better support stable co-adaptation than one that depends on maintaining consistent isolation of individual neurons across sessions [55]. This finding does not necessarily generalize to tasks requiring discrimination of fine-grained single-neuron tuning differences, but it substantially weakens the argument that spike sorting is a universal prerequisite for high-performance motor BMI systems.

Deep Learning Decoders: Gains and Open Questions

The same study that challenged the spike-sorting orthodoxy also delivered one of the more rigorous comparisons of classical and deep learning decoders for motor kinematics [6]. The quasi-recurrent neural network (QRNN) architecture significantly outperformed all classical decoders—Kalman filter, Wiener filter, Wiener cascade filter, and unscented Kalman filter—as well as other recurrent architectures including simple RNNs, LSTMs, and GRUs. The QRNN’s advantage appears to derive from its capacity to model long-range temporal dependencies in neural spike trains more efficiently than standard LSTMs, while maintaining lower computational overhead than full attention mechanisms. This offline superiority of deep learning over classical methods has been replicated across motor, speech, and language decoding domains [5, 23], with reviews of human motor decoding noting that deep neural networks consistently outperform linear and Kalman-based approaches across intracortical, ECoG, and peripheral recording modalities [5].

In the language decoding literature, [23] proposed a three-stage developmental framework—signal interpretation, dynamic communication, and intelligent interaction—that organizes the field’s progression from simple command decoding toward naturalistic linguistic interaction. A notable empirical signature of this progression is the dramatic reduction in processing latency: from 3–26 seconds in early systems to 50–235 milliseconds in current architectures, despite the latter handling substantially more complex linguistic structures. This trajectory underscores how algorithmic and hardware co-evolution has enabled deep learning pipelines to approach real-time operation even for speech decoding. Complementary work on stable speech BCI decoding has further demonstrated that deep learning models can maintain high performance over extended periods without recalibration [24], reinforcing their practical viability for clinical deployment.

For non-invasive decoding, the challenge shifts from signal degradation over time to the inherent low spatial resolution and susceptibility to noise of scalp EEG [16]. [63] addressed this directly for speech imagery decoding by proposing a multi-view graph fusion model (MVGSF) that integrates features from multiple signal domains—time, frequency, and spatial—while adaptively weighting their contributions. This approach reflects a broader trend toward ensemble feature representations in EEG-based BCI [16, 64], acknowledging that no single feature domain captures sufficient discriminative information for reliable speech imagery classification. The absence of consensus on optimal EEG feature representations for this task [63] signals that the field remains in an exploratory phase for non-invasive language decoding.

Persistent Gaps and Unresolved Tensions

Despite these advances, several foundational limitations remain—many of which are clarified and unified by the co-adaptive framework developed above. The demonstrated offline superiority of deep learning decoders over classical methods has not been systematically validated in real-time closed-loop operation, where factors such as non-stationarity, latency constraints, and the co-adaptive dynamics between decoder and user introduce challenges that offline benchmarks cannot capture. As the neurofeedback literature makes explicit, feedback timing is not merely a computational convenience but a determinant of the user’s learning efficacy: delays beyond 250–350 milliseconds compromise the operant conditioning mechanisms through which users refine their neural control strategies [68]. The computational and energy costs of running LSTM or QRNN models on embedded, implantable hardware represent a genuine barrier to clinical translation that the literature has not adequately addressed [16]—and the co-adaptive perspective sharpens this concern, because a decoder that is computationally powerful but introduces latency may paradoxically impair the user-side learning it depends on for sustained performance.

The co-adaptive framework also reframes the problem of cross-session and cross-individual decoder adaptation. Neural signals drift across days and weeks not only because of electrode migration and biological processes [72] but also because the user’s cortical representations are themselves adapting—sometimes in productive directions that improve control, sometimes in directions that diverge from the decoder’s expectations. Long-term recalibration work has shown that unsupervised alignment of low-dimensional neural activity subspaces can partially mitigate this drift [55, 70], yet these approaches still do not fully account for the volitional, learning-driven component of representational change. The Cybathlon longitudinal data [54] demonstrate that when mutual adaptation is explicitly managed—with coordinated updates to the decoder, the application interface, and the user’s training regimen—dramatic performance gains are achievable over months. Conversely, when only the decoder adapts (or only the user adapts), the resulting mismatch between the two adaptive agents can produce instability rather than improvement. Transfer learning and cross-session decoder adaptation—critical for any system that must remain functional across months or years of use—are thus not purely algorithmic challenges; they are co-adaptation management problems that require explicit modeling of the user’s learning trajectory alongside the decoder’s parameter updates.

Peripheral nerve decoding, identified by [5] as a promising but underdeveloped modality, continues to suffer from signal-to-noise ratios compromised by tissue attenuation and foreign-body inflammatory responses at the electrode interface [72], with signals in the 5–20 μV range that remain difficult to decode reliably compared to surface electromyography [43]. The 15–30% BCI illiteracy rate reported across the broader literature [10] further complicates generalization: if a substantial minority of potential users cannot generate sufficiently robust signals to participate effectively in co-adaptive learning, then decoder sophistication alone will not achieve universal coverage, and the field must develop complementary strategies—potentially including neurostimulation-based priming [69, 65] or alternative signal paradigms [25]—for this population. These gaps collectively suggest that while the algorithmic frontier has advanced considerably, the path from laboratory demonstration to durable clinical utility requires as much progress in managing the coupled dynamics of user-decoder co-adaptation as in raw decoding accuracy.

6. Neural Interface Hardware, Biocompatibility, and Device Longevity

The engineering of neural interfaces sits at an intersection of materials science, neurobiology, and surgical robotics, where hardware design choices cannot be disentangled from the biological failure modes they must overcome. Understanding how these two dimensions interact has been a central preoccupation of the field for decades, but the complexity of that interaction—and the gap between laboratory promise and clinical durability—has only become clearer as long-term implant data have accumulated. What early work established as fundamental trade-offs, more recent evidence has reframed as solvable engineering problems, even as it has revealed new challenges that remain unsolved.

Established Platforms and Their Limitations

The Utah Array and Michigan-style silicon probes established the foundational vocabulary of intracortical recording: discrete electrode shanks penetrating cortical laminae to isolate single-unit activity with high spatiotemporal resolution [73, 3]. Yet these rigid silicon platforms carry an inherent contradiction. As [43] quantified, silicon and metallic electrode materials exhibit Young’s moduli that are orders of magnitude greater than brain gray matter—silicon at ~130–170 GPa versus cortical tissue at roughly 0.5–1 kPa [74]—which has mechanical compliance closer to gelatin than to any conventional engineering material. This stiffness mismatch is not a minor inconvenience; it drives a sustained neuroinflammatory cascade that progressively degrades signal quality. [7] identified glial scarring and micro-motion—the small but continuous relative displacement between a rigid implant and pulsating, moving neural tissue—as the primary mechanisms underlying signal instability in microelectrode arrays. The foreign body response unfolds in two phases: an acute inflammatory reaction in the days following implantation, followed by chronic glial encapsulation in the central nervous system and fibrotic encapsulation in the peripheral nervous system, ultimately electrically isolating the electrode from the neurons it was meant to record [13, 72]. Signal degradation timelines vary considerably across studies, with some reporting significant decline within months and others demonstrating multi-year functional recording—Hughes et al. [41], for instance, documented stable neural stimulation and recording performance in human somatosensory cortex over 1,500 days—a variance that likely reflects differences in implant geometry, surgical precision, and individual biological variability [1, 7, 42].

The Materials Turn: Flexible Substrates and Nanomaterials

Recognition that mechanical mismatch was a root cause of failure drove a sustained shift toward compliant materials [42, 73]. The orders-of-magnitude difference in elastic modulus between stiff silicon or tungsten probes and soft brain tissue (~1 kPa) generates chronic micromotion stress at the implant–tissue interface, accelerating neuronal loss and glial encapsulation [74, 73]. Flexible polymer probes—fabricated from polyimide, parylene-C, and related substrates—represent the most clinically advanced expression of this strategy [75]. Neuralink’s published system [76] demonstrated ultra-fine polymer threads just 4–6 μm thick, each carrying 32 electrodes, designed explicitly to minimize the immune footprint of the implant by approaching the mechanical and dimensional scale of neural tissue itself. Critically, this design acknowledged that flexibility creates an insertion problem: threads too compliant to penetrate cortex under their own rigidity require a precision robotic delivery system. The Neuralink surgical robot [76] addressed this by automating thread implantation at up to six insertions per minute while using real-time vascular imaging to avoid surface blood vessels—a capability that reflects how implantation mechanics have become inseparable from materials selection. Recent work on flexible polymer electrodes has further demonstrated stable, chronic visual percepts in animal models, suggesting that polymer substrates can maintain functional performance over behaviorally relevant timescales [77]. Nevertheless, [13] cautioned that flexible probes, while reducing chronic glial scarring, introduce distinct long-term reliability concerns including lead fracture, delamination at material interfaces, and electrochemical degradation of electrode coatings—failure modes that rigid arrays, paradoxically, handle more predictably [67].

Nanomaterials offer a complementary and in some respects more radical approach. Foundational work by [78] established that carbon nanotubes (CNTs), graphene, and semiconductor nanowires possess a combination of electrical, mechanical, and chemical properties that are uniquely suited to the neural interface context: their nanoscale dimensions are commensurate with cellular machinery, their electrical impedance is low enough to record small-amplitude signals with high fidelity, and their charge injection capacities exceed those of conventional metals. Organic and nanostructured electrode coatings more broadly have been shown to reduce impedance by one to two orders of magnitude relative to bare metal, directly improving signal-to-noise for single-unit recording [79]. CNTs in particular were shown to support robust neuronal growth and synaptic function while simultaneously reducing electrode impedance and inflammation [78]. Graphene electrodes have since demonstrated these properties in the context of micro-electrocorticography, where [44] reported chronic recordings sustained for up to 70 days, combined with optical transmittance above 90% that enables simultaneous electrophysiology and optogenetic or imaging techniques—a multimodal capability not accessible with opaque metal electrodes. Despite this promise, long-term chronic data for nanomaterial-based electrodes in human subjects are essentially nonexistent, and the specific interactions between nanomaterial surface chemistry and the chronic immune response remain poorly characterized [78, 13].

High-Density Systems and the Resolution Frontier

While flexible and nanomaterial approaches address biocompatibility, a parallel effort has targeted recording density. Neuropixels probes, developed originally for rodent neuroscience [42], represent a step change in the number of simultaneously sampled neurons. [80] reported the first successful intraoperative deployment of Neuropixels 1.0 probes in human cortex, isolating an average of 201±151 unit clusters per participant, with approximately 52% of units showing deflections across multiple recording channels—a spatial footprint that provides information about cortical column organization inaccessible to sparser arrays. This adaptation from animal to human use required addressing sterilization, probe geometry, and the constraints of intraoperative neurosurgical settings, and the recordings were necessarily acute rather than chronic. The scalability of such high-density silicon probes to long-term implants therefore remains an open question, one that intersects directly with the mechanical mismatch and foreign body response problems that have already constrained simpler silicon platforms [7, 43]. Foreign body reaction—characterized by glial encapsulation and progressive neuronal loss that accelerates once the electrode-to-neuron distance exceeds approximately 140–150 µm—is recognized as the primary limiting factor for chronic silicon probe performance [42, 72], and CMOS-based high-density architectures have yet to demonstrate stable multi-month recordings in the human brain.

Micro-ECoG: Reduced Invasiveness, Contested Resolution

Micro-electrocorticography (micro-ECoG) arrays occupy an intermediate position in the invasiveness spectrum, resting on the cortical surface rather than penetrating it and thereby avoiding the most severe phases of the foreign body response—particularly the chronic fibrotic encapsulation and neuronal loss associated with penetrating shanks [72]. [44] argued that micro-ECoG achieves spatial resolution approaching that of intracortical arrays—a claim supported by the spatial sampling density of modern micro-ECoG grids—while substantially reducing surgical risk and implant-related tissue damage. However, direct performance comparisons between micro-ECoG and intracortical arrays in matched BCI decoding tasks are limited [5], and the signal quality advantage of penetrating arrays for isolating single-unit spikes is not replicated by surface recordings, which primarily capture local field potentials and high-gamma activity [43]. The extent to which these signal types are functionally equivalent for BCI applications depends on the decoder architecture [6], but the absence of controlled comparative studies leaves the invasiveness-performance trade-off empirically underspecified.

Biocompatible Coatings, Failure Modes, and the Durability Problem

The most comprehensive recent synthesis of hardware failure modes, [13], identified that neural interface systems fail through multiple concurrent mechanisms: not only the biological response of the tissue, but also material-level processes including corrosion of metallic conductors, electrochemical degradation of electrode coatings under chronic stimulation, delamination of insulating layers, and packaging failures that allow biological fluids to reach electronics [75]. Anti-inflammatory coatings, drug-eluting surfaces, and zwitterionic polymer modifications have been explored as mitigation strategies [72, 79], but evidence of their efficacy beyond acute and subacute timeframes in humans remains thin. Conductive polymer coatings such as PEDOT, while capable of enhancing charge storage capacity by an order of magnitude over bare platinum and reducing electrochemical impedance substantially [79], are nonetheless prone to delamination in vivo and have demonstrated limited long-term stability in chronic implant studies. Similarly, hydrogel coatings address mechanical mismatch but can suffer from swelling-induced increases in electrode-tissue distance, redistributing rather than eliminating failure risk [79, 74]. This reflects a broader limitation acknowledged across the literature: standardized accelerated lifetime testing protocols for neural implants do not yet exist [13], making it difficult to predict in vivo longevity from bench measurements or to compare durability across platforms. The signal instability that [1] and [7] identified as a translational barrier—rooted in biological, mechanical, and electrochemical processes operating simultaneously—has not been eliminated by any current approach; it has at best been deferred or redistributed across different failure modes. Resolving this will require not only improved materials and surgical techniques but also the longitudinal human implant datasets that, as of the most recent literature [13, 80], remain largely unavailable for next-generation devices.

7. Clinical Translation, Neurorehabilitation, and Ethical-Regulatory Dimensions

The pathway from laboratory innovation to clinical deployment of brain–computer interfaces represents one of the most complex translational challenges in modern neuroscience and biomedical engineering. BCIs sit at the intersection of neurosurgical risk, adaptive computational systems, regulatory novelty, ethical controversy, deeply unequal global research capacity, and an underdeveloped health economic infrastructure. Understanding how technical progress converts into real-world benefit requires examining not only engineering milestones but also clinical trial frameworks, rehabilitation synergies, health economic barriers, usability dynamics, ethical dimensions, and structural inequities that shape who ultimately benefits from this technology and when.

From Laboratory to Clinic: A Translational Bottleneck

Early reviews of implantable BCI technology identified the brain’s reactive glial response to chronically implanted microelectrode arrays as a primary obstacle to durability, with signal degradation typically compromising device performance within approximately one year of implantation [81]. Silicon-based electrode arrays have been observed to function for roughly six to twelve months before substantial signal loss, with only occasional cases sustaining performance for up to three years [48]. Concurrent analyses emphasized that microelectrode arrays were inherently vulnerable to signal instability arising from both glial scarring and electrode micro-motion [7]—the latter driven by respiration, heartbeat, and cerebrospinal fluid pressure fluctuations that generate mechanical shearing between rigid electrodes and compliant neural tissue [13]. These biological constraints established an early tension between the resolution necessary for sophisticated motor decoding and the long-term reliability demanded by clinical applications [48].

By the early 2010s, first-in-human studies demonstrated that cortical signals could support increasingly complex motor tasks. A tetraplegic participant achieved progressive robotic arm control spanning three to seven degrees of freedom through iterative decoder training [49], and subsequent work showed that handwriting and speech could be decoded from cortical activity, enabling direct text communication [50]. Yet even as these demonstrations grew more sophisticated, the clinical population receiving implantable BCIs remained extraordinarily small. A 2025 systematic review of 112 studies published between 2000 and 2024 estimated that only approximately 80 individuals worldwide had ever received an implantable BCI system [8]—a figure that underscores how nascent clinical translation remains despite two decades of technical progress.

The Outcome Measure Disconnect

The same systematic review found that clinical outcome measures appeared in only 17.9% of implantable BCI publications, with the overwhelming majority of studies relying exclusively on engineering-focused metrics such as decoding accuracy, bits per second, or degrees of freedom achieved [8]. A complementary systematic review of outcome measures across 77 implantable BCI studies likewise found that accuracy was reported in 76% of publications while only 22.1% included any clinical outcome measure—and among those that did, 20 different instruments were employed with considerable heterogeneity, making cross-study comparison difficult and rendering existing measures unsuitable as generalized regulatory endpoints [12]. Regulatory agencies require evidence of clinical benefit expressed in terms meaningful to patients and clinicians—functional independence, quality of life, patient-reported outcomes—rather than signal classification performance, and this evidentiary misalignment constitutes a structural barrier to approval that technical refinements alone cannot resolve [8, 7]. The methodological weaknesses are not unique to BCIs: a consensus initiative examining spinal cord stimulation trials identified pervasive deficiencies across 34 randomized controlled trials, including inadequate blinding and insufficient handling of missing data, cautioning that bias in neuromodulation trials can be of comparable magnitude to genuine treatment effects [82].

That engineering metrics are insufficient proxies for real-world utility has been empirically confirmed. Kübler and colleagues [27] evaluated four BCI-controlled applications with nineteen severely motor-impaired end-users and found striking dissociations between technical performance and user-perceived value: an SMR-BCI application averaged only 60% accuracy yet workload ratings were broadly comparable to a P300 application exceeding 80%, and entertainment applications sustained engagement even under degraded accuracy. This result challenges the assumption that higher decoding performance straightforwardly translates into greater user benefit [27], and reinforces the concern that studies which do report clinical outcomes rarely employ validated usability instruments capable of capturing effectiveness, efficiency, and satisfaction. A systematic review of usability evaluation practices across BCI research found that while NASA Task Load Index and Visual Analogue Scale were the most commonly deployed subjective instruments, the overwhelming majority of studies relied on closed, experimenter-defined tasks conducted in laboratory settings, with only a minority evaluated in real-world daily-life environments—severely limiting ecological validity [39].

This gap is particularly significant given the regulatory complexity surrounding adaptive, learning-capable neural interfaces whose behavior changes post-implantation. Existing frameworks were not designed for systems that continuously update decoding parameters in response to neural plasticity, raising unresolved questions about how safety and efficacy should be demonstrated across a device’s operational lifespan [81, 50]. U.S.-based programs such as BrainGate have provided some evidence that implantable BCIs can achieve a safety profile comparable to other chronically implanted medical devices [38], offering a partial precedent, but comprehensive regulatory pathways specific to adaptive neural interfaces remain underdeveloped.

Neurorehabilitation Synergies and the Standalone Intervention Problem

While BCI clinical trials have largely evaluated devices as standalone interventions, a contrasting and arguably more neurobiologically informed framework emerged from neurorehabilitation research. A foundational synthesis argued that combining neurobiological interventions—including neurotrophic factors, monoamine pharmacotherapy, and cellular regeneration strategies—with neural interface training and task-specific rehabilitation produces synergistic recovery exceeding any individual treatment alone [83]. Evidence that epidural spinal stimulation combined with monoamine pharmacotherapy and goal-directed training enabled patients with complete chronic spinal cord injury to regain voluntary movement and autonomic functions supported this claim—outcomes that neither pharmacological nor electrophysiological intervention alone had achieved [83].

Earlier work established the neurobiological premise underpinning these approaches. Research on BMI-driven stroke neurorehabilitation demonstrated that BMI-trained participants outperformed sham-control groups in controlled trials while articulating a consequential distinction between assistive BCIs, which reroute motor intent around damaged pathways, and rehabilitative BCIs, which target cortical reorganization to restore intrinsic motor function [84]—a distinction with direct implications for health economic modeling, since the two modes generate fundamentally different cost-benefit profiles over a patient’s lifetime. A proof-of-principle study evaluating an EEG-based BCI for post-stroke arm rehabilitation in an inpatient setting found high patient motivation, compliance, and satisfaction, with performance correlating positively with self-reported motivation [85]—underscoring that psychological engagement modulates BCI efficacy beyond the purely neurophysiological. More recent multimodal initiatives have begun to operationalize these principles: the NeuroSuitUp clinical trial, for instance, integrates motor imagery, robotic interfaces, brain-machine interfaces, and augmented reality into a unified platform targeting chronic cervical spinal cord injury, explicitly hypothesizing that synergistic neuroplasticity across modalities will exceed what any single component achieves alone [86].

This multi-intervention paradigm has not been widely adopted in BCI clinical trial design. A systematic review of implantable BCI clinical trials spanning 112 studies and approximately 80 unique participants found that outcome reporting remains predominantly engineering-focused—with decoding and task accuracy metrics appearing in the majority of studies—while validated clinical outcome measures were used in only 17.9% of publications [8]. The absence of standardized, clinically meaningful outcomes has been identified by regulatory bodies including the FDA as a critical barrier to translation [8], and the field’s heavy geographic concentration and small enrolled populations further constrain the generalizability of existing evidence [8]. The tension between the neurorehabilitation literature’s advocacy for synergistic, multi-modal approaches and the BCI field’s predominantly device-centric trial frameworks represents a significant gap in translational strategy. Incorporating BCIs into comprehensive rehabilitation protocols alongside spinal stimulation, pharmacological priming, and task-specific training may be necessary to demonstrate the functional gains that regulatory bodies and healthcare systems require, yet the organizational complexity of such trials has impeded their execution.

Accessibility considerations further complicate the clinical pathway. Modular, at-home BCI platforms represent an important step toward reducing clinic-dependence [38], but widespread deployment requires reimbursement structures, caregiver training, and remote monitoring capabilities that remain largely undeveloped. The BackHome project [34] deployed a P300-based BCI across five homes in two countries, achieving mean task accuracy of 76%—lower than laboratory benchmarks but judged functionally meaningful by users—with adoption critically enabled by wireless dry-electrode EEG hardware that eliminated conductive gel preparation and substantially reduced caregiver burden. ALS patients independently created and publicly exhibited artwork through Brain Painting, illustrating how BCI adoption is inseparable from questions of selfhood, creative agency, and social participation that purely technical metrics cannot capture [34]. A parallel home-based neurofeedback study for central neuropathic pain [35] found that fifteen of twenty participants successfully learned independent system operation within four sessions, though 10–20% of home-recorded data was judged excessively noisy. These deployments confirm that ecological usability—encompassing preparation burden, caregiver requirements, and aesthetic acceptability—constitutes a translational barrier at least as consequential as signal quality or decoding accuracy.

Health Economic Barriers and Reimbursement Pathways

The translational challenges documented above also constitute the primary obstacles to establishing the health economic evidence base that payers and health technology assessment bodies require. No robust cost-effectiveness evidence exists for implantable BCIs across any clinical indication, and no harmonized reimbursement pathway governs these devices in any major health system. The low clinical outcome measure adoption rate documented by [8] means that existing BCI trial data cannot populate cost-utility models in a form recognizable to HTA agencies, which require quality-adjusted life year estimates anchored to validated clinical and patient-reported endpoints [12]. This evidentiary gap cannot be resolved by technical refinement—it is a structural misalignment between how the BCI field generates evidence and how health systems evaluate it.

Loeb’s critique of the field’s historical development argued that economic viability must be treated as a foundational design criterion co-equal with technical feasibility, rather than a downstream commercial consideration [14]. Historically, neural prosthetic development has prioritized technical demonstration, producing devices that achieve impressive laboratory performance but cannot be manufactured, reimbursed, or sustained within real-world health system budgets [14]. Deep brain stimulation offers both instructive precedents and sobering cautions: despite decades of clinical use, DBS continues to face structural weaknesses in its health economic evidence base. Methodological reviews of cost-effectiveness modelling for device-aided therapies in Parkinson’s disease highlight persistent heterogeneity in utility measurement approaches, limited long-term follow-up data, and wide variation in model assumptions that collectively undermine the comparability of economic estimates across studies [87, 33]. The Fourth Annual DBS Think Tank documented an international registry for Tourette syndrome encompassing only 149 cases from 16 institutions, with incomplete longitudinal outcomes and higher-than-anticipated device explantation rates [88], while the Eleventh Annual DBS Think Tank highlighted the continued need for governance frameworks capable of keeping pace with AI-guided autonomy and novel reimbursement questions [89]. The DBS trajectory demonstrates that reimbursement for implantable neurotechnologies is achievable through sustained registry infrastructure investment—investment the BCI field has not yet undertaken.

A further challenge concerns the inadequacy of standard cost-utility instruments for BCI populations. Conventional QALY tools were validated primarily in ambulatory populations and produce floor effects that obscure clinically meaningful functional gains in individuals with severe motor disability [87]. For a locked-in patient who regains the ability to communicate, the QALY increment captured by generic preference-based measures may be trivially small despite the intervention being transformative. This problem is compounded by documented tendencies for caregivers and physicians to systematically underestimate patient-reported quality of life in populations such as late-stage ALS [20], meaning proxy-reported utility values are likely biased downward. Multi-criteria decision frameworks offer a partial response: structured approaches such as the Hierarchical Decision Model have demonstrated that patient perspective—encompassing usability, engagement, and perceived benefit—can emerge as the dominant adoption criterion, ranking ahead of technical performance and financial considerations [90]. However, these approaches have not yet been applied specifically to implantable BCIs, and their integration with conventional HTA processes remains methodologically unresolved.

Usability, Adoption, and the Risk of Assistive Technology Abandonment

Assistive technology abandonment—the discontinuation of device use by individuals for whom it was prescribed—is predictable and preventable, driven by poor device-person fit, inadequate user involvement, unmet functional expectations, and insufficient post-adoption support. A systematic review of AT acceptance among users with motor disabilities found that dropout rates reach as high as 75%, largely attributable to insufficient user involvement during development [29]. These dynamics carry direct implications for BCIs, which impose substantially greater setup complexity and caregiver burden than most conventional assistive devices.

Shah and Robinson [91], synthesising 24 studies, established that medical device users constitute a deeply heterogeneous group whose involvement in development remains systematically concentrated in design and usability testing phases, while concept generation and post-deployment stages are comparatively neglected. This is directly relevant to BCIs, where post-deployment support infrastructure is minimal and where populations with progressive conditions such as ALS face evolving needs that initial design consultations cannot fully anticipate. A scoping review of 73 publications employing social research methods in BCI contexts [30] confirmed a consequential divergence: potential users—typically non-impaired research participants—prioritise usability features such as ease of use and accuracy, whereas actual users living with physical impairments articulate expectations centred on independence, social participation, and self-expression [30]—dimensions that standard engineering metrics do not capture.

Validated frameworks for assessing device-person fit, including the Matching Person and Technology model, the Quebec User Evaluation of Satisfaction with Assistive Technology, and tools operationalising the DIN 92419 ergonomic design principles [92], remain critically underutilised in neurotechnology research. Chaudhry et al. [93] further identify the Matching Person and Technology Model as a particularly valuable but underdeployed lens for implantable and neural technologies, noting that conventional acceptance frameworks such as TAM inadequately capture the ethical, affective, and physiological dimensions specific to body-integrated devices. Their systematic integration into BCI trial protocols would substantially improve the field’s capacity to identify adoption risks before they manifest as abandonment. Home deployment evidence reinforces why usability assessment must extend beyond laboratory settings: preparation burden emerges as a decisive factor in whether home users can realistically deploy a BCI system independently [34, 35]. Work on human-centred design for immersive virtual reality neurorehabilitation [94], which employed 192 hours of ethnographic observation alongside expert co-design, exemplifies the depth of participatory engagement that BCI development has not yet routinely achieved. Embedding validated usability instruments alongside clinical and health economic endpoints from the inception of trial design is not an optional refinement; it is a prerequisite for generating evidence that regulators, payers, clinicians, and patients require to evaluate whether a system will be adopted and sustained in practice.

Ethical Dimensions: Privacy, Autonomy, and Dual Use

Neural data—continuous recordings of cortical activity reflecting intentions, emotions, and cognitive states—represents a qualitatively novel category of personal information whose privacy implications are not addressed by existing health data frameworks [45, 38]. The potential for neural signals to be decoded or accessed by third parties raises questions about cognitive liberty and mental privacy that existing informed consent processes do not adequately address [95, 23]. Critically, advances in neural decoding now enable reconstruction of imagined speech, visual imagery, and affective states from raw cortical signals [23], substantially expanding the surface area of potential privacy violation beyond what conventional biomedical data governance anticipates. Long-term psychosocial impacts of chronic BCI use on personal identity, sense of agency, and mental health remain almost entirely unexplored—a gap that [30] identified as particularly consequential, given that quantitative attitude surveys rarely allow users to articulate how BCI use affects their sense of self or their relationships.

The informed consent challenge is especially acute for populations with progressive communication loss. Quality of life can be maintained even under severe motor impairment yet is systematically underestimated by caregivers and physicians [20]. Between 10 and 17 percent of ALS patients cannot use eye-gaze communication technology due to progressive oculomotor impairment, making implantable BCIs potentially the only remaining communication pathway [20]. Standard consent frameworks are poorly calibrated for this population: patients facing complete locked-in states may have diminishing capacity to re-evaluate their wishes over time, and proxy decision-making introduces risks of projection and underestimation of patient wellbeing [95, 20]. Ethical reviews of BCI trials have highlighted that surrogate consent in this context is particularly susceptible to conflation of caregiver distress with patient preference [95].

Survey evidence from 155 Italian healthcare professionals across rehabilitation and assistive robotics contexts revealed that respondents rated their university-level ethics training below threshold across all modules, while 71.61% identified the replacement of healthcare professionals by robotic systems as their foremost ethical concern [96]—illustrating how workforce anxieties and patient-care ethics are entangled in ways that may shape technology adoption. Military and dual-use dimensions add further complexity: systems developed for therapeutic purposes can be adapted for performance enhancement in healthy individuals, raising concerns about coercion in military contexts and the appropriate limits of human augmentation [81, 97]. The boundary between restoration of lost function and enhancement beyond baseline capacity is often technically ambiguous [97, 95], and governance frameworks specifically addressing BCI enhancement remain absent [38, 45].

Global Research Equity and the Bibliometric Landscape

Bibliometric analysis of 25,336 Scopus publications found that China surpassed the United States as the leading annual publisher of BCI research beginning in 2019, signalling an increasingly multipolar research landscape [45]. More troubling is the near-total absence of African research contributions: Africa accounts for less than one percent of global BCI publications despite housing over 80 million people with disabilities who could potentially benefit from such technologies [45]. The conditions producing the highest burden of neurological disability—stroke, spinal cord injury, and traumatic brain injury from conflict or road traffic accidents—are disproportionately prevalent in low- and middle-income countries [8]; indeed, neurological conditions represent the leading cause of ill health worldwide, often leaving patients decades dependent on care following injury [8]. Yet these populations are effectively excluded from both BCI research participation and the knowledge production that shapes future device development [45, 8]. This geographic concentration is further underscored by the finding that the entire global implantable BCI literature—spanning 112 studies conducted between 2000 and 2024—encompasses only approximately 80 implanted participants, the overwhelming majority based in the United States [8], leaving profound questions about whose needs and neural profiles the technology is actually designed to serve.

Synthesis and Outstanding Challenges

Taken together, the translational literature reveals a field in which technical ingenuity has consistently outpaced the clinical, regulatory, economic, and ethical infrastructure required to convert laboratory demonstrations into broadly accessible therapies. The persistence of engineering-dominated outcome metrics [8, 7], the mismatch between device-centric trial designs and evidence for rehabilitation synergies [83, 84], the absence of health economic evidence compatible with HTA requirements [14, 88], the systematic neglect of validated usability and adoption frameworks [30, 91, 27], and the stark geographic concentration of research capacity [45, 38] collectively define an agenda that is as much organizational and ethical as it is scientific. The near-absence of social research methods and structured end-user consultation in BCI studies [30] compounds these deficits, as does the relative scarcity of frameworks addressing the ethical dimensions of neurotechnology deployment at scale [95]. Progress toward clinical application will further require coordinated standardization efforts spanning device performance, regulatory pathways, and trial methodology [20, 98]. Closing these gaps will require deliberate collaboration between engineers, clinicians, health economists, regulatory scientists, ethicists, and the patient communities whose voices remain largely absent from the literature that purports to serve them [95, 25]—a challenge that the following discussion examines in the context of the field’s broader trajectory.

8. Discussion

The past three years have produced a genuine inflection point in brain-computer interface research. What was, until recently, a field defined largely by proof-of-concept demonstrations has begun producing results that demand engagement from clinicians, regulators, engineers, health economists, and ethicists simultaneously. This review reveals not a single linear advance but a convergent pressure across five interdependent dimensions — clinical performance, decoding algorithms, hardware longevity, non-invasive alternatives, and translational infrastructure — that are simultaneously maturing and straining against one another. A unifying thread that emerges across these dimensions is the recognition that BCI operation constitutes a coupled co-adaptive system, and that many of the field’s most persistent challenges — the offline-online performance gap, cross-session decoder instability, user training demands, and the functional necessity of sensory feedback — are manifestations of this single underlying dynamic rather than independent problems requiring independent solutions.

A Field Converging Toward Clinical Seriousness

Perhaps the most important shift since 2022 is the movement from demonstrating that BCIs can work to characterizing how well and for whom they work under real-world constraints. High-channel-count intracortical arrays and ECoG grids have now enabled communication at speeds and vocabularies that begin to approach functional utility rather than laboratory novelty [24, 23]. Simultaneously, decoder architectures built on transformer and recurrent neural network foundations have compressed the gap between neural signal quality and output fidelity [23, 6], making algorithmic sophistication a genuine force multiplier on hardware that has not itself changed dramatically. The practical implication is that clinical BCI performance is no longer primarily a hardware bottleneck; it is increasingly a systems integration problem, one in which signal acquisition, computational decoding, user training, and device maintenance must all be optimized together [8]. That reframing has consequences for how research programs should be structured and how regulatory review should be scoped.

Yet the foundation on which these claims rest remains fragile. Approximately 80 individuals worldwide have received implantable BCIs [12] — a number that makes every subgroup analysis provisional and every generalization premature. Long-duration stability data beyond three to five years are essentially absent for most platforms, with only isolated reports of sustained recording performance extending to 1,500 days or more [41]. The field has, so far, compared platforms against baseline disability rather than against each other; no rigorous head-to-head comparison of intracortical arrays, ECoG grids, and endovascular approaches exists under matched task conditions [12, 8]. These are not minor methodological lacunae. They mean that the clinical positioning of different invasive modalities — Utah arrays for high-bandwidth motor and speech decoding [18], ECoG for chronic stability [44], stentrodes for lower-risk deployment [20] — rests more on theoretical reasoning and early-phase case series than on comparative evidence.

The Non-Invasive Gap Remains Real but Is Narrowing

Non-invasive BCI research has accelerated substantially, driven by consumer-grade EEG hardware improvements [99], hybrid fNIRS-EEG architectures [100, 39], and the same decoder advances that benefit invasive systems. Imagined speech decoding and passive cognitive state monitoring represent genuinely new capabilities relative to three years ago [64, 23, 15]. However, a structural limitation persists: the overwhelming majority of non-invasive BCI studies are conducted in controlled laboratory environments, and evidence for at-home, longitudinal, unsupported use remains sparse [35, 34]. For populations most likely to benefit — individuals with progressive neurological conditions who cannot attend regular clinical visits — this gap is not merely academic [25]. Standardized benchmarks across non-invasive language and speech BCI studies are also absent [23, 16], making it nearly impossible to assess whether improvements in one laboratory’s paradigm represent generalizable progress or optimized performance on a specific protocol. Closing this gap requires infrastructure investment — shared datasets, agreed evaluation metrics, and regulatory guidance for software-only BCI products — as much as it requires technical innovation.

Decoder Generalization as a Fundamentally Co-Adaptive Problem

Cross-session and cross-individual decoder adaptation emerges from this review as the single most consequential unsolved technical problem, and the co-adaptive framework developed in Section 5 clarifies why this challenge has proven so resistant to purely algorithmic solutions. Neural signals drift across days and weeks due to electrode displacement, impedance changes, and genuine neurophysiological non-stationarity [16, 70]; current decoders often require recalibration that is clinically impractical at scale. The conventional framing treats this drift as a nuisance variable to be corrected — a form of domain shift amenable to transfer learning or adaptive filtering. But the co-adaptive perspective reveals that signal drift during closed-loop operation is not merely electrode degradation or biological noise; it is partly the signature of a user whose cortical representations are actively reorganizing in response to the decoder’s behavior [1, 54]. Attempting to correct for this “drift” without modeling the user’s learning trajectory risks chasing a moving target — adapting the decoder to neural patterns that are themselves shifting in response to the adaptation. Work on neural manifold alignment offers one principled approach: by stabilizing decoders within a low-dimensional subspace of neural activity that remains consistent across sessions, some methods have achieved multi-week decoder stability without full recalibration [55], though generalization to the broader population remains limited.

The Cybathlon longitudinal data provide the most compelling empirical illustration of this dynamic: when mutual adaptation was explicitly managed through coordinated updates to the decoder, the application interface, and the user’s training regimen, command accuracy improved from below 54% to above 93% over several months, with corresponding neurophysiological evidence of increasingly discriminable sensorimotor rhythm modulation [54]. This result demonstrates that the co-adaptive problem is solvable in principle — but only when both sides of the adaptive loop are addressed simultaneously. Transfer learning approaches show early promise but have not been validated across the population heterogeneity and device variation that real deployment would involve [15, 16]. The 15–30% BCI illiteracy rate [10] adds a further complication: if a substantial fraction of users cannot effectively participate in co-adaptive learning due to neurobiological or psychological constraints, then no amount of decoder sophistication alone will achieve universal coverage. Strategies for addressing this population — including neurostimulation-based cortical priming [69, 65], alternative signal paradigms, and personalized training protocols informed by individual neurobiological predictors [10] — deserve substantially greater investment. Notably, BCI user learning itself appears to involve active cortical reorganization: longitudinal studies demonstrate that effective BCI operation recruits new neural activity patterns over weeks of training [101, 64], further coupling the decoder adaptation problem to the trajectory of user plasticity.

Until decoders can adapt robustly to the coupled dynamics of user learning and signal drift without burdening users or clinicians, the operational lifespan of any BCI system will be constrained by maintenance demands that are incompatible with independent living. This problem sits at the intersection of computational neuroscience, control theory, and practical engineering, and the co-adaptive framework suggests it will not be solved by treating the decoder as an isolated optimization target. Research programs that jointly model user plasticity and decoder adaptation — and that explicitly measure the bidirectional dynamics of co-learning rather than reporting only decoder-side metrics — represent the most promising path forward.

Economic Viability as a Binding Constraint on Deployment

This review identifies a dimension of the translational challenge that the BCI literature has historically underweighted: the absence of health economic evidence and reimbursement infrastructure as a binding constraint on deployment that is at least as consequential as unresolved engineering problems. The approximately 80 individuals worldwide who have received implantable BCIs represent the limits not only of clinical evidence but of economic evidence [8, 12]. No cost-effectiveness analysis grounded in validated clinical endpoints exists for any implantable BCI system, and the 17.9% clinical outcome measure adoption rate across the published literature [8] means that existing trial data cannot populate the QALY-based models that health technology assessment bodies require. This is not a gap that further decoding improvements or biocompatibility advances can close — it reflects a structural misalignment between how the field generates evidence and how health systems make coverage decisions. Early health technology assessment, which embeds economic endpoints prospectively in development pipelines, has been identified as a mechanism for closing precisely this kind of misalignment [9], yet it remains largely absent from BCI research programs.

The deep brain stimulation experience provides both encouragement and caution. DBS achieved reimbursement through iterative evidence accumulation over decades, beginning with narrow indications and expanding as registry-based real-world evidence accumulated [88, 89]. Yet even after thousands of implantations, DBS continues to face reimbursement friction attributable to incomplete longitudinal data, inconsistent outcome measurement, and the difficulty of evaluating adaptive devices within frameworks designed for fixed-parameter technologies [88]. The methodological challenges in modelling cost-effectiveness for neurological device therapies — including how to handle utility decrements, caregiver burden, and non-linear disease trajectories — remain incompletely resolved even in the relatively mature DBS literature [87]. For BCIs, which are far earlier in their evidence trajectory and serve far smaller patient populations, the timeline from first implantation to sustainable reimbursement could be considerably longer unless the field adopts a fundamentally different approach to evidence generation — one in which health economic endpoints are embedded prospectively in trial design rather than retrofitted after technical milestones have been achieved.

The measurement problem is not solely one of evidence quantity. Standard cost-utility instruments are poorly calibrated for the severe disability populations that BCIs serve, frequently producing floor effects that obscure transformative functional gains [20]. The documented tendency for caregivers and clinicians to underestimate patient-reported quality of life in late-stage ALS compounds this problem [20], meaning that proxy-reported utility values — often the only values available for locked-in patients — may systematically understate the benefit of restoring communicative autonomy. Until condition-specific preference-based instruments are developed and validated for BCI-relevant populations, economic evaluations will remain vulnerable to the charge that they undervalue the very outcomes that matter most to patients.

The neural prosthetics literature’s argument that economic viability must be treated as a constitutive design criterion rather than a derivative commercial consideration [14] carries direct implications for how BCI research programs should be structured. Development pipelines that optimize for technical performance metrics without simultaneously generating the clinical and economic evidence that payers require are, in effect, building devices that may work brilliantly in the laboratory but cannot be reimbursed in any health system. The DBS field’s belated recognition of this dynamic — manifested in the registry fragmentation and evidence gaps that Think Tank proceedings continue to document years after initial clinical adoption [88, 89] — should serve as an explicit warning rather than a parallel that BCIs repeat.

Usability and Adoption: The Assistive Technology Abandonment Problem

Technical capability, clinical efficacy, and even economic viability do not, individually or collectively, guarantee that a BCI system will be adopted and sustained by the individuals it is designed to serve. This review’s engagement with the broader assistive technology literature reveals a pattern that should temper optimistic deployment projections: assistive devices developed without sustained end-user participation face predictable abandonment, and the factors driving discontinuation — poor device-person fit, setup complexity, caregiver burden, aesthetic concerns, and insufficient post-adoption support — are largely independent of technical performance [91, 30]. Studies of assistive technology acceptance among users with motor disabilities confirm that psychosocial factors, including perceived social stigma and impact on self-image, frequently outweigh functional performance in determining sustained use [29]. The BCI field is not exempt from these dynamics; if anything, the complexity of neural interface systems amplifies every adoption barrier that simpler assistive technologies encounter.

The empirical evidence from BCI-specific usability research is instructive. When BCI applications were evaluated using ISO 9241-210 user-centred design frameworks rather than engineering metrics alone, striking dissociations emerged between technical performance and user-perceived value: applications with substantially lower classification accuracy elicited comparable workload and satisfaction ratings to higher-performing alternatives, while the experiential dimensions that users valued most — creative self-expression, social participation, and a sense of agency — were invisible to standard performance reporting [27]. Home deployment studies revealed that the preparation burden of electrode application, the aesthetic intrusiveness of visible hardware, and the requirement for caregiver assistance constituted decisive adoption barriers that laboratory evaluations could not detect [34, 35]. A scoping review of social research methods in BCI contexts [30] further documented a consequential divergence in the priorities of different user groups: non-impaired research participants — who constitute the majority of BCI study samples — prioritise ease of use and accuracy, whereas individuals living with physical impairments articulate expectations centred on independence, social participation, and self-expression. This divergence implies that acceptance models and design priorities derived from healthy volunteer studies may systematically misrepresent the adoption dynamics of the clinical populations for whom BCIs carry the greatest stake.

These findings converge on a single implication: the approximately 80 individuals who have received implantable BCIs worldwide — a figure drawn from a systematic registry of 112 clinical studies spanning 2000–2024, which identified only 57 unique participants in the published literature and a further 23 through unpublished sources [8] — represent not only the limits of clinical and economic evidence but also the limits of usability evidence. Deployment beyond highly motivated research participants into broader clinical populations will therefore encounter adoption dynamics that the current evidence base is structurally unable to predict. The same systematic review found that validated clinical outcome measures appeared in fewer than one in five iBCI publications, with engineering metrics dominating reporting across nearly three-quarters of studies [8] — a pattern that mirrors the broader methodological skew documented by the BCI social research literature [30]. That scoping review documented a further concern: BCI social research remains dominated by quantitative feasibility studies, with qualitative accounts of how neural interfaces affect identity, selfhood, and lived experience largely absent from the literature. This methodological gap means that the field lacks the evidence base needed to anticipate how chronic BCI use will be experienced by users over months and years — precisely the timeframe over which assistive technology abandonment typically occurs.

Validated adoption and usability frameworks exist and have been successfully deployed in adjacent domains, yet their systematic integration into BCI development remains the exception rather than the norm. The Matching Person and Technology model, the Quebec User Evaluation of Satisfaction with Assistive Technology (QUEST), and the DIN 92419 ergonomic assessment standard operationalised through dedicated measurement instruments [92] offer concrete, field-tested tools for identifying adoption risks before they manifest as discontinuation [29]. Human-centred design methodologies employing sustained ethnographic engagement — such as the 192-hour observational co-design study conducted in immersive virtual reality neurorehabilitation [94] — demonstrate the depth of participatory inquiry required to surface the contextual constraints that quantitative performance evaluations routinely miss. The failure to incorporate these approaches is not attributable to their unavailability; it reflects a disciplinary orientation in which engineering metrics have historically been treated as sufficient proxies for real-world utility. The evidence reviewed here demonstrates conclusively that they are not.

Implications for Regulation and Equity

The regulatory pathway for BCIs has clarified meaningfully, with iterative approval frameworks and breakthrough device designations reducing some uncertainty [8, 98]. What has not kept pace is the ethical, economic, and equity infrastructure. BCI access remains concentrated in high-income academic medical centers, with cost, geographic availability, and institutional expertise acting as compounding barriers to broader uptake [38, 11, 45]. Informed consent frameworks have not fully accommodated the possibility of personality or agency changes from chronic neural modulation [95, 102], nor the specific challenges of obtaining informed consent from individuals with progressive communication loss [20]. Long-term device explantation responsibilities remain underspecified — a gap that becomes clinically salient as implanted systems age, degrade, or are discontinued by manufacturers [13, 17]. Adaptive reimbursement mechanisms — including conditional coverage with evidence development and risk-sharing agreements — represent viable pathways for enabling access while managing the uncertainty inherent in early-stage neurotechnologies [9, 103], but their implementation requires sustained engagement between manufacturers and payers that has not yet become routine in the BCI sector [8]. Early health technology assessment, increasingly used in adjacent device markets, offers one model for structuring this engagement prospectively [9]. As deployment scales — and the trajectory of current industry programs suggests it will — these gaps will become acute rather than theoretical.

Future Directions

The most productive near-term research investments appear to be: longitudinal registry studies capable of generating multi-year stability data across platforms [41, 12]; the development and validation of BCI-specific clinical outcome measures and condition-specific preference-based instruments suitable for QALY estimation in severely disabled populations [20, 8]; prospective embedding of health economic endpoints — including validated functional outcomes, patient-reported measures, caregiver burden, and long-term cost tracking — in all clinical BCI trials from inception [14]; the systematic integration of validated usability and adoption instruments — including the Matching Person and Technology model, the Quebec User Evaluation of Satisfaction with Assistive Technology, and ISO 9241-210-aligned evaluation protocols — into all clinical BCI trials from inception [27, 92]; participatory co-design methodologies that position end-users, including those with severe motor disabilities, as co-creators of BCI systems rather than passive evaluators of pre-determined designs, drawing on established human-centred design frameworks that emphasise sustained engagement across concept generation, development, deployment, and post-deployment phases [91, 94, 30]; standardized benchmarking infrastructure for non-invasive BCIs [66, 98]; co-adaptive decoder architectures that explicitly model user learning dynamics alongside algorithmic adaptation, informed by the dual-loop framework that distinguishes brain-state and task-level feedback channels [62]; federated learning approaches to decoder adaptation that preserve privacy while enabling cross-site generalization [45, 23]; systematic investigation of neurostimulation-based priming and training interventions to expand the population of effective BCI users [69, 65, 10]; flexible nanomaterial electrode validation in chronic human implants [78, 72, 42]; and early engagement with health technology assessment bodies and payers to establish adaptive reimbursement pathways capable of accommodating iterative device evolution [89, 88]. More broadly, the field would benefit from explicitly comparative trial designs that can answer the question clinicians, patients, and payers most need answered — not whether BCIs work, but which approach works best for a specific person with a specific condition at a specific point in disease progression, and at what cost relative to alternative interventions [12, 25]. That question is now answerable in principle. Making it answerable in practice is the defining challenge of the next phase of BCI research.

9. Conclusions

This systematic review of 51 papers, organized across five thematic domains, synthesizes the current state of brain-computer interface technology and charts the landscape from laboratory demonstration to clinical translation. The conclusions drawn from this body of evidence are as follows.

What BCIs Can Currently Achieve

Invasive BCIs, particularly those employing intracortical electrode arrays, have demonstrated clinically meaningful restoration of motor and communication function in paralyzed individuals, enabling cursor control [3], robotic limb operation [104], and speech decoding at rates approaching natural communication. For instance, participants using 96-channel intracortical arrays in the BrainGate2 trial achieved typing rates of up to 39.2 correct characters per minute — representing a 3–4× improvement over previously reported BCI systems — with information throughput reaching 4.2 bits per second [18]. Non-invasive systems, while more limited in bandwidth and spatial resolution, have achieved reliable control of assistive devices and established proof-of-concept for rehabilitation applications [99, 60]. The evidence confirms that BCIs are no longer experimental curiosities but constitute a legitimate therapeutic modality for select patient populations [12].

Technical Barriers That Remain

Signal resolution, device longevity, and data bandwidth represent the primary engineering constraints limiting broader deployment. Chronic implants face well-documented foreign body responses that degrade electrode performance over months to years [72, 13]: following implantation, an acute inflammatory phase — driven by macrophage recruitment and reactive oxygen species — transitions into a chronic fibrotic encapsulation that physically attenuates recorded signals and dissipates stimulation currents [72]. Mechanical mismatch between stiff electrode materials and soft neural tissue exacerbates this response, making long-term signal fidelity a persistent unsolved problem [73]. Wireless transmission of high-channel-count neural data remains separately constrained by power budgets and regulatory exposure limits, and scaling to the electrode densities demonstrated in research platforms [76, 105] without a commensurate increase in wireless throughput has yet to be achieved. Non-invasive approaches sacrifice the spatial and temporal resolution that invasive methods achieve, creating a persistent trade-off that no current platform has fully resolved. A cross-cutting challenge that connects these hardware-level constraints to system-level performance is the co-adaptive nature of BCI operation: because the user’s neural representations and the decoder constitute coupled adaptive agents, long-term system stability depends not only on electrode durability and signal fidelity but also on the capacity of the decoder to accommodate the user’s ongoing neural plasticity [55, 70] — a coupled optimization problem that current systems manage poorly [54]. Until biocompatible, stable, high-bandwidth interfaces are reliably demonstrated across multi-year timescales, and until decoder architectures can robustly manage co-adaptive dynamics, clinical scalability will remain limited.

Positioning of Different Approaches

Utah arrays remain the most extensively validated intracortical platform for high-resolution motor decoding [5, 3], but are constrained by longevity concerns stemming from foreign body responses and electrode impedance degradation over time [67, 41]. Neuralink and similar next-generation systems offer substantially higher electrode counts and fully integrated wireless capability [76], positioning them for more ambitious applications, though long-term clinical data remain sparse. Electrocorticography occupies an important middle ground, offering greater signal stability and spatial coverage than intracortical arrays with lower invasiveness [44, 56], making it well suited to epilepsy monitoring and emerging communication applications [24]. Non-invasive modalities, including EEG and fNIRS-based systems [100, 21], are best positioned for rehabilitation, neurofeedback, and consumer applications where surgical risk is unacceptable and moderate performance is sufficient [25, 15].

Regulatory and Surgical Pathways

Regulatory frameworks, particularly those governing first-in-human trials and long-term implant approval, represent a significant and often underappreciated bottleneck. Notably, no implantable BCI has yet received regulatory approval or entered standard clinical practice [8], despite the global recipient count numbering only around 80 individuals across all published and unpublished trials [8]. Surgical risk profiles differ substantially across modalities — from penetrating microelectrode arrays to endovascular approaches — and must be matched carefully to patient benefit calculations, informed by systematic analyses of hardware complication rates across established intracranial technologies [17]. Expanded deployment will require not only technical maturation but also adaptive regulatory pathways that can accommodate iterative device updates without demanding full re-approval cycles, including standardized outcome measures that regulatory bodies such as the FDA can act upon [8].

Economic Viability and Reimbursement

The absence of robust health economic evidence constitutes a structural barrier to BCI deployment that is independent of, and in its current form at least as constraining as, unresolved engineering challenges. No cost-effectiveness analysis grounded in validated clinical endpoints exists for any implantable BCI [12], and standard QALY instruments are poorly calibrated for the severe disability populations these devices serve — a problem compounded by well-documented “response shift” effects, whereby patients with severe motor impairment report subjectively positive quality of life that generic health-utility scales systematically fail to capture [20]. A 2024 systematic review found that only 22.1% of implantable BCI studies reported any clinical outcome measure at all, with those studies employing 20 different instruments — a heterogeneity that renders cross-study economic comparison effectively impossible [12]. Until prospective trials systematically embed health economic endpoints — including validated functional outcomes, patient-reported measures, caregiver burden, and long-term cost tracking — and until manufacturers engage payers through adaptive reimbursement mechanisms, the economic case for BCI coverage will remain unanswerable, and access will remain confined to the academic medical centers able to absorb the cost of uncompensated innovation.

Usability and Adoption

Technical performance and economic viability are necessary but not sufficient conditions for BCI impact at scale. The assistive technology literature documents abandonment rates that frequently exceed fifty percent [29], driven not by device failure but by inadequate fit between device design and the lived contexts of users. BCIs are not immune to this dynamic. Systems that perform impressively in laboratory conditions have repeatedly failed to achieve comparable performance in real-world settings [27, 34] — a gap attributable in part to factors such as variable user mental states, environmental noise, and the absence of trained support personnel that controlled trials systematically exclude [71]. The absence of systematic usability evaluation — using validated instruments such as the Matching Person and Technology model, the Quebec User Evaluation of Satisfaction with Assistive Technology, and ISO 9241-210-aligned protocols — from the majority of clinical BCI trials reflects a structural gap in how the field measures success [27, 39]. Participatory co-design methodologies that position end-users as co-creators rather than passive evaluators [30, 91] offer a path toward devices whose real-world adoption rates match their technical promise. Integrating these frameworks into BCI development and trial design from inception, rather than as post-hoc additions, is a condition for realizing the population-level benefits the technology can in principle deliver.

Synthesis

The evidence reviewed across these five domains supports a single integrated claim: brain-computer interfaces have crossed from experimental demonstration into legitimate clinical translation, but the conditions for broad, equitable, and sustained deployment remain only partially in place. A systematic review of the global iBCI literature identified only approximately 80 implanted participants across 112 published studies — a figure that underscores how nascent real-world deployment remains despite rapid research growth, with nearly half of all publications appearing since 2020 [8]. Outcome measure heterogeneity across those studies further complicates cross-trial comparison and the accumulation of evidence needed to support regulatory and reimbursement decisions [12]. Ethical and clinical analyses further confirm that equitable access, valid informed consent, and post-trial device continuity remain unresolved prerequisites for responsible scale-up [20, 45]. Usability and human-centred design represent an equally underaddressed dimension: user-centred evaluation frameworks have been proposed but remain inconsistently applied across the field [27], and the integration of social research methods into BCI studies — essential for capturing lived experience — is still limited [30]. Technical maturation, adaptive regulation, credible health economic evidence, and rigorous usability integration are not sequential milestones but parallel requirements [7] — progress on any one dimension is insufficient if the others lag. The field stands at a genuine inflection point, one at which coordinated investment across engineering, clinical science, regulatory science, health economics, and human-centred design will determine whether BCIs fulfil their transformative potential for the populations who need them most, or remain a technology whose benefits are real but whose reach is narrow [38].

---

References

[1] Lebedev, M., Nicolelis, M. (2017). Brain-Machine Interfaces: From Basic Science to Neuroprostheses and Neurorehabilitation. Physiological Reviews. [2] van Gerven, M., Farquhar, J., Schaefer, R., et al. (2009). The brain-computer interface cycle. Journal of Neural Engineering. [3] Hochberg, L., Serruya, M., Friehs, G., et al. (2006). Neuronal ensemble control of prosthetic devices by a human with tetraplegia. Nature. [4] Santhanam, G., Ryu, S., Yu, B., et al. (2006). A high-performance brain-computer interface. Nature. [5] Tam, W., Wu, T., Zhao, Q., et al. (2020). Human motor decoding from neural signals: a review. BMC Biomedical Engineering. [6] Ahmadi, N., Constandinou, T., Bouganis, C. (2021). Robust and accurate decoding of hand kinematics from entire spiking activity using deep learning. Journal of Neural Engineering. [7] Murphy, M., Guggenmos, D., Bundy, D., et al. (2016). Current Challenges Facing the Translation of Brain Computer Interfaces from Preclinical Trials to Use in Human Patients. Frontiers in Cellular Neuroscience. [8] Dohle, E., Swanson, E., Jovanović, L., et al. (2025). Toward the Clinical Translation of Implantable Brain-Computer Interfaces for Motor Impairment: Research Trends and Outcome Measures. Advanced Science. [9] IJzerman, M., Koffijberg, H., Fenwick, E., et al. (2017). Emerging Use of Early Health Technology Assessment in Medical Product Development: A Scoping Review of the Literature. PharmacoEconomics. [10] Saha, S., Mamun, K., Ahmed, K., et al. (2021). Progress in Brain Computer Interface: Challenges and Opportunities. Frontiers in Systems Neuroscience. [11] Cruz, M., Jamal, S., Sethuraman, S. (2025). A Comprehensive Survey of Brain–Computer Interface Technology in Health care: Research Perspectives. Journal of Medical Signals & Sensors. [12] Dohle, E., Swanson, E., Yusuf, S., et al. (2024). Outcome measures in implantable brain-computer interface research: a systematic review. [13] Dalrymple, A., Jones, S., Fallon, J., et al. (2025). Overcoming failure: improving acceptance and success of implanted neural interfaces. Bioelectronic Medicine. [14] Loeb, G. (2018). Neural Prosthetics: A Review of Empirical vs. Systems Engineering Strategies. Applied Bionics and Biomechanics. [15] Edelman, B., Zhang, S., Schalk, G., et al. (2025). Non-Invasive Brain-Computer Interfaces: State of the Art and Trends. IEEE Reviews in Biomedical Engineering. [16] Rashid, M., Sulaiman, N., Majeed, A., et al. (2020). Current Status, Challenges, and Possible Solutions of EEG-Based Brain-Computer Interface: A Comprehensive Review. Frontiers in Neurorobotics. [17] Bullard, A., Hutchison, B., Lee, J., et al. (2020). Estimating Risk for Future Intracranial, Fully Implanted, Modular Neuroprosthetic Systems: A Systematic Review of Hardware Complications in Clinical Deep Brain Stimulation and Experimental Human Intracortical Arrays. Neuromodulation Technology at the Neural Interface. [18] Pandarinath, C., Nuyujukian, P., Blabe, C., et al. (2017). High performance communication by people with paralysis using an intracortical brain-computer interface. eLife. [19] Ajiboye, A., Willett, F., Young, D., et al. (2017). Restoration of reaching and grasping movements through brain-controlled muscle stimulation in a person with tetraplegia: a proof-of-concept demonstration. The Lancet. [20] Vansteensel, M., Klein, E., van Thiel, G., et al. (2022). Towards clinical application of implantable brain–computer interfaces for people with late‑stage ALS: medical and ethical considerations. Journal of Neurology. [21] Värbu, K., Muhammad, N., Muhammad, Y. (2022). Past, Present, and Future of EEG-Based BCI Applications. Sensors. [22] Padfield, N., Zabalza, J., Zhao, H., et al. (2019). EEG-Based Brain-Computer Interfaces Using Motor-Imagery: Techniques and Challenges. Sensors. [23] Qiu, Y., Liu, H., Zhao, M. (2025). A Review of Brain–Computer Interface-Based Language Decoding: From Signal Interpretation to Intelligent Communication. Applied Sciences. [24] Luo, S., Angrick, M., Coogan, C., et al. (2023). Stable Decoding from a Speech BCI Enables Control for an Individual with ALS without Recalibration for 3 Months. Advanced Science. [25] Simon, C., Bolton, D., Kennedy, N., et al. (2021). Challenges and Opportunities for the Future of Brain-Computer Interface in Neurorehabilitation. Frontiers in Neuroscience. [26] Biasiucci, A., Leeb, R., Iturrate, I., et al. (2018). Brain-actuated functional electrical stimulation elicits lasting arm motor recovery after stroke. Nature Communications. [27] Kübler, A., Holz, E., Riccio, A., et al. (2014). The User-Centered Design as Novel Perspective for Evaluating the Usability of BCI-Controlled Applications. PLoS ONE. [28] Millán, J. (2010). Combining brain-computer interfaces and assistive technologies: state-of-the-art and challenges. Frontiers in Neuroscience. [29] Ventura, S., Ottoboni, G., Pappadà, A., et al. (2023). Acceptance of Assistive Technology by Users with Motor Disabilities Due to Spinal Cord or Acquired Brain Injuries: A Systematic Review. Journal of Clinical Medicine. [30] Kögel, J., Schmid, J., Jox, R., et al. (2019). Using brain-computer interfaces: a scoping review of studies employing social research methods. BMC Medical Ethics. [31] Müller-Putz, G., Leeb, R., Tangermann, M., et al. (2015). Towards Noninvasive Hybrid Brain-Computer Interfaces: Framework, Practice, Clinical Application, and Beyond. Proceedings of the IEEE. [32] Eggington, S., Valldeoriola, F., Chaudhuri, К., et al. (2013). The cost-effectiveness of deep brain stimulation in combination with best medical therapy, versus best medical therapy alone, in advanced Parkinson’s disease. Journal of Neurology. [33] Śmiłowska, K., van Wamelen, D., Pietrzykowski, T., et al. (2020). Cost-Effectiveness of Device-Aided Therapies in Parkinson’s Disease: A Structured Review. Journal of Parkinson s Disease. [34] Miralles, F., Vargiu, E., Dauwalder, S., et al. (2015). Brain Computer Interface on Track to Home. The Scientific World JOURNAL. [35] Al-Taleb, M., Purcell, M., Fraser, M., et al. (2019). Home used, patient self-managed, brain-computer interface for the management of central neuropathic pain post spinal cord injury: usability study. Journal of NeuroEngineering and Rehabilitation. [36] Saibene, A., Caglioni, M., Corchs, S., et al. (2022). EEG-based BCIs on motor imagery paradigm using wearable technologies: A systematic review. Preprints.org. [37] Saibene, A., Caglioni, M., Corchs, S., et al. (2023). EEG-Based BCIs on Motor Imagery Paradigm Using Wearable Technologies: A Systematic Review. Sensors. [38] Sonko, S., Fabuyide, A., Ibekwe, K., et al. (2024). Neural interfaces and human-computer interaction: A U.S. review: Delving into the developments, ethical considerations, and future prospects of brain-computer interfaces. International Journal of Science and Research Archive. [39] Choi, I., Rhiu, I., Lee, Y., et al. (2017). A systematic review of hybrid brain-computer interfaces: Taxonomy and usability perspectives. PLoS ONE. [40] Birbaumer, N., Cohen, L. (2007). Brain–computer interfaces: communication and restoration of movement in paralysis. The Journal of Physiology. [41] Hughes, C., Flesher, S., Weiss, J., et al. (2021). Neural stimulation and recording performance in human somatosensory cortex over 1500 days. [42] Perna, A., Angotzi, G., Berdondini, L., et al. (2023). Advancing the interfacing performances of chronically implantable neural probes in the era of CMOS neuroelectronics. Frontiers in Neuroscience. [43] Uddin, E., Gao, P. (2021). Signal Generation, Acquisition, and Processing in Brain Machine Interfaces: A Unified Review. Frontiers in Neuroscience. [44] Shokoueinejad, M., Park, D., Jung, Y., et al. (2019). Progress in the Field of Micro-Electrocorticography. Micromachines. [45] Maiseli, B., Abdalla, A., Massawe, L., et al. (2023). Brain-computer interface: trend, challenges, and threats. Brain Informatics. [46] Mervyn Jun Rui Lim, Yu Tung Lo, Yong Yi Tan, et al. (2025). The state-of-the-art of invasive brain-computer interfaces in humans: a systematic review and individual patient meta-analysis. Journal of Neural Engineering. [47] Carmena, J., Lebedev, M., Crist, R., et al. (2003). Learning to Control a Brain–Machine Interface for Reaching and Grasping by Primates. PLoS Biology. [48] Ryu, S., Shenoy, K. (2009). Human cortical prostheses: lost in translation?. Neurosurgical FOCUS. [49] Bensmaı̈a, S., Miller, L. (2014). Restoring sensorimotor function through intracortical interfaces: progress and looming challenges. Nature reviews. Neuroscience. [50] Adrien Rapeaux, Timothy G. Constandinou (2021). Implantable brain machine interfaces: first-in-human studies, technology challenges and trends. Current Opinion in Biotechnology. [52] Benabid, A., Costecalde, T., Eliseyev, A., et al. (2019). An exoskeleton controlled by an epidural wireless brain-machine interface in a tetraplegic patient: a proof-of-concept demonstration. The Lancet Neurology. [53] Samejima, S., Khorasani, A., Ranganathan, V., et al. (2021). Brain-Computer-Spinal Interface Restores Upper Limb Function After Spinal Cord Injury. IEEE Transactions on Neural Systems and Rehabilitation Engineering. [54] Perdikis, S., Tonin, L., Saeedi, S., et al. (2018). The Cybathlon BCI race: Successful longitudinal mutual learning with two tetraplegic users. PLoS Biology. [55] Degenhart, A., Bishop, W., Oby, E., et al. (2020). Stabilization of a brain-computer interface via the alignment of low-dimensional spaces of neural activity. Nature Biomedical Engineering. [56] Caldwell, D., Ojemann, J., Rao, R. (2019). Direct Electrical Stimulation in Electrocorticographic Brain–Computer Interfaces: Enabling Technologies for Input to Cortex. Frontiers in Neuroscience. [57] Zhao, Z., Nie, C., Jiang, C., et al. (2023). Modulating Brain Activity with Invasive Brain-Computer Interface: A Narrative Review. Brain Sciences. [58] Caiado, F., Уколов, А. (2025). The history, current state and future possibilities of the non-invasive brain computer interfaces. Medicine in Novel Technology and Devices. [59] Naser, M., Bhattacharya, S. (2023). Towards Practical BCI-Driven Wheelchairs: A Systematic Review Study. IEEE Transactions on Neural Systems and Rehabilitation Engineering. [60] Lazarou, I., Nikolopoulos, S., Petrantonakis, P., et al. (2018). EEG-Based Brain-Computer Interfaces for Communication and Rehabilitation of People with Motor Impairment: A Novel Approach of the 21st Century. Frontiers in Human Neuroscience. [61] Rudroff, T. (2025). Non-Invasive Brain Stimulation and Artificial Intelligence in Communication Neuroprosthetics: A Bidirectional Approach for Speech and Hearing Impairments. Brain Sciences. [62] Zrenner, C., Belardinelli, P., Müller‐Dahlhaus, F., et al. (2016). Closed-Loop Neuroscience and Non-Invasive Brain Stimulation: A Tale of Two Loops. Frontiers in Cellular Neuroscience. [63] Zhao, Z., Li, Y., Peng, Y., et al. (2025). Multi-view graph fusion of self-weighted EEG feature representations for speech imagery decoding. Journal of Neuroscience Methods. [64] Bhadra, K., Giraud, A., Marchesotti, S. (2025). Learning to operate an imagined speech Brain-Computer Interface involves the spatial and frequency tuning of neural activity. Communications Biology. [65] Kusano, K., Hayashi, M., Iwama, S., et al. (2025). Improved motor imagery skills after repetitive passive somatosensory stimulation: a parallel-group, pre-registered study. Frontiers in Neural Circuits. [66] Paek, A., Brantley, J., Ravindran, A., et al. (2021). A Roadmap Towards Standards for Neurally Controlled End Effectors. IEEE Open Journal of Engineering in Medicine and Biology. [67] Szostak, K., Grand, L., Constandinou, T. (2017). Neural Interfaces for Intracortical Recording: Requirements, Fabrication Methods, and Characteristics. Frontiers in Neuroscience. [68] Enriquez‐Geppert, S., Huster, R., Herrmann, C. (2017). EEG-Neurofeedback as a Tool to Modulate Cognition and Behavior: A Review Tutorial. Frontiers in Human Neuroscience. [69] He, B., Baxter, B., Edelman, B., et al. (2015). Noninvasive Brain-Computer Interfaces Based on Sensorimotor Rhythms. Proceedings of the IEEE. [70] Wilson, G., Willett, F., Stein, E., et al. (2024). Long-term unsupervised recalibration of cursor BCIs. [71] Myrden, A., Chau, T. (2015). Effects of user mental state on EEG-BCI performance. Frontiers in Human Neuroscience. [72] Carnicer‐Lombarte, A., Chen, S., Malliaras, G., et al. (2021). Foreign Body Reaction to Implanted Biomaterials and Its Impact in Nerve Neuroprosthetics. Frontiers in Bioengineering and Biotechnology. [73] Szostak, K., Grand, L., Constandinou, T. (2017). Neural Interfaces for Intracortical Recording: Requirements, Fabrication Methods, and Characteristics. Frontiers in Neuroscience. [74] Spencer, K., Sy, J., Ramadi, K., et al. (2017). Characterization of Mechanically Matched Hydrogel Coatings to Improve the Biocompatibility of Neural Implants. Scientific Reports. [75] Yang, W., Gong, Y., Li, W. (2021). A Review: Electrode and Packaging Materials for Neurophysiology Recording Implants. Frontiers in Bioengineering and Biotechnology. [76] Musk, E., Neuralink (2019). An Integrated Brain-Machine Interface Platform With Thousands of Channels. Journal of Medical Internet Research. [77] Orlemann, C., Boehler, C., Kooijmans, R., et al. (2024). Flexible Polymer Electrodes for Stable Prosthetic Visual Perception in Mice. Advanced Healthcare Materials. [78] Kotov, N., Winter, J., Clements, I., et al. (2008). Nanomaterials for Neural Interfaces. Advanced Materials. [79] Aregueta‐Robles, U., Woolley, A., Poole‐Warren, L., et al. (2014). Organic electrode coatings for next-generation neural interfaces. Frontiers in Neuroengineering. [80] Paulk, A., Kfir, Y., Khanna, A., et al. (2023). Large-scale neural recordings with single-cell resolution in human cortex using high-density Neuropixels probes. [81] Kotchetkov, I., Hwang, B., Appelboom, G., et al. (2010). Brain-computer interfaces: military, neurosurgical, and ethical perspective. Neurosurgical FOCUS. [82] Katz, N., Dworkin, R., North, R., et al. (2021). Research design considerations for randomized controlled trials of spinal cord stimulation for pain: Initiative on Methods, Measurement, and Pain Assessment in Clinical Trials/Institute of Neuromodulation/International Neuromodulation Society consensus recommendations. Pain. [83] Krucoff, M., Rahimpour, S., Slutzky, M., et al. (2016). Enhancing Nervous System Recovery through Neurobiologics, Neural Interface Training, and Neurorehabilitation. Frontiers in Neuroscience. [84] Soekadar, S., Birbaumer, N., Slutzky, M., et al. (2014). Brain-machine interfaces in neurorehabilitation of stroke. Neurobiology of Disease. [85] Morone, G., Pisotta, I., Pichiorri, F., et al. (2015). Proof of Principle of a Brain-Computer Interface Approach to Support Poststroke Arm Rehabilitation in Hospitalized Patients: Design, Acceptability, and Usability. Archives of Physical Medicine and Rehabilitation. [86] Athanasiou, A., Mitsopoulos, K., Praftsiotis, A., et al. (2022). Neurorehabilitation Through Synergistic Man-Machine Interfaces Promoting Dormant Neuroplasticity in Spinal Cord Injury: Protocol for a Nonrandomized Controlled Trial. JMIR Research Protocols. [87] Dams, J., Zapp, J., König, H. (2023). Modelling the Cost Effectiveness of Treatments for Parkinson’s Disease: An Updated Methodological Review. PharmacoEconomics. [88] Deeb, W., Giordano, J., Rossi, P., et al. (2016). Proceedings of the Fourth Annual Deep Brain Stimulation Think Tank: A Review of Emerging Issues and Technologies. Frontiers in Integrative Neuroscience. [89] Johnson, K., Dosenbach, N., Gordon, E., et al. (2024). Proceedings of the 11th Annual Deep Brain Stimulation Think Tank: pushing the forefront of neuromodulation with functional network mapping, biomarkers for adaptive DBS, bioethical dilemmas, AI-guided neuromodulation, and translational advancements. Frontiers in Human Neuroscience. [90] Liliya Hogaboam (2000). Assessment of Technology Adoption Potential of Medical Devices: Case of Wearable Sensor Products for Pervasive Care in Neurosurgery and Orthopedics. [91] Shah, S., Robinson, I. (2006). User involvement in healthcare technology development and assessment: Structured literature review. International Journal of Health Care Quality Assurance. [92] Flávia Xavier Macedo de Azevedo, Rüdiger Heimgärtner, Karsten Nebe (2022). Development of a metric to evaluate the ergonomic principles of assistive systems, based on the DIN 92419. Ergonomics. [93] Chaudhry, B., Shafeie, S., Mohamed, M. (2023). Theoretical Models for Acceptance of Human Implantable Technologies: A Narrative Review. Informatics. [94] Cucinella, S., de Winter, J., Grauwmeijer, E., et al. (2024). Towards personalized immersive virtual reality neurorehabilitation: a human-centered design. Journal of NeuroEngineering and Rehabilitation. [95] Burwell, S., Sample, M., Racine, É. (2017). Ethical aspects of brain computer interfaces: a scoping review. BMC Medical Ethics. [96] Morone, G., Pirrera, A., Meli, P., et al. (2022). Ethics and Automated Systems in the Health Domain: Design and Submission of a Survey on Rehabilitation and Assistance Robotics to Collect Insiders’ Opinions and Perception. Healthcare. [97] Cinel, C., Valeriani, D., Poli, R. (2019). Neurotechnologies for Human Cognitive Augmentation: Current State of the Art and Future Prospects. Frontiers in Human Neuroscience. [98] Paek, A., Brantley, J., Ravindran, A., et al. (2021). A Roadmap Towards Standards for Neurally Controlled End Effectors. IEEE Open Journal of Engineering in Medicine and Biology. [99] Jamil, N., Belkacem, A., Ouhbi, S., et al. (2021). Noninvasive Electroencephalography Equipment for Assistive, Adaptive, and Rehabilitative Brain–Computer Interfaces: A Systematic Literature Review. Sensors. [100] Chen, J., Xia, Y., Zhou, X., et al. (2023). fNIRS-EEG BCIs for Motor Rehabilitation: A Review. Bioengineering. [101] Oby, E., Golub, M., Hennig, J., et al. (2019). New neural activity patterns emerge with long-term learning. Proceedings of the National Academy of Sciences. [102] Ramirez–Zamora, A., Giordano, J., Gunduz, A., et al. (2018). Evolving Applications, Technological Challenges and Future Opportunities in Neuromodulation: Proceedings of the Fifth Annual Deep Brain Stimulation Think Tank. Frontiers in Neuroscience. [103] Dreger, M., Eckhardt, H., Felgner, S., et al. (2021). Implementation of innovative medical technologies in German inpatient care: patterns of utilization and evidence development. Implementation Science. [104] Velliste, M., Perel, S., Spalding, M., et al. (2008). Cortical control of a prosthetic arm for self-feeding. Nature. [105] Sahasrabuddhe, K., Khan, A., Singh, A., et al. (2020). The Argo: A 65,536 channel recording system for high density neural recording in vivo.