Advances in clinical brain–computer interfaces for assistive substitution and rehabilitation: A rapid scoping review

2.1 Literature search strategy

Relevant literature was identified through electronic database searches in PubMed, IEEE Xplore, Scopus, and Google Scholar, covering the period up to September 2023. Keywords used in the search included ‘brain–computer interface’, ‘BCI’, ‘neuroprosthetics’, ‘neural signal processing’, ‘neurorehabilitation’, ‘neuromodulation’, and ‘closed-loop stimulation’. Articles were also identified by manually reviewing the reference lists of relevant publications.

2.2 Inclusion and exclusion criteria

To map the breadth and trends of BCI literature, we included peer-reviewed studies irrespective of source, including primary research, reviews, and nonempirical evidence published in English. Articles were included if they discussed BCI technologies, signal acquisition methods, data processing techniques, BCI applications in assistive communication or rehabilitation, or ethical issues related to BCIs. A particular emphasis was placed on recent advances in BCI systems, especially those involving advanced feature selection and neural decoding algorithms, as well as clinical trials of adaptive neuromodulation. Studies including local research contributions from Hong Kong were highlighted. Given the scoping nature of the review, studies were selected based on their relevance and contribution to the understanding of BCI technologies and applications.

2.3 Data extraction and synthesis

Data from the selected articles were extracted and organised into thematic areas corresponding to the sections of the review: signal acquisition methods, signal processing and enhancement, effectors and applications, and ethical considerations. Timelines of key developments, findings, and advancements were synthesised to provide an informed discussion of current and prospective applications of BCIs.

3.1 Non-invasive modalities

One of the most common non-invasive techniques is EEG. While having reasonable temporal resolution, its spatial resolution is limited by poor signal attenuation through physical barriers of the CSF, skull, and scalp.² EEG recordings are also susceptible to mechanical, electromyographic, and electrooculographic artefacts.¹²

Magnetoencephalography (MEG) records post-synaptic activity with magnetic fields; functional magnetic resonance imaging detects oxygenation within brain regions as a surrogate marker for neuronal activity with magnetic fields; near-infrared spectroscopy detects oxygenation by photon activity. Non-invasive techniques are relatively cheap and have fewer risks to the patients, but suffer from low signal specificity, low signal-to-noise ratio, and signal distortion.¹³

3.2 Invasive modalities

Neuroprostheses may be implanted surgically to obtain neural data from the central or peripheral nervous system. With surgical risks come more sensitive and specific data (Figure 1).

Electrocorticography (ECoG) describes signals acquired from electrodes placed either subdurally or epidurally.¹⁴ Local field potentials are detected as summative neuronal processes surrounding the recording electrode. An example of the robustness of ECoG technologies can be found during the 2022 Annual Scientific Meeting of the Hong Kong Neurosurgical Society, where Professor Chang⁹ from UC San Francisco introduced his work in speech rehabilitation, allowing speech to be decoded in real time with local field potentials acquired from ECoG electrodes implanted across temporal and frontal cortices. Signal activity down to the individual neuron could be recorded using microelectrodes. The smaller the calibre of the electrode, the higher the selectivity, but also the higher the impedance. Microelectrodes at diameters of less than 100 μm are commonly found to be optimal.¹⁵

The first microelectrodes were conductive microwires insulated except at the tip. As they only record activity at the exposed tip, increasing the number of recording sites would require increasing the number of electrodes, resulting in a linear increase in probe size and thus neural trauma and gliosis. Silicon-based probes were created as a solution to increase the number of recording sites without increasing the overall probe size. Metal is melded through silicon coating with patterns that expose the metal to form recording sites, read-out pads for connecting to external circuitry, and interconnecting traces between the recording sites and read-out pads.¹⁶

Such rigid interfaces, however, suffer from gradual recording quality degradation, which is thought to arise from tissue damage and the ensuing immune response. Softer and more flexible neural probes could be fabricated in a polymeric fashion to mitigate the textural mismatch between probe and brain. Temporary stiffness that may be required during implantation could be imparted by additional surgical shuttles or biodegradable coatings.^{17, 18}

Although sensitive, the quality of signals acquired from more invasive techniques may be affected by foreign body response. The blood–brain barrier and absence of immune cells within the central nervous system mean that it responds differently against foreign bodies. It is thought that the disruption of blood–brain barrier induces reactive gliosis, where microglia and astrocytes activate and encapsulate said foreign body. With time, the electrode may be subject to buckling, corrosion, and degradation, while the brain may suffer from injury due to micromotion of the electrodes.¹³ Hybrid systems could potentially incorporate the strengths of different technologies while minimising their drawbacks. Different signal inputs from the brain could be integrated simultaneously or sequentially, while signals may be extracted from even outside the brain in neurovisceral circuitry.

4.1 Sources of noise

EEG signals collected from the surface of the scalp are highly contaminated with various types of internal and external artefacts and background noise. The recording unit, including amplifiers and cables, and subject actions, including breathing and blinking, contribute towards disturbances of EEG data.^{21, 22} EEG data are particularly prone to such disturbances because of their very small amplitude.²³

Myopotentials caused by muscle contraction around the temple and forehead are the most common physiological interferences in EEG.²⁴ These artefacts, including extraocular and tongue movements, are also among the most easily identified in EEG recordings due to their expressive amplitudes compared with neuronal potentials and distinctive shape. Extraocular movements generate an alternating current with a high amplitude which is visible on electrodes around the eyes. Unique patterns, meanwhile, are often observed in movement-related disorders (eg, rhythmic sinusoidal artefacts of 4–6 Hz for PD).²⁵ EEG and MEG signals can also encounter strong power line interference at 50 or 60 Hz.²⁶ Although shielded rooms can theoretically reduce the impact of interference on recordings, the mobility of EEG systems in clinical BCI systems makes effective shielding impractical .^{27, 28}

4.2 Denoising techniques

Without proper noise removal techniques, brain signals may be completely suppressed and illegible for further qualitative or quantitative analysis.²⁹ The field of denoising techniques designed for either single- or multi-channel data is rather mature.^{30, 31} The major techniques are adaptive filters, wavelet transform (WT), or independent component analysis (ICA), the characteristics of which are briefly summarised in Table 1. Certain techniques, including empirical mode decomposition and time–frequency (T–F) image dimensionality reduction, are excluded from further discussion as they cannot be implemented in real time, which heavily restricts their feasibility within BCI systems.^{32, 33}

Traditional regression methods have been in use for many years. Regression analysis first defines the amplitude relation between a reference channel and the channel of interest, then subtracts estimated artefacts from the raw signal. This algorithm thus requires exogenous reference channels (ie, electrooculography, electrocardiogram) to omit different artefacts. Mutual contamination between the signal and the reference channel, however, may reduce demising effectiveness.³⁴

Often, reference signals are lacking, and the involved neurophysiological processes behind artefacts are complex. Denoising can therefore be considered a blind source separation (BSS) problem, where the original sources underlying the signals are approximated, without a priori knowledge about the physiological sources themselves, and without knowledge about the mixing of such sources. Four main techniques have emerged to specifically tackle the difficult problem of denoising in a BSS context, including WT, ICA, canonical component analysis (CCA), and empirical mode decomposition.

WT was conceptualised in the early 1980s and is performed by deconstructing a signal into the time and frequency domains. Thresholding is then applied to discard the artefacts from the raw signal, where the frequencies of artefacts from experience or from literature are used as the threshold. The remaining data are subsequently added to reconstruct the clean signal.³⁵

The concept of ICA was introduced in 1986 especially to solve the BSS problem.³⁶ ICA operates under three assumptions: that source signals exhibit statistical independence and are instantaneously mixed; the observed signal must be equal to or greater than that of the source signal³⁷; and that sources are either non-Gaussian or only one source is Gaussian. That the true signal is assumed to be mixed with signal artefacts at each time point independently allows the observed signal to be decomposed into independent components. In general, artefacts such as eye movements, eye blinks, and cardio activities originate from mutually independent sources, and volume conduction is linear, rendering the first assumption reasonable. However, physiological signals in the brain are not mixed and propagated instantaneously in a linear fashion; therefore, a convolutive ICA approach, which takes into account weighted and delayed contributions of signals, has also been proposed.³⁸

CCA, by contrast, describes a method to measure the linear relationship between two multi-dimensional random variables.³⁹ The use of CCA in removing muscle artefacts was first described in 2006, where it outperformed low-pass filtering and ICA techniques.⁴⁰ CCA further differs from ICA in that the former brings shorter computational time compared with the use of higher-order statistics in ICA.

In recent years, neural networks have been increasingly adopted in denoising due to their real-time applicability and high accuracy. One of the first deep convolutional auto-encoders designed to eliminate ocular and jaw-clenching artefacts demonstrated superiority over traditional filtering methods.⁴¹ The dedicated removal of ocular artefacts using deep learning was further explored in Yang et al.⁴² Further research demonstrated promise for neural networks to remove muscular and cardiac artefacts. Neural plasticity as a result of pathological events such as epilepsy, drug alterations, rehabilitation regimens, and biological cycles (eg, circadian rhythms) might alter the dynamics of the brain and require adaptive retraining of neural networks before generalising to unseen signals.^{43, 44}

Hybrid methods have also been proposed, offering better performance at the cost of increased computational complexity. Several studies combined WT and ICA to separate artefact-independent components, resulting in clean data within improved computational times.⁴⁵ For removing ocular artefacts, the combination of WT-based and adaptive noise cancelling–based methods showed promise. Jafarifarmand et al.⁴⁶ investigated real-time processing and ocular artefact removal using ICA and adaptive noise cancelling, focusing on a small number of EEG channels, making it suitable for BCI applications.

4.3 Feature extraction

Broadly, BCI signal processing, following denoising, involves feature extraction and classification (Figure 3). Task-specific features are extracted from EEG/ECoG signals using various methods. Feature extraction, in essence, involves frequency filtering, windowing, feature extraction, and feature selection which outputs selected features for the classifier. In BCIs, band power features represent the power of signals (relative amplitude) for a frequency band (a certain range of frequencies) averaged over a time window. Common feature extraction techniques are fast Fourier transform, autoregressive model, WT (shared between denoising and extraction techniques), and common spatial pattern. Although a brief comparison of the techniques is shown in Table 2, review articles with a specific technical focus on such techniques may provide the reader with an enhanced level of understanding.^{47, 48}

The main consideration underlying various techniques is the forced trade-off between time resolution and frequency resolution. The trade-off exists because when the time window for analysis is fixed, the longer the length of the window used, the higher the frequency resolution and the lower the achieved time resolution. This trade-off is named the uncertainty principle of signal analysis, analogous to Heisenberg's Uncertainty Principle commonly applied in quantum mechanics.⁴⁹

4.4 Classification

Once features are extracted and selected, classification can be conducted to predict target variables from those discriminative features. Linear discriminative analysis is a popular classifier in EEG-based BCI applications due to its low computational requirements and has been successfully used for classifying right- and left-hand MI. However, it performs poorly on complex non-linear EEG data.⁵⁰ Support vector machine is widely used in BCI research, as it is known for its good generalisation properties across individuals and performance with high-dimensional data. Meanwhile, neural networks offer a reasonable trade-off between accuracy and speed, making them extensively used in BCI research. Deep learning further improves accuracy in classification, by overcoming the obstacle of arbitrarily selecting weights within traditional neural networks. A neural network is a series of nodes. Within each node is a set of inputs and corresponding weights. Weights convey the importance of that corresponding feature in predicting the final output (classification), and deep learning attempts to optimise the weights themselves. Recent applications of deep learning include the classification of multiclass MI tasks.⁵¹

5.1 Directions for BCI applications

The current BCI research revolves around two arms, propelled by both publicly and privately sponsored research efforts. The Brain Research through Advancing Innovative Neurotechnologies (BRAIN) Initiative of the National Institutes of Health (NIH) has funded hundreds of academic, neuroscientific, and neuroprosthetic research programs, yielding major advances in neural interface technologies. Numerous companies (Emotiv Inc., Neuralink Corp., Kernel, Paradromics Inc., InBrain Neuroelectronics) and corporate ventures (Google Brain, Facebook Reality Labs) have also been established seeking to enable volitional interaction with prostheses and computers beyond pure clinical applications.⁵²

5.2 Assistive communication and substitution

The first arm of BCI applications revolves around assistive communication or achieving the substitution of sensory and motor functions, and the second arm relates to closed-loop rehabilitation or neuromodulation. In assistive communication, brain activity is processed and used to generate a feed-forward route to control external devices. The prosthesis, the final stage of the BCI, can take many forms from motor to visual. A surrogate arm and hand may directly substitute biological functional deficits for patients with amputated arms, while a tool, such as a computer or a cursor, provides special functionality to the patient.

The identification of strong, reproducible brain signals can be used as the basis for driving a BCI, and foundational studies have demonstrated the ability to control various robotic arms with decoder outputs.^{8, 53} Motor neuroprostheses in humans have progressed dramatically in the past decades, with the milestones briefly summarised in Table 3. The field has developed from slow predictions to systems achieving the substitution of all limbs of the body.

^aBCI, brain–computer interface.

5.3 Motor imagery

A primary modality of assistive motor devices is based on MI, which involves imagining the execution of a movement, accessing motor representation consciously.^{62, 63} This can be done from a first-person (internal) or third-person (external) perspective. Research has shown that MI and motor representation share similar functional relationships and mechanisms, activating the same brain areas during actual and imagined movement.^{64, 65} Analysing EEG signals from the primary sensorimotor cortex during executed or imagined movement reveals activation of certain frequencies.^{66, 67} The variations in amplitude of such frequency bands, called event-related desynchronisation and synchronisation, correspond to specific states of movement.⁶⁸ MI training improves both motor execution and neuroplasticity.⁶² However, MI abilities are subjective, requiring extensive assessment and training before application.^{69, 70}

Hochberg's team⁵³ first applied such findings to an individual with tetraplegia with a 96-channel MEA implanted in the primary motor cortex's hand area. Using algorithms to decode 2D movement intentions, the patient was able to open and close a prosthetic hand, control a computer cursor, and operate software and discrete neural states. This technology was further expanded to control robotic arms with up to three degrees of freedom (DOF), allowing hand grasping and self-feeding functionality.⁸ Other groups have devised similarly successful motor prostheses based on MEAs in the motor cortex, including success up to seven DOF.⁷¹ In 2019, Benabid and colleagues⁷² provided a proof-of-concept demonstration of a four-limb exoskeleton controlled by a fully implantable ECoG recording system with 64 epidural electrodes placed bilaterally over the motor cortices. The patient with tetraplegia achieved up to eight DOF without any recalibration over 7 weeks.

5.4 Functional electrical stimulation

Functional electrical stimulation (FES) is a technology that can reanimate paralysed muscles by delivering electrical stimulation to a muscle or the nerve innervating a muscle. Coordinated electrical stimulation can return function to any DOF of the body via synchronised muscle groups. FES can be deployed as the output of a BCI system to reanimate the upper extremity, using pattern recognition from residual functional muscles.^{36, 73, 74} Recent work has progressed further into individual reanimation of digits with a BMI-controlled FES system, using MI of the finger movements as a direct control signal for stimulation.^{75, 76}

5.5 Communication prostheses

While most BCIs have primarily focused on motor replacement, recent advancements have expanded into interfaces for speech and communication restoration, which would be transformative for people who are unable to communicate because of neurological impairments.^77-80 Early attempts of EEG spellers revolved around expanding simpler MI of arm and hand movements to create interfaces for speech and communication replacement, achieving 1 to 5 characters per minute.^{81, 82} Subsequent iterations utilising visually evoked potentials to guide cursors or keyboard inputs have achieved speeds of 60 characters per minute, but the need for constant gaze direction and panels of flashing lights on a screen remain key obstacles to naturalistic real-life applications.⁸³ Henderson, Shenoy, Hochberg, and colleagues⁸⁰ subsequently demonstrated that high-performance assistive communication systems were possible through higher-density intracortical 96-channel silicon MEAs (Utah arrays) implanted in the hand area of the motor cortex, and even faster communication was achievable by decoding rapid sequences of highly dexterous MI, such as handwriting.⁷⁹ The key criteria for clinical applications were now met, including communication being entirely self-paced with the eyes free to move, and sufficient accuracy offline with a large-vocabulary language model. While these strategies achieve speeds comparable to able-bodied smartphone typing speeds (12–19 words per minute),^{84, 85} this remains inferior to the average of 150 words per minute of natural speech.⁸⁶

Going forward, natural speech–based approaches that directly translate neural activity into speech appear to hold the key to restoring intuitive communication. Initial strategies for neural decoding of speech production focused on the direct classification of speech segments such as phonemes or words.⁸⁷ However, these attempts were limited to small vocabulary sizes and slow communication rates. By contrast, in addition to the capability to communicate unconstrained vocabularies at a natural speaking rate, direct speech synthesis can capture prosodic elements of speech that are not available with text output, such as pitch intonation.⁸⁸ In 2019, Edward Chang and colleagues¹⁰ designed a neural decoder that leverages representations of vocal tract articulatory movements encoded in the speech motor cortex to synthesise high-fidelity intelligible speech. This was subsequently translated to a paralysed patient with anarthria, who used attempted speech to control algorithms that decoded intended words and sentences in real time. Decoding neural-activity patterns in the ventral sensorimotor cortex combined with a natural-language model providing next-word probabilities given the preceding words in a sequence enabled higher communication rates and accuracy.⁸⁹ A follow-up study demonstrated the continued stability of the relevant cortical activity beyond 2 years following device implantation, and that silently attempted speech could be as effective as an alternative behavioural strategy to imagined speech and overt-speech attempts.⁹⁰ Recent studies have further advanced the field, with decoding algorithms achieving higher communication rates at 62 to 78 words per minute,^{91, 92} and a bilingual speech neuroprosthesis capable of decoding cortical activity into sentences in both English and Spanish.⁹³ The existence of shared cortical articulatory representations across languages may enable decoding multiple languages without the need to train language-specific models. This study also demonstrated stable decodable cortical activity 4 years after implantation. However, challenges remain in generalising these technologies to patients fully locked in with negligible residual vocal tract function, beyond the spectrum of imagined speech, silently attempted speech, and overt-speech attempts currently studied in patients with some residual vocal cord function. Imagined speech involves internally simulating speech without actual articulatory movements, which may engage different neural networks compared with overt or attempted speech. Given the diversity of pathologies that result in dysarthria, speech-decoding studies may also consider recording from intact speech production regions such as the posterior superior temporal gyrus (STG), supramarginal gyrus (SMG) afferent signal, or preserved aspects of the precentral gyrus to accommodate cortical forms of dysarthria.⁹⁴

5.6 Closed-loop rehabilitation and neuromodulation

The other arm in BCI applications is closed-loop rehabilitation or neuromodulation, where the dynamic feedback loop aims to restore neural plasticity through reinforcement training or the modulation of brain processes. These closed-loop systems which respond dynamically to neurophysiological changes can also be termed adaptive neuromodulation.

A leading application of adaptive neuromodulation is in aDBS, also known as closed-loop DBS, which delivers stimulation in response to a specific electrophysiological brain marker that represents periods of aberrant neural activity in which stimulation would be needed, such as resolving freezing of gait in PD.^95-98 Although conventional DBS has been clinically validated to reduce motor fluctuations in medically refractory advanced PD, many patients require a combination of levodopa and stimulation for best function. Constant stimulation disregards peaks and valleys in the effective brain concentration of levodopa, so patients may experience residual motor fluctuations that come with high or low dopaminergic states. aDBS could allow a reduction in stimulation current without loss of therapeutic benefit, reducing stimulation-induced adverse effects.

Traditionally, subthalamic beta oscillations (13–30 Hz) have been used as control signals due to their association with motor symptoms in PD. In proof-of-principle studies, Simon Little and colleagues⁹⁹ demonstrated not only the feasibility of successful BCI-controlled DBS in patients with PD, but also the associated improvement in motor scores and reduction in necessary stimulation amplitude, which has an impact on both side effects and the lifetime of implanted battery systems. Alberto Priori's group¹⁰⁰ subsequently studied aDBS effectiveness across 8 hours, demonstrating the potential for hours of optimised unrestricted patient activity, and in conjunction with levodopa. These initial studies recorded subthalamic oscillations via externalised leads and used external customised neurostimulators, until Helen Bronte-Stewart and colleagues⁹⁷ demonstrated the tolerability and efficacy of aDBS using a fully implanted neurostimulator capable of concurrent neural sensing and stimulation. The common approach involved the use of beta oscillations as a control signal; however, during active stimulation, beta oscillations may become less reliable due to stimulation-induced suppression of beta activity.^101-103 In a blinded, randomised, cross-over feasibility trial, Starr and colleagues¹¹ demonstrated the efficacy of aDBS using stimulation-entrained gamma oscillations as control signals in patients with PD. The study found that stimulation-entrained gamma oscillations, centred at half the stimulation frequency (~65 Hz), in either the subthalamic nucleus or the motor cortex, served as optimal biomarkers of motor function fluctuations during ongoing stimulation. These gamma oscillations were robust across varying stimulation amplitudes and correlated with patients’ dopaminergic states, outperforming traditional beta-band signals in predicting motor symptoms. Multi-site brain recordings, including ECoG from the motor cortex, provided additional neural signals that were not detectable from subcortical recordings alone. By utilising these signals, the aDBS system could adjust stimulation parameters in real-time, improving motor symptoms and quality of life compared with clinically optimised standard stimulation. In some patients, cortical stimulation-entrained gamma activity was the only reliable biomarker for adaptive control, underscoring the added value of cortical recordings in aDBS, and may signal surgical implications for future iterations.

The potential to tailor stimulation parameters based on real-time neural feedback holds promise for enhancing therapeutic outcomes while minimising adverse effects in patients with PD and other neurological disorders. Similar aDBS modalities have been investigated in other conditions with varying levels of success, including essential tremor,¹⁰⁴ obsessive-compulsive disorder,^{105, 106} and epilepsy.^{107, 108}

5.7 BCIs as neurofeedback

In rehabilitation, traditional therapies for restoring motor function focus on the forced use of the impaired limb, for improving muscle strength and performance of activities of daily living.^{109, 110} However, the success of rehabilitation relies on the patient having some residual function. Alternative methods for restoring function have thus been studied, using BCIs as a neurofeedback tool to engage neuroplasticity along the corticospinal pathway. Users receive real-time feedback on particular neural activity features, such as oscillatory signals from the cortex, and learn to control the target neural signal. Literature has shown that functional recovery is often associated with cortical activity returning to a state close to that of an unimpaired individual.^{111, 112} Neurofeedback-driven BCI training can directly impact cortical activity and possibly restore regular cortical activation patterns, potentially benefiting downstream neuromuscular systems.

Recent studies have investigated the feasibility of restoring movement using BCI training in patients with chronic stroke. In early studies, patients learned to modulate their brain activity in order to control an orthosis that opens and closes the hand, which led to improvements in motor function.^113-116 In one study, participants with chronic hand weakness secondary to stroke underwent training to modulate sensorimotor rhythm amplitudes activated during imagery of movement of the paralysed hand, as measured by an MEG-based BCI system.¹¹⁷ Drawing inspiration from Hebb's rule for neuroplasticity where ‘the neurons that fire together, wire together’,¹¹⁸ the sensorimotor rhythm amplitude was set to control the vertical position of a computer cursor. In addition, the same neural signals were simultaneously used to control the movement of a hand orthosis. Remarkably, most patients were able to achieve control with MEG sensors at sensorimotor areas ipsilateral to their subcortical stroke lesions, suggesting that imagery training can lead to cortical reorganisation and enhanced neurocognitive rehabilitation when coupled with somatosensory feedback.

Another feasibility study tested an EEG-based BCI system that used signals related to affected hand MI, recorded from the unaffected hemisphere, to control the affected hand via a powered exoskeleton.¹¹⁹ The unaffected hemisphere demonstrated strong potential in the BCI rehabilitation system which improved patients’ Action Research Arm Test (ARAT) scores surpassing the minimal clinically important difference.¹²⁰ In a transcranial magnetic stimulation–based study, neurofeedback increased the corticospinal excitability of affected muscles in patients with stroke with severe upper limb paralysis.¹²¹ Taken together, these results build on previous evidence that BCI-controlled rehabilitation systems can facilitate motor recovery.¹²²

5.8 Prospective rehabilitation in neuropsychiatry

Closed-loop BCIs may have further applications in neuropsychiatric illnesses. It has been shown that sensory-affective experiences can be modulated by a BCI in real time in freely behaving rats.¹²³ By coupling neural codes for nociception directly with optogenetic activation of the prelimbic prefrontal cortex, the BCI effectively inhibited affective behaviours caused by a multitude of pain, including acute mechanical or thermal pain, and chronic inflammatory or neuropathic pain. Furthermore, modelling frameworks have recently been developed to decode mood state variations from intracranial recordings in patients with epilepsy.^{124, 125} Similar to how motor BCIs encourage patients to learn to better control them through neural adaptation, driven by somatosensory feedback of the decoded output, providing explicit feedback of the decoded mood state to the patient may lead to successful rehabilitation where patients learn to control their own mood states.^{126, 127}

7.1 Advancements in BCI electrodes

The Utah array has been widely used in human BCIs with 48 implants as of 2018.^{138, 139} While no replacement designs come close to demonstrating similar robustness and clinical dependability over a decade, pre-clinical testing has revealed promising electrode designs that could soon enter clinical use. Devices such as Neuropixels with 384 channels and other high-density silicon probes allow far greater channel count in rodent experiments.^{16, 140, 141} Invasive intracranial electrodes currently used in BCIs are metal based, which presents significant distortion when capturing the full spectrum of electrophysiological signals, where low frequencies tend to be attenuated and distorted.¹⁴² Field-effect transistors have been developed as a non-inferior alternative. As they are active devices, they can act as inherent signal amplifiers to increase the signal-to-noise ratio without interference from pre-amplified noise from electrical connections.¹⁴³ Novel materials promise to improve electrochemical inertness, electrical conductivity, and biocompatibility to produce flexible and ultrathin devices. These include, among others, graphene,¹⁸ carbon fibres,¹⁴⁴ and silicon carbide.¹⁴⁵ These electrode materials may allow designs smaller than neurons promise to drastically reduce bleeding and gliosis, opening further possibilities for high-channel-count chronic BCI implants.¹⁴⁶

7.2 Challenges in BCI adoption

In April 2021, the FDA authorised the first brain-controlled hand exoskeleton for neurorehabilitation, IpsiHand by Neurolutions. This device represents a step towards integrating BCIs into rehabilitative care. However, several challenges must be addressed before wider adoption can occur, such as patient stratification and personalised treatment, integration into existing treatment plans, and commercialisation of BCIs.

Patients selected for BCIs, including survivors of stroke and spinal cord injury, exhibit a high degree of heterogeneity, requiring personalised rehabilitation plans.¹⁴⁷ Factors such as BCI training duration and intensity should be tailored to individual patients—yet evidence is currently limited as to which patients benefit most from BCI-enabled neurorehabilitation training given publication bias towards successful cases, and more research is needed to establish optimal training parameters. Clear definitions of inclusion and exclusion criteria are a crucial foundation for comparisons across BCI frameworks, and to determine a significant level of efficacy.

Integrating BCIs into existing treatment plans remains a key challenge, as rehabilitation approaches vary across different centres. Implants for neurological conditions are also frequently viewed as last-resort options, limiting sample size. Early adopters, who are open to innovation, are crucial in promoting BCI use among clinicians and patients as they may help shape future workflows.¹⁴⁸ Similar to biomedical devices such as open-loop deep brain stimulators, the commercialisation of advanced BCIs faces several obstacles, including complex certification processes, which vary geographically and have limited distribution pathways. Reimbursability is also a critical factor, as health insurances require clear, evidence-based benefits before covering costs for medical devices or treatments.

Technical hurdles must be overcome to demonstrate clinical efficacy and safety in sufficiently large cohorts. Obtaining safety and efficacy data is a long and costly process, and the commercialisation of academic work is hampered by long regulatory processes and technology risks. Furthermore, academic funding, degree cycles, and publishing dynamics rarely offer support to these multiyear studies with uncertain success. Despite challenges, recent studies have successfully demonstrated the clinical efficacy of select implantable neural interfaces, highlighting the need for sustained public funding and partnerships between academia and industry to advance the translation of BCIs. Scalable manufacturing design, quality control comprehension, and engineer-friendly system designs are crucial investments to further BCIs beyond proof-of-concept studies.^149-151