This narrative review provides a broad yet critical evaluation of the state-of-the-art BCI-based ARTs, employing a flexible method for reviewing and selecting the literature. Research databases, including CINAHL, Embase, Engineering Village, IEEE Xplore Digital Library, MEDLINE, and PubMed, were utilized to search for relevant literature. Publications addressing the current state of BCI-based ARTs were selected and included in this review. Search terms were used related to: (1) individualization of BCI; (2) user-centric design strategies for ARTs; and (3) to fulfil the evolving functional and developmental needs of users with disabilities.

This review contributes to the field by providing a holistic overview of BCI-based ARTs in terms of user-centric design philosophy to fulfil individualized user requirements of people with disabilities. Studies have investigated the importance of an inclusive, personalized approach for the development of communication BCI for children and adults with disabilities, not only due to inter-individual variability, but also due to the importance of including opinions from all stakeholders (Shrivastava et al., 2025; Branco et al., 2025; Russo et al., 2025; French et al., 2024; Ivanov et al., 2023; Branco et al., 2023). (Edelman et al. 2024) have recently evaluated current advancements in noninvasive BCI technologies. At the same time, other reviews have focused on invasive options (Dunlap et al., 2020; Ferguson et al., 2019). In contrast, this review compares both invasive and noninvasive neuroimaging techniques for their suitability in diverse ARTs for people with disabilities, with a perspective on user needs. (Patrick-Krueger et al. 2024) have explored the current state of clinical trials of implantable BCIs. Some reviews have investigated different digital signal processing, machine learning, and artificial intelligence techniques, as well as BCI decoding algorithms (Saha and Baumert, 2020; Azab et al., 2018; Lotte et al., 2018; Jayaram et al., 2016; Krusienski et al., 2011; Bashashati et al., 2007; Lotte et al., 2007). (Saha et al. 2021) published a review on BCI advancements, discussing the diverse application areas of BCI in general, unlike this review, which addresses individualized applications for people with disabilities. On the contrary, most, if not all, state-of-the-art reviews aim to consider a particular user group or a specific challenge or opportunity in terms of neuroimaging, signal processing, pattern recognition, clinical and socioeconomic aspects of BCI development (Patrick-Krueger et al., 2024; Edelman et al., 2024; Bergeron et al., 2023; Kaongoen et al., 2023; Karikari and Koshechkin, 2023; Lopez-Bernal et al., 2022; Karlsson et al., 2022; Keser et al., 2022; Fry et al., 2022; Gao et al., 2022; Hramov et al., 2021; Mrachacz-Kersting et al., 2021; Orlandi et al., 2021; Dunlap et al., 2020; Duun-Henriksen et al., 2020; Martini et al., 2020; Micera et al., 2020; Saha and Baumert, 2020; Ferguson et al., 2019; Azab et al., 2018; Deffieux et al., 2018; Lotte et al., 2018; Jayaram et al., 2016; Klein and Ojemann, 2016; Krusienski et al., 2011; Bashashati et al., 2007; Lotte et al., 2007; Dobkin, 2007).

The practical usability of a BCI depends on various factors, primarily the user-centric design of ARTs (Shrivastava et al., 2025; Russo et al., 2025; Branco et al., 2025; French et al., 2024; Ivanov et al., 2023; Kübler et al., 2014). The user-centric design is an iterative approach that includes users' perspectives while developing BCI-based ARTs to fulfil individualized user needs. The ICF framework aligns closely with user-centric design principles in BCI development. By recognizing the influence of environmental and personal factors, such as caregiver support, living conditions, and user preferences, the ICF supports the creation of adaptable technologies that reflect real-world diversity in user needs and contexts. It provides a holistic method to evaluate the usability of ARTs for people with disabilities. Some studies proposed user-centric design-based techniques for assessing usability of BCI applications in real-life settings, going beyond developer-centric evaluation criteria such as classification accuracy and information transfer rate (Ivanov et al., 2023; Kübler et al., 2014). Other studies conducted surveys on people with disabilities and their caregivers seeking their opinions on BCI-based ARTs (Shrivastava et al., 2025; Russo et al., 2025). Overall, BCIs are acceptable; however, the usability will depend on how developers integrate the input from the users in the development of BCI-based ARTs.

The clinical distinction of neurological variability defines disability types and promotes individualization of BCI-based ARTs. While one group of users encounters developmental disabilities like people with cerebral palsy, others experience acquired disabilities due to neurological incidents such as stroke, amyotrophic lateral sclerosis, and spinal cord injury (Metzger et al., 2023; Bekteshi et al., 2023; Hallett et al., 2022; Monforte et al., 2021; Makris et al., 2021; Micera et al., 2020; Goulet et al., 2019; Milekovic et al., 2018; Goldstein and Abrahams, 2013). Due to the diverse types and severities of impairments that reflect individualized user needs (Seghier and Price, 2018; Park et al., 2016), the user-centric and personalized design of BCI-based ARTs may better fulfil the individualized requirements of people with disabilities (French et al., 2024; Branco et al., 2023; Saha et al., 2021; Martin et al., 2018).

The principles of BCI-based assistive or communication technologies rely on how brain signals are translated into machine commands, for example, using two distinct brain activity patterns to control a switch (on/off) (Cao et al., 2014; Mitchell et al., 2023). These can take active, reactive and passive forms of translation through a BCI. The active BCI translates users' intentions, such as imagined movements and speech, into control signals or messages on a computer screen utilizing sophisticated decoding algorithms and software (Willett et al., 2023, 2021; Saha et al., 2021; Moses et al., 2021; Acı et al., 2019; Aricò et al., 2018; Zander et al., 2009). However, the success of these applications relies on how accurately the decoding algorithms can classify brain activities (Lotte et al., 2018; Krusienski et al., 2011; Bashashati et al., 2007; Lotte et al., 2007). Visual and auditory evoked potentials are signals due to external stimulation and define reactive BCI applications. Reactive BCIs are suitable for people who have severe disabilities, for example, individuals in locked-in states (Chaudhary et al., 2022). Depending on the disability type and severity, the specific signal type is selected to meet individualized user needs. For example, a person with a visual impairment is not likely to use a visually evoked potential-based BCI but may find an auditory cue-based stimulation more useful. While reactive BCIs are time-synchronized to external stimuli, active ones offer intuitive use for functional autonomy. For visually evoked potential, studies found that not all individuals with disabilities perform equally for varied types of external visual stimulation, e.g., steady-state visual evoked potential vs. P300, and checkerboard flashing vs. row-column flashing-based BCIs (Severens et al., 2014; Combaz et al., 2013; Townsend et al., 2010). Thus, a BCI system is unlikely to work well for all users, and poor performance with one system does not mean no system would work for that user. Finally, a passive BCI is applicable when monitoring users' affective states, such as detecting drowsiness. There is no one-size-fits-all; nonetheless, BCI-based ARTs require assessments of users' perspectives of gaining functional autonomy and codesign to best fulfil individualized needs (French et al., 2024; Huggins et al., 2015; Nijboer et al., 2014; Huggins et al., 2011). Individualized BCI design can be guided by the ICF's emphasis on activity limitations and participation restrictions. For example, a user's ability to engage in daily tasks or social roles may be constrained by both impairments and contextual barriers, which BCI-based ARTs can help overcome.

The principles of rehabilitative BCIs rely on promoting neuroplasticity, which refers to the adaptive behavior of central and peripheral neural networks. Neuroplasticity is a critical ingredient of the motor learning process and is at the center of BCI-driven rehabilitation (Flesher et al., 2021; Gao et al., 2022; Saha and Baumert, 2020; Dobkin, 2007). Neurostimulation modalities are the main components of neural encoding that externally modulate neural activities in target brain areas or peripheral neural networks to repair paretic functional abilities by rendering neuroplasticity (Flesher et al., 2021; Mrachacz-Kersting et al., 2016). They are feedback elements of some closed loop or open loop rehabilitative BCIs, mostly restimulating impacted neural networks in the brain and central and peripheral nervous systems. Studies proposed BCI-based rehabilitation strategies featuring transcranial magnetic stimulation and direct/alternating current stimulation for repairing stroke lesions in the brain (Keser et al., 2022; Van der Groen et al., 2022; Yang et al., 2021; Antal and Herrmann, 2016). Some implantable options include deep brain stimulation and intracortical microstimulation (Allert et al., 2018; Willett et al., 2023, 2021; Oxley et al., 2016). While the primary applications of neurostimulation techniques are in neurorehabilitation, they can also be used to improve signal quality and users' ability to operate a BCI (Won et al., 2023). (Lorach et al. 2023) developed a rehabilitative BCI featuring epidural nerve stimulation for rectifying spinal cord injury. Other studies presented BCI-driven functional electrical stimulation for rehabilitating impaired upper or lower limb functions (Brunner et al., 2024; Jovanovic et al., 2021; Mrachacz-Kersting et al., 2021, 2016). BCI-driven neurofeedback has been reported to regulate the cortical-subcortical networks and assists in modulating brain signals for cognitive or functional recovery (Silversmith et al., 2021; Sitaram et al., 2017; Orsborn et al., 2014). Notably, both neurofeedback and neurostimulation may involve decoding users' intentions from brain activities to operate ARTs. Neurostimulation is an essential element of a bidirectional BCI, specifically for neural encoding; however, a more detailed investigation on neurostimulation is beyond the scope of this review, and the subsequent sections will emphasize neuroimaging modalities.

Neuroimaging is a critical element of neural decoding, which defines the appropriateness of BCI-based ARTs across diverse user groups with different types of disabilities (Patrick-Krueger et al., 2024; Edelman et al., 2024; Karikari and Koshechkin, 2023; Moghimi et al., 2013). Typically, no single modality offers the spatial, spectral and temporal resolution for mapping all required complex brain functions, resulting in limited BCI control signals (Saha et al., 2021; Martini et al., 2020; Min et al., 2010). However, recently developed implantable speech BCIs have demonstrated up to 50-word decoding and constructing thousands of sentences by integrating language models (Willett et al., 2023, 2021; Moses et al., 2021). Electroencephalography (EEG), magnetoencephalography (MEG), electrocorticography (ECoG) and microelectrode arrays (MEA) capture fine temporal features (i.e., electrical activities). However, the sensors are either sparsely distributed (spatially) or localized within an area of interest only. Functional magnetic resonance imaging (fMRI) records high-resolution spatial features of the brain, but the unmanageable size and slow hemodynamic signals are unsuitable for many real-time applications (Sitaram et al., 2007). As an alternative, functional near-infrared spectroscopy (fNIRS) has lower spatial resolution but is more practical, mainly due to its portability and ease of use (Naseer and Hong, 2015). On the other hand, electric and magnetic field-based neuroimaging techniques, i.e., EEG and MEG, offer better speed and bandwidth. MEG lacks portability and is barely usable outside a specialized setting (Bu et al., 2023). Table 2 illustrates the key characteristics of the neuroimaging modalities and their applicability in BCI-based ARTs (Saha et al., 2021; Hramov et al., 2021; Deffieux et al., 2018; Nicolas-Alonso and Gomez-Gil, 2012; Min et al., 2010).

From a translational perspective, EEG is still one of the most viable neuroimaging modalities due to its portability, easy maintenance and low cost. It does not involve any surgical procedure, like a craniotomy for implantable ECoG and MEA. Moreover, EEG could enable wearable BCIs using tiny recording setups like ear EEG (Kaongoen et al., 2023). However, EEG signals are highly nonstationary and nonlinear due to time-variant and subject-specific anatomical, psychological, physiological, and environmental factors (Saha et al., 2023, 2021; Saha and Baumert, 2020; Saha et al., 2019, 2017a,b). Furthermore, EEG signals do not offer the required spatial resolution for mapping localized brain functions because the signals attenuate through the outer brain layers, skull, scalp, skin and hairs (Miinalainen et al., 2019; Cohen, 2017). This issue is somewhat alleviated with implantable ECoG and MEA, which record functionally relevant localized signals (Vansteensel et al., 2016). However, the higher risk associated with craniotomy diminishes the utility of these neuroimaging modalities. For example, in cases where the long-term trajectory of user benefits and associated risk factors remains unclear to stakeholders. There are implanted alternatives that use less extensive surgery, such as Stentrode, a new type of neuroimaging introduced by Synchron^TM (Oxley et al., 2016). This system does not require craniotomy. Instead, it is computer-guided, with a minimally invasive procedure used to guide the Stentrode into a blood vessel within the brain. Studies have found that the signals are comparable to their counterparts with more extensive surgery, such as ECoG and MEA. (Mitchell et al. 2023) discovered Stentrode to be a safe implantable endovascular BCI for people with severe paralysis. However, a limitation of Stentrode is that it records brain signals from only the major blood vessels; thus, a question remains if this technology is scalable to any brain area. Sub-scalp EEG has recently become another viable option for recording finer resolution signals than EEG and requires less maintenance than ECoG and MEA (Duun-Henriksen et al., 2020). Generally, signal quality tends to improve with increasing extensive surgery in current neuroimaging modalities of BCI (Figure 3). For example, the invasive Utah intracortical eletrode array and minimally invasive stentrode offer significantly superior signal quality as compared to EEG and is suitable for diverse BCI-based ARTs (Mitchell et al., 2023; Grani et al., 2022; Oxley et al., 2016; Maynard et al., 1997).

The most advanced BCIs utilize the brain's electrical activities recorded through neuroimaging techniques such as EEG, ECoG and MEA (Edelman et al., 2024; Willett et al., 2023, 2021; Moses et al., 2021; Sponheim et al., 2021; Saha et al., 2021; Edelman et al., 2019; Oxley et al., 2016; Rao et al., 2014; LaFleur et al., 2013). Other state-of-the-art modalities include MEG to measure magnetic fields produced by the brain's electrical currents and fNIRS to record the brain's hemodynamic activities (Bu et al., 2023; Naseer and Hong, 2015). These methods output time-domain signals corresponding to the user's intentions or cognitive states. Traditionally, various digital signal processing techniques filter out undesired noises or artefacts from raw neural signals to enhance the detectability of user's intentions or cognitive states (Lotte et al., 2018; Krusienski et al., 2011; Bashashati et al., 2007; Lotte et al., 2007). Then, handcrafted feature extraction follows a classifier to evaluate the decoding accuracy of a BCI. However, with the recent developments in artificial intelligence-based algorithms, nonlinear activation functions replace handcrafted feature engineering for attributing more generic features associated with users' intentions or cognitive states (Barmpas et al., 2023; Roy et al., 2020; Fahimi et al., 2019). Convolutional neural network (CNN)-based architectures are popular among the state-of-the-art artificial intelligence techniques. The raw neural signals may require preprocessing to enhance signal quality and transformation to represent the signals as compatible with CNN or other architectures (Kamble et al., 2023a,b; Garćıa-Salinas et al., 2023). For example, raw neural signals are directly compatible as input to a 1-dimensional CNN or long short-term memory network architecture; however, the time-domain signals are incompatible as input to 2-dimensional CNN architectures. In this case, converting time-domain neural signals into 2-dimensional images through time-frequency analyses such as short-time Fourier transform and wavelet transform is typically performed.

Recent advancements in BCI technologies have significantly improved their performance. An implantable MEA has demonstrated an excellent 94% decoding accuracy for a 50-word classification task for a speech neuroprosthesis (Willett et al., 2023). Non-surgical EEG systems, while less precise, show around 60% average accuracy for decoding up to 6 words (Bhalerao and Pachori, 2025). A non-implantable EEG-based speller BCI can also provide high accuracy, such as 82% average accuracy, featuring visually evoked potential and few-shot learning (Miao et al., 2024). However, decoding using the non-surgical EEG-based speller paradigm is much slower than word decoding by implantable BCIs. The trade-off between speed and accuracy is important while selecting an appropriate BCI type for a user, by considering the disability and user-centric technical specifications fulfilling individualized user needs and preferences. Although surgical options offer superior performance to non-surgical BCIs, surgical options are highly individualized and tailored to demonstrate user-specific high decoding accuracy. On the contrary, EEG signals are more accessible to many users with advanced decoding algorithms, some of which are generalizable across users.

The time-domain neural signals fluctuate over time and across users due to diversity in anatomical, physiological, psychological and cognitive characteristics (Saha and Baumert, 2020). These inter-session and inter-subject variabilities cause differences in training and testing feature distributions, leading to covariate shifts. Covariate shift occurs when training and testing feature domains differ due to inherent variability in data characteristics, resulting in poor decoding accuracy (Saha and Baumert, 2020; Azab et al., 2019, 2018; Jayaram et al., 2016). For covariate shift adaptation, a BCI typically requires tedious calibration using data from a new target subject or the same subject on a new session. Studies demonstrated significant inter-session and inter-subject variabilities, even in healthy user cohorts (Saha et al., 2019, 2017b; Kang and Choi, 2014; Kang et al., 2009). The impact of these variabilities should be more prominent for users with disabilities due to the diversity in impairments post-neurological incidents (Park et al., 2016).

Studies proposed various transfer learning strategies in BCI decoding algorithms for covariate shift adaptation (Azab et al., 2019, 2018; Jayaram et al., 2016). Inter-session and inter-subject domain adaptation is an example of transfer learning in BCI. However, most transfer learning techniques require at least a few training data for decoding algorithm calibration while the goal is a fully zero-training BCI that promotes wide dissemination of this technology across diverse user groups. Inter-subject associative BCI is feasible requiring no training data from the target user when the training and testing subjects share similar brain dynamics (Saha et al., 2023, 2019, 2017a,b). Quantifying inter-subject associativity is a way to predict the performance of fully zero-training BCI. From a translational perspective, generalized decoding algorithms can disseminate BCI-based ARTs to a large cohort by minimizing the calibration requirements (Roy et al., 2020; Saha et al., 2017b). Even so, a large cohort of people (15%–30%) encounter BCI deficiency, which refers to BCI system's inability to interpret users' brain signals (Park and Jun, 2024; Bamdadian et al., 2015; Vidaurre and Blankertz, 2010).