Visual evoked potentials (VEPs) rely on the conscious recognition of a target in the user’s visual field. VEPs depend on external stimuli, which generally consist of lights flickering with different temporal profiles. These profiles fall into five broad categories: 1) frequency modulation (f-VEP), 2) time modulation (t-VEP), 3) code modulation (c-VEP), 4) phase modulation (p-VEP), and 5) motion-onset (m-VEP). In all of these cases, each stimulus carries a unique temporal pattern according to the modulation scheme that can also be detected in the scalp EEG to indicate the selected stimulus. For example, in the case of t-VEPs and c-VEPs, different stimuli are presented at varying delays from trial onset, which can then be detected in the time-domain EEG signal [57], [58], [59]. However, due to noisy time-domain signals, these VEP variations require a high number of trial averages for robust results. By contrast, f-VEP and p-VEP stimuli flicker at different frequencies and/or phases. Attending to a particular target increases the band power of the corresponding oscillation in electrodes covering the occipital cortex. Thus, f-VEPs and p-VEPs are collectively also known as steady-state VEPs (SSVEPs). In practice, SSVEP-based BCIs provide some of the highest information transfer rates as the combination of frequency and phase modulation can be decoded with high accuracy [60]. This combination also now allows for the successful detection of >100 target options [61], [62], which can be further expanded in hybrid systems when complementary signal types are utilized [63]. Finally, the m-VEP stimuli consist of moving lines across a virtual keyboard to induce visual motion-based event-related potentials [64] to form a BCI speller. In particular, the N200 component of these evoked signals is considered to be primarily related to motion-onset visual processing. These fast and high-dimensional BCIs work well with a variety of applications, including spellers and, less often, spatial navigation, where individual commands need to be correctly identified during relatively narrow time windows.

Overt spatial attention (OSA) tasks leverage activity patterns generated in the parietal cortex in response to the user’s spatial attention. Similar to motor imagery (MI) tasks, OSA signals do not require external stimuli and generate focal patterns of alpha band activity. This makes OSA tasks well suited to complement other endogenous signals, such as those generated from performing MI tasks [65], [66], by increasing control dimensions without sacrificing overall performance of the BCI. While this relatively new BCI paradigm shows promise for higher dimensional control, users and experimenters must still consider the interaction effects of overt gaze control and other BCI-related tasks [67], [68].

The P300 is an endogenous event-related potential (ERP) within the EEG, and is detected in electrodes covering the parietal cortex. P300 signals occur in the context of an oddball paradigm [69], and rely on a user’s implicit ability to distinguish a rarely presented target stimulus from other more common non-target stimuli. This structure makes such an event-related response useful for spelling applications where a specific letter must be chosen from a larger set of irrelevant letters. Most P300-based BCIs use the visual row/column paradigm, in which a matrix (e.g., 6 × 6 cells or variable) containing the alphabet, numbers, and other items is presented to the user for selection [70], [71]. A P300 response is then elicited when the rows and columns of the matrix, flashing in random order, converge on the desired item being attended to. The P300 was documented well before BCI use and has long been implicated as a biomarker of various clinical disorders, including depression [72] and schizophrenia [73]. While these signals can be gaze- and space-dependent, advances in stimulus design such as variations of the rapid serial visual presentation approach have significantly improved information transfer rates as well as task reliability and performance [74], [75].

Other broadband time-domain strategies that are favored for the identification of both cued and continuous mental states involve slow cortical potentials, otherwise known as movement-related cortical potentials (MRCPs). When closely examining voluntary movement, changes in low-frequency activity (delta band, ~1 – 4 Hz) occur in the movement planning and execution stages and are time- and phase-locked to movement onset [76], [77]. After averaging across numerous trials, a so-called MRCP becomes visible; the main components are the “Bereitschaftspotential” or readiness potential, the motor potential (negative peak), and the positive rebound potential. It has been shown in the literature that such an MRCP occurs after different voluntary actions and can be derived from sensorimotor areas. Over the last decade, MRCP features have been heavily exploited for decoding movement detection [78], [79], [80], [81], [82], [83], [84], but also for the classification of movement-related parameters such as speed [85] and force [79].

While the success rate of the MRCP-based motor control is modest compared to some other signal types, the burden of imagining different types of movement is dramatically reduced. For example, during the MoreGrasp project [86], study participants described a low burden during the calibration and use of the BCI, suggesting that a successful transfer of MRCPs towards natural control can be achieved. Nevertheless, the false positive and success rates can still be improved, which is a matter of ongoing and future work (see [87] for a comprehensive review).

Event-related desynchronization (ERD) or event-related synchronization (ERS) refers to a band power decrease or increase, respectively, relative to a reference period [88], [89], [90]. These phenomena are usually induced in response to a motor task and produce a reliable contralateral ERD and ipsilateral ERS in the alpha/mu (8–13 Hz) and/or beta (14–30 Hz) bands of scalp EEG. Thus, projecting the different ERD/ERS traces into a time-frequency plot for each electrode can further generate so-called ERD maps [91] to identify the spatial, temporal, and spectral characteristics of these signals. These features are based on a traditional neuro-structural model and have been robustly exploited using a simple task paradigm illustrated in Fig. 3 [92].

Fig. S1 provides an overview of the typical ERD/ERS patterns that are generated during various movement conditions for both able-bodied and participants with spinal cord injury (SCI). Here, we differentiate these conditions into: a) active, voluntary movement; b) passive movement and movement induced by functional electrical stimulation (FES); c) complex finger movement and peripheral stimulation; d) continuous finger, hand/arm movements; and e) motor imagery. We also indicate the same conditions for people with SCI, however, due to physical limitations of these individuals, there is no counterpart to c), and f) describes attempted movement instead of real movement.

The most classical ERD/ERS patterns appear in able-bodied participants when performing voluntary hand movements (Fig. S1a, h) [90], [93]. These contexts often produce a mu/beta ERD before and during movement, as well as a post-movement beta ERS (PMBS), also called the beta rebound. Active foot movement (Fig. S1a, f) can also induce a clear beta ERD followed by a PMBS [94], [95], [96]. These oscillatory activities are often also accompanied by an MRCP [97], [98].

During passive hand movements (Fig. S1b, h), there is no pre-movement ERD, however, a mu/beta ERD is present during hand movement (either completely passive, or facilitated by FES), followed by a PMBS [93]. A beta ERD and PMBS are also observed during passive dorsiflexion of both feet (Fig. S1b, f) [95]. A similar pattern occurs in the case of complex finger movement (Fig. S1c, h; cube manipulation) and additional median nerve stimulation. These tasks also induce a contralateral mu/beta ERD; where ipsilateral beta ERS returns quickly to baseline, the mu ERD/ERS oscillates [99]. These tasks also produce sensory evoked potentials that are much larger than the rest condition [100].

In addition to changes in oscillatory activity, movement trajectories (Fig. S1d, h) can be decoded from delta band activity [101]. For example, in [102], a mu ERD and characteristic delta band activity was observed when participants performed rhythmic finger movements (flexion, extension, repetitively). Along these lines, low-frequency time domain (LFTD) signals have been shown to encode hand/arm kinematic movement [103], [104], [105]. Importantly, for these and other works [106], there also exists a mu and beta ERD during movement as well as PMBS after the termination of movement.

Very early after the investigation of executed movements in EEG research, imagined movement or motor imagery (MI) (Fig. S1e, h) played a major role in BCI development. Mu and beta ERD occur during MI [90], [107], and are often followed by a PMBS [96], [108], [109] (Fig. S1e, f). These ERD/ERS signals are often focal phenomena that arise according to the topographic organization of the motor homunculus, similar to what was previously observed for motor execution tasks. For example, multimodal work using EEG source imaging and fMRI has revealed strong co-localization of sensorimotor cortex activation for hand and foot movement execution and imagery tasks [110], [111]. Similarly, common patterns of BOLD elevation observed in fMRI have also been observed using EEG source imaging [112] in the form of ERD in the sensorimotor cortex. Based on these findings, MI tasks involving body parts with spatially separated motor cortex representations often define the number of control signals that can be used. Many online BCI paradigms include four or more MI tasks of different body parts to control an end effector in multiple dimensions. Common tasks include imagining the action of the right hand, left hand, both hands, rest, foot, and tongue.

Studies have shown that the speed of hand movement or imagined hand movement is also correlated to EEG activity [113]. Many of these tasks can be easily detected and separated by a standard set of electrodes covering sensorimotor areas [114], [115], however, including subject-specific features from larger groups of electrodes can improve both offline and online performance [116], [117], [118]. Furthermore, successful discrimination of more specialized MI tasks of the same body part (i.e., hand gestures, limb joints) has shown promise in offline scenarios and may, in the future, be integrated into online control as well [119], [120], [121], [122].

In people with SCI, MRCPs are apparent for different movement attempts (Fig. S1f, h) as well as a beta ERD and, in some cases, a beta ERS [94], [123]. In a study on foot movement attempts, only a slight ERS was visible, which indicates that the PMBS is not solely triggered by afferent input but also by internal processes [95]. Passive foot movement in SCI patients (Fig. S1g, f) exhibits no visible patterns, which makes sense since there is no afferent feedback to the brain in these individuals [95]. In incomplete SCI patients, however, there may exist slight ERD/ERS activity.

Overall, when examining the existence of ERD/ERS patterns in people with SCI, several concerns need to be considered. First, attempting continuous hand/arm movement can lead to activity in sensorimotor and centro-parietal areas (hand-eye coordination). For example, in Pulferer et al. [124], delta band activity could be used for kinematic decoding in a SCI patient. Furthermore, a mu and beta ERD can also be observed in SCI patients performing hand and feet MI (Fig. S1i, f,h) [125], [126], [127], [128], [129], [130]. Finally, these patterns vary widely across patients and depend on the severity of their injury, the time since injury [94], [131] and also whether the individual had previous BCI training experience.

While some non-invasive BCI signals can be inherently produced, others must be learned throughout an often long training period. Some users never learn to successfully control a BCI despite lengthy training experience, which is often referred to as “BCI illiteracy” for a specific BCI paradigm. Thus, it is important to identify human factors that cause illiteracy and formulate strategies that promote BCI skill acquisition and optimal performance. One popular theory within the field posits that user attention plays a key role in the success of user training and achieving BCI control. This is often portrayed in terms of the Yerkes-Dodson law [133], which describes an inverted U-shape relationship between a user’s engagement and habit/skill formation. In this sense, tasks or contexts that are too simple can cause boredom, and those that are too difficult can cause stress, both of which negatively impact performance (Fig. 4). To address this, researchers have developed novel or adaptive tasks that appropriately engage subjects and improve BCI training and performance [134].

Along similar lines, many complementary technologies have been employed in BCI research to increase user engagement and avoid reduced performance due to boredom. Traditional BCI tasks that rely on endogenous signals utilize simple visual feedback paradigms involving a computer cursor, however, recent trends show an increased usage of virtual reality (VR) [135] and anthropomorphic feedback using computerized arms [10]. Previous cognitive studies have found that virtual hands and arms can elicit a strong sense of embodiment that is use-dependent [136], and suggest that these tools may also benefit BCI training. In fact, VR feedback with MI tasks has been commonly explored for motor recovery training in stroke patients with overall positive, but still varying levels of success [10], [14], [137]. These studies often focus on motor function rather than BCI performance and must accommodate highly variable brain damage, making it difficult to draw conclusions regarding the efficacy of VR aids.

Nevertheless, the use of VR tools with healthy individuals has shown much more promise by providing immersive contexts that enhance the embodiment of virtual limbs. Many of these studies have shown an increased amplitude of the respective ERD/ERS signals and a further improvement in online BCI performance compared to classic computer cursor feedback [138], [139], [140]. Using VR is currently a relatively low-cost addition to BCI systems; however, its full occupancy of the visual channel makes this technology somewhat impractical for future integration into daily life. By contrast, augmented reality (AR) is more compatible with everyday tasks and has been shown to be compatible with at least SSVEP-based BCIs [141], [142]. However, these contexts can create crowded visual scenes that must be considered so as not to overwhelm and distract the user [143].

At the other end of the spectrum, anxiety has been shown to impart a significant negative bias on BCI learning and performance for both MI and SSVEP-based BCIs [144], indicating that the self-regulation of mental state can lessen these effects. This factor has recently been creatively addressed through the use of meditation practice that can help alleviate elevated anxiety levels. Exemplary meditation-based strategies such as mind-body awareness training (MBAT) and mindfulness-based stress reduction (MBSR) training [145], [146], [147] have improved MI-based BCI performance. Overall, this suggests that acute and chronic stress levels impact BCI performance. Therefore, the ability to control a BCI may scale with the amount of stress-reduction training experience and the improved ability to produce stronger EEG signals associated with MI tasks.

Modern brain-based navigation of external devices includes the control of computer cursors, quadcopters/drones, wheelchairs, and robotic arms. Computer cursor paradigms remain robust testbeds for exploring new decoding algorithms, tasks/signal types, and more. Currently, non-invasive cursor control can be performed in two [107] and three dimensions using MI tasks [116], [148], [149] or hybrid systems that utilize MI and OSA [66]. More sophisticated navigational control has also been achieved using virtual helicopters or drones [150], [151]. However, this application comes with additional complexity as continuous flight often requires an asynchronous control strategy with ongoing commands/tasks rather than discrete trials, such as the center-out paradigm. BCI-controlled continuous flight of a physical quadcopter in three-dimensional space was first demonstrated using ERD/ERS signals corresponding to MI tasks [152]. This feat was achieved using right- and left-hand MI for one dimension of control, and MI of both hands and rest for the second [152]. More recently, a combination of SSVEP and MI tasks was also used to control various aspects of quadcopter flight [153].

Another important application of BCI is the brain-based control of a wheelchair [154], [155]. Such a task allows BCI technology and the use of various brain signals, such as P300 or SSVEP responses, in addition to MI tasks, to assist impaired patients in situations resembling daily life [156], [157], [158]. A recent study demonstrated that tetraplegic patients with SCI can operate a self-paced wheelchair during complex navigation tasks using an MI protocol [155].

In recent years, multiple non-invasive BCI studies have focused on robotic arm control. Controlling robotics is far more complex than the previously described devices, as the increased physical skill required for grasping tasks requires a corresponding mental strategy. Initial demonstrations of robotic arm control via scalp EEG focused on endpoint velocity or position control. However, as more detailed control of reach-and-grasp tasks practical for daily assistance became desired, this work branched into additional threads that focus on individual task segments. These segments can be broadly categorized into navigational robotic device control (“reach”) and dexterous hand commands (“grasp”). Along these lines, the reaching aspect was initially demonstrated in three-dimensional movement decoding of center-out and motor execution tasks [159] and has since greatly expanded [160], [161], [162].

Other than movement execution, MI paradigms have been pursued using ERD/ERS signals to control a robotic’s arm reach, grasp, and movement. He and colleagues have developed a unique and comprehensive framework for reach and grasp using MI. Along these lines, Meng et al. first demonstrated online non-invasive BCI control of a robotic arm in a group of human subjects, in whom MI tasks were used to reach for and grasp a Lego block (Fig. 5) [163]. Eight subjects completed all tasks and demonstrated well over 80% accuracy for completing reach tasks requiring two-dimensional navigation. Similar performance was achieved for more complex sequences of tasks in three-dimensional space involving grasping blocks off of a shelf, reaching over 70% accuracy. Edelman et al. further demonstrated the continuous reaching of a robotic arm through a combination of source-based decoding and a continuous pursuit control paradigm in a group of six human subjects [164], [165]. This work first demonstrated that the extended target-chasing task significantly improved MI-based BCI learning compared to traditional discrete trial protocols by increasing user engagement throughout training. By integrating real-time electrical source imaging, BCI performance was further enhanced across the skill spectrum in naive and experienced users. These works represent major advancements in the reach and grasp of BCI robotic arm control. When combined, this framework may, in the future, enable individuals with neuromuscular dysfunctions to autonomously perform daily tasks using only MI-based brain signals.

Robotic arm control has also been pursued from decoding movement execution. As such, to decode control initiation, Pereira et al. [166] developed a paradigm for online detection of self-initiated movements in a realistic scenario. EEG signals were time-locked to saccade onset, with participants shifting their gaze during goal-directed reach-and-grasp tasks. This strategy enabled relevant feature extraction and gaze incorporation with careful artifact handling [167]. A hierarchical classification approach utilizing LFTD features achieved a 54% true positive rate (compared to a chance level of 12%). This work expands possibilities for realistic settings, incorporating motor and visual processing in real-world tasks. Kobler et al. [103] further designed an experiment that involved center-out and continuous goal-directed movements, and considered two conditions (observation, execution) to study different volitional states. In either condition, the participants fixated on a target stimulus with their gaze in a 2D workspace. During motor execution trials, participants controlled a cursor by moving their right arm (on a flat surface) in a pursuit tracking task (PTT) to investigate the tuning characteristics of position and velocity. These temporal tuning characteristics indicated that neural activity preceded cursor velocity by approximately 150 ms [103]. These observations from EEG are in agreement with the tuning characteristics of spiking activity of M1 neurons [168].

In a follow-up study, Mondini et al. [105] investigated the feasibility of continuously decoding voluntary hand/arm movement trajectories from EEG to achieve closed-loop online control of a robotic arm. This work also implemented a PTT where the participants were asked to track a moving object on the screen by controlling a robotic arm. Kobler et al. [104] then suggested that integrating information about non-directional kinematics (e.g., distance, speed) in the decoding model can alleviate the amplitude mismatch problem. This non-directional kinematic information has been found in both ECoG [169] and MEG activity [170]. Provided that distance and speed are nonlinearly related to position and velocity, the previously introduced partial least squares Kalman filtering (PLSKF) approach [105] was extended to an unscented Kalman Filter (PLSUKF). Kobler et al. [104] re-analyzed EEG [103] and MEG data [170] to evaluate the performance of the PLSUKF with respect to both the PLS and the PLSKF during both observed and executed movements. Indeed, the correlation between the executed and the decoded trajectories was stronger for PLSUKF compared to other algorithms. Further, with the integration of non-directional kinematics in the decoding model, the amplitude mismatch between recorded and decoded trajectories could be reduced, overcoming the limitations of previous work [105]. Follow-up studies by Müller-Putz et al. and Pulferer et al. investigated the viability of decoding attempted movement, mimicking the limited motor function of SCI patients (by strapping each participant’s dominant arm to a chair) [86], [171]. They examined potential learning effects that may arise when a user performs the same motor control task multiple times [171]. Within five days, ten able-bodied participants performed three sessions. The selected timeframe aimed to allow participants to recover from the mental strain of various tasks while ensuring they maintained a distinct memory of previous sessions. After calibration, EEG decoding was gradually included in online feedback until it reached 100% EEG control. Overall, this longitudinal study revealed learning effects across the various days of training, indicating a positive effect of training on decoding performance.

While extensive effort has been dedicated to improving the navigation aspect of robotic arm control, considerable work has also focused on incorporating EEG-based grasping signals into a BCI. First attempts to restore hand movements in paralyzed persons were purely pragmatic in that BCIs were directly connected to an orthosis or the patients’ arm using FES [127], [128], [129]. These works demonstrated the possibility of connecting a BCI with closed-loop FES, recording EEG from the scalp, and stimulating a body part. However, they remained proof-of-concept studies. Study subjects were able to switch through pre-defined grasping patterns, but BCI control remained rather artificial in that they used the imagination of different limbs (foot and right hand) [129], or only the left hand [126] to trigger the different neuroprosthesis steps.

Nevertheless, in recognition of the various complexities involved with grasping objects of different sizes, shapes and orientations, further studies have since investigated how relevant grasping actions and features are encoded in scalp EEG. In particular, the use of MRCPs has helped support the hypothesis of natural control of a neuroprosthetic device. MRCPs vary across different types of movements or grasps, which may be useful as a natural control signal for a BCI-based neuroprosthetic. In fact, studies with able-bodied participants have shown that different movement types (right arm: hand open, hand close, pronation, supination, elbow flexion, elbow extension), grasp actions (i.e., palmar, pincer, lateral, rotation), grasp laterality (i.e., unimanual vs bimanual) and grasp force (i.e., power vs precision) can be successfully decoded from scalp EEG using MRCPs [172], [173], [174], [175], [176], [177], [178], [179]. Importantly, similar differences across multiple grasp types at the single-trial level could also be observed in individuals with SCI [123].

While these studies show promise for neurorobotic control, they were primarily employed in offline classification settings. Translating these offline studies into online performance greatly increases the potential utility in patient-oriented BCIs. Working towards this goal, some studies that integrate decoding paradigms that simulate online control, have shown successful above-chance performance for the execution of three different grasp types (i.e., palmar, lateral, wrist supination) [180]. More recent work has successfully moved to complete online control of four grasp types (i.e., cylindrical, spherical, lateral, pincer) using MI tasks [181].

This demonstration indicates that dexterous grasping can be performed in real-time using scalp EEG, however, it is difficult to suppress reaching signals during the grasp portion of the task, and vice versa. In this case, mode switching would be greatly beneficial in informing the BCI whether the user is focusing on positioning the arm or grasping an object. Similar to the previously described movement onset detection, various BCI “switches” have been proposed, but have also yet to be seamlessly integrated into online control paradigms [182]. Therefore, an easier solution to robotic arm control for daily tasks that has been widely adopted by the field is to utilize shared control strategies [183], [184], [185], [186]. In this sense, shared control refers to partial autonomous control by the user and partial automatic task recognition and execution by the BCI [187]. In these cases, such a strategy allows the user to control the BCI in up to three dimensions using a variety of MI and SSVEP signals (often hybrid) to select a pre-established task/action that is then automatically carried out. Nevertheless, the degree and complexity of shared control vary. Some systems utilize custom image processing pipelines for object identification using off-the-shelf camera components such as the now-discontinued Xbox Kinect. However, this approach can be limited to objects with similar shapes consistently viewed in the same orientation [184], [185], [188]. Other strategies utilize depth cameras and more advanced pose estimation methods [189] or simple QR codes [186] to identify objects that may be approached from different angles or perspectives, potentially offering a more versatile solution.

To ensure effectiveness and safety, the user and machine must consider practical considerations when utilizing a real-life robotic arm. Recent work on robotic arm control utilizes SSVEP signals for at least one control dimension. However, the visual attention required for these signals can be distracting when integrating these systems into everyday life. Similar issues have been observed when controlling a robotic arm using MI tasks – while the visual channel is reserved for observing the robotic arm rather than the flickering stimuli, impaired sightlines can still cause a slight drop in performance [165]. It has also been shown that other cognitive distractions accompany the user when controlling an interactive device, which can reduce its effectiveness [190], [191]. Another practical consideration for robotic device control that has been recently addressed is the need to impose physical constraints such as obstacle avoidance or trajectory correction [192], [193] to ensure the user’s and others’ safety. Controlling a robotic arm is a highly complex cognitive process, and various precautions and aids should be added to BCIs to maximize their effectiveness.

Non-invasive BCIs have emerged as a transformative technology with significant clinical applications, and carry immense potential for improving the quality of life for individuals with various neurological conditions and disabilities. Using BCI technology for clinical applications can complement traditional strategies for disease diagnosis, treatment, and rehabilitation. For example, BCIs facilitate the early detection of EEG biomarkers for epilepsy monitoring and intervention. In addition, BCIs interact directly with the brain, enabling precise targeting of neural pathways. This can lead to more accurate and effective treatments of specific bodily functions and provide immediate feedback based on brain activity. Finally, the interactive nature of BCIs can be more engaging and motivating for patients than repetitive physical exercises.

BCI technology has two primary clinical applications: 1) assistive BCI and 2) rehabilitative BCI [194]. Assistive BCI seeks to bypass damaged neural pathways, providing a continuous and permanent means of communication and control of external devices. Assistive BCI systems are designed for individuals with severe motor impairments. These systems use SSVEP, P300, or MI signals to control various external devices, such as spellers [195], computer cursors [196], robotic devices [197], or speech synthesizers [198], thereby improving the user’s quality of life and independence. Rehabilitative BCI aims to recover damaged neural connections, restoring impaired functional abilities by effectively facilitating neuroplasticity. This approach offers comprehensive and personalized functional, cognitive, and affective rehabilitation therapies. Restoring motor function is one of the most promising clinical applications of BCIs for individuals with SCI, stroke, or neurodegenerative diseases such as ALS. Since conditions like stroke do not affect the ability to form motor intentions, BCIs can decode neural signals related to these activities. This allows for neurofeedback training through assistive devices such as robotic systems, virtual reality, and functional electrical stimulation. Despite the significant attention BCI-based rehabilitation has received for motor function, applications in cognitive training remain underexplored. Evidence suggests that BCI-based neurofeedback training can improve cognitive functions in conditions like attention-related hyperactivity disorder [199] and mild cognitive impairment [200], and may similarly benefit post-stroke cognitive impairments [201]. These findings highlight the potential of neurofeedback-based cognitive training to induce neuroplastic changes essential for rehabilitation. BCI-based interventions have also demonstrated significant potential in emotion and mood regulation, offering solutions to address critical challenges such as depression. By leveraging real-time neural feedback, these interventions enable individuals to consciously modulate their brain activity and promote more positive emotional states. Advanced BCI systems can detect and respond to specific neural patterns associated with mood disorders, providing personalized therapeutic feedback that encourages neural reorganization and emotional resilience. Such interventions not only complement traditional treatments but also offer a non-invasive, self-regulated approach to managing and mitigating depressive symptoms, thus opening new avenues for mental health care. We believe the future of BCI for clinical applications lies in integrating functional, cognitive, and affective rehabilitation into a holistic approach. This comprehensive strategy enhances motor recovery and supports cognitive and emotional health, paving the way for more robust and inclusive rehabilitation solutions.

Compared to BCIs targeted at healthy users, BCIs designed for clinical applications still present several unique challenges. Patients usually show higher individual variability in neurological conditions. Patients may also experience higher levels of fatigue and cognitive load when using BCI systems. Furthermore, neurological impairments can affect the quality of neural signals that the BCI relies on. Many neurorehabilitation patients may have limited mobility or dexterity, making it difficult for them to use BCI hardware comfortably. Consequently, recent advancements propose the use of advanced decoding methods [202], [203], [204], [205] and the integration of other medical interventions, such as transcranial direct current stimulation (tDCS), transcranial magnetic stimulation (TMS), and transcranial focused ultrasound stimulation (tFUS) [10], [26], [27], [206], [207], [208] to develop more robust, calibration-free, and user-friendly BCIs for clinical applications.

EEG decoding involves developing algorithms and models to analyze recorded EEG signals, with the aim of understanding and interpreting the underlying neural processes. EEG decoding is crucial for BCIs because it enables the translation of neural signals into actionable commands, realizing communication and control between the human brain and external devices. Brain signal decoding is always a challenging part of BCI, especially for non-invasive EEG. As EEG signals travel from their source in the brain to the scalp surface through the low-conductive skull, they encounter various instances of noise caused by both biological and non-biological factors.

While EEG signal decoding can be viewed as a specific type of classification problem, this topic presents unique complexities. In typical classification tasks, larger datasets enhance performance by providing richer information. However, in EEG classification, more data necessitates longer acquisition times, which can lead to participant fatigue and distraction. This degrades data quality, introduces noise and inconsistencies, and negatively affects model training. Therefore, proper EEG classification needs a balance between data quantity and quality. It is generally recommended to limit EEG recording sessions to 30–60 minutes, with regular breaks, to maintain high data quality. For intra-subject training, a model is trained and tested on data from the same subject. By contrast, inter-subject training involves training a model on data from multiple subjects and testing it on different subjects. This requires a large and diverse dataset to capture the variability in EEG signals across individuals, along with more complex operations to create a model capable of handling EEG data with different distributions.

The traditional classification pipeline in BCI typically encompasses an initial filtering process applied to EEG signals, followed by the classification of the filtered data. This filtering operation is designed to extract pertinent information from temporal, spectral, or spatial dimensions, depending on the specific BCI paradigms. However, traditional methods face many challenges when decoding EEG. First, the feature extraction and classification processes in traditional methods are optimized separately using distinct objective functions. This could result in a suboptimal solution for the decoding process. Second, traditional methods strongly rely on hand-crafted features, a manual selection process that can be time-consuming and subjective. Finally, while these approaches prove effective in a specific subject, their performance tends to decline when applied in subject-independent applications. Therefore, more advanced methods are explored to address these limitations.

In this section, we comprehensively explore recent progress in decoding methods for EEG-based BCI. We outline the principles underlying these methods, how these methods relate to one another, and provide guidance on their appropriate application. The methods discussed in this section are summarized in Fig. 6 and Table I.

Toolboxes relevant to BCI development provide functions that simplify the development of an individual BCI system. BCI toolboxes are usually developed on top of commercial high-level platforms (MATLAB, SIMULINK, or LabVIEW) or, more recently, using scripting languages such as Python.

MATLAB provides the basis for a number of BCI toolboxes for the presentation of behavioral paradigms (e.g., Psychophysics Toolbox [278]) and the post-hoc analysis and visualization of biosignals (e.g., EEGLab, [279]; BioSig [280]; Brainstorm [281]; FieldTrip [282]; and MNE [283]). EEGLab, BioSig and FieldTrip have been extended into toolboxes for rapidly prototyping and evaluating online BCIs, called BCILAB, rtsBCI and FieldTrip buffer, respectively.

SIMULINK provides the basis for g.BCIsys [284], a commercial SIMULINK-based toolbox that provides a front-end interface for rapid BCI research and development to users of g.tec hardware. LabVIEW has been the basis for several BCI demonstrations and has been used to implement the BCI toolbox Craniux [285]. The popularity of Python in the BCI community has increased over the past decade, and quite a few toolboxes have been implemented to facilitate BCI development. They include BciPy [286]; PyBCI [287]; MEDUSA [288]; Lab Streaming Layer [289]; pySPACE [290]; PsychoPy (for presentation of stimuli) [291]; and a set of toolboxes from the Berlin BCI group: Pyff – a framework for feedback applications and stimulus presentation [292]; Mushu – a toolbox for BCI signal acquisition [293]; and Wyrm – a BCI analysis toolbox [294].

Just like BCI toolboxes, BCI software platforms offer functions that facilitate the development of BCI systems. To do this, they contain support for different signal acquisition hardware, signal processing and visualization routines, and different types of user feedback. They also have complete implementations of different BCI approaches that have been validated in different contexts, contain ample documentation, are maintained consistently to adapt to variations in a device’s software interface or operating systems, and provide a whole ecosystem of tools to interact online and offline with other tools or programming languages.

BCI software platforms are usually implemented in low-level programming languages (e.g., C++), compiled into native machine code, and executed without dependencies on commercial libraries. Mature platforms also support economical development and deployment across many laboratories and users. While many general-purpose BCI software platforms have been proposed, e.g., BCI++ [295], xBCI [296], TOBI [297], MindDesktop [298], and UniPA BCI [299], only BCI2000 [300], [301] and OpenViBE [302] have been consistently developed for many years, and have enjoyed widespread adoption across many laboratories.

The initial and important choice for designing BCI software is whether to implement it from scratch using a particular programming language, or whether to proceed using the existing toolboxes or BCI software platforms described above in order to reduce the time, complexity, and cost associated with BCI software development [313].

Implementing BCI software from scratch can be appropriate (and is only practical) if the technical demands are relatively low, such as, for example, in a student project that aims to use a consumer BCI headset to create simple visualizations/feedback of a particular brain signal. It may also be the best choice for clinical BCI systems that need to pass regulatory approval and whose functionality is relatively fixed.

Using BCI toolboxes can be appropriate when the technical demands are somewhat more complex but when there is an expectation that BCI software design, data collection, and data analysis are performed by the same person. Example situations are the development of a BCI system to support one research study by an individual graduate student.

Using BCI software platforms is most appropriate when the technical demands are complex (such as the simultaneous use of different hardware), and when there is an expectation that over the course of several years, many different BCI systems are needed to support diverse experimentation executed by different people in potentially different locations.

The ultimate purpose of BCI software is to facilitate the implementation and testing of different BCI approaches. Thus, the success of BCI software in achieving this purpose can be measured by the number of scientific studies that they have supported. The illustration in Fig. S2 indicates the number of publications that have been supported by different BCI software packages. The benefit of the use of BCI software platforms is an improvement in practicality or cost associated with BCI system development [313].