Stroke is among the most common causes of disability worldwide and is the chronic neurological disease with the greatest impact at this level [1,2]. Among the multiple consequences of stroke, upper limb (UL) impairments are one of the most prevalent, with around 70% of stroke survivors showing altered arm function, 40% maintaining persistent functional disability [3], and only 5%-20% showing complete dexterity 6 months after stroke [4]. Motor control impairments result in altered movement dynamics, characterized by muscle synergies that reduce optimal motor performance and induce movement compensations [5,6].

Upper limb rehabilitation (ULR) after stroke is recognized as a complex challenge due to different factors, namely the multifactorial nature of the condition (eg, spasticity, decreased muscle strength, changes in somatosensation and perception) [7-9] and the nonlinear progression of disability over time. In addition, the complexity underlying ULR is also compounded by the need to define multiple parameters, including the type of intervention, dosage (such as duration, difficulty, and intensity), application format, and the underlying factorial paradigm, which involves different degrees of parameter interaction [10].

Multiple types of motor rehabilitation interventions have been studied throughout the years with the goal of reducing motor impairments and improving functional activities, making use of learning- and use-dependent mechanisms [11] to promote motor learning. Considering UL interventions, traditional or modified constraint-induced movement therapy, mental practice, robot-assisted movement therapy, and neuromuscular electrical stimulation are currently among the most recommended strategies [11-13]. Neuroplasticity is the structural or functional changes (or both) within neurons affecting their connectivity with each other that occur during learning after injury, being a key element of motor rehabilitation and motor acquisition [14]. Motor rehabilitation allows patients to learn,optimize and adapt their motor, sensory, and cognitive functioning through dosed, repetitive, goal-oriented, progressive, task- and context-specific training [11]. The combination of traditional therapy methods with technologies that enable the use of enriched environments, sensory stimulation, and task-specific training is a neuroplasticity promoter, enhancing functional recovery for stroke survivors [15].

Brain-computer interfaces (BCIs) have been increasingly studied over the last 20 years as a way of enhancing ULR by being a facilitator of intensive and repetitive practice [16]. BCI is a technology that allows the acquisition, analysis, and decoding of brain activity, translating it into communication and control commands for an interface [17].

Measuring and analyzing brain activity data is often done using electroencephalography (EEG) due to its noninvasive, portable, low-cost nature and high temporal resolution [16]. EEG enables the detection of various brain activity patterns, such as event-related potentials (eg, P300), steady-state visual evoked potentials, and event-related desynchronization (ERD) and event-related synchronization (ERS) that are associated with motor-related tasks [18]. Common paradigms for intervention using BCI are motor imagery (MI) and motor attempt, also associated with motor observation [7]. Studies using MI tasks frequently rely on ERD and ERS as a reliable method for detecting motor-related brain activity [19]. MI-based BCI operation is then associated with the ERD phenomena and the modulation of sensorimotor rhythms, linked to imagined limb movements. By providing feedback based on the patient’s brain activity—that is, neurofeedback—BCIs close the motor-sensory feedback loop during MI training, allowing users to actively perceive and modulate their brain activity [20]. Repeated practice combined with real-time feedback promotes neuroplasticity and facilitates functional recovery [21]. Kruse et al [22] conducted a systematic review and meta-analysis showing evidence that the addition of MI-based BCI to conventional therapy may enhance the improvement of UL function and brain function recovery. Although meta-analysis on different design characteristics of BCI interventions favors motor intention over mental practice [17,23], MI offers a unique opportunity for stroke survivors who are unable to move their extremities by attempting to stimulate the brain regions responsible for movement [22,24].

MI, also referred to as mental practice and recently action imagery [25], is described as training movements or tasks in an imagined way and without producing movement, with the specific aim of improving performance [26]. Recently, Bach et al [27] proposed a conceptualization of MI, not as purely motoric and based on the execution of the movement, but in its anticipated perceptual effects, suggesting motor imagination (or “effect imagery” as proposed by the authors) is driven by the desired perceptual outcomes of actions rather than the movements themselves.

MI application is extensively recommended as an adjunct to conventional therapy [28-32]. Neuroimaging findings provided evidence that MI and physical practice are functionally equivalent, that is, they recruit overlapping brain regions within the brain motor networks underlying motor preparation and execution, including premotor, parietal, primary somatosensory, and motor cortices [33,34]. Both MI and planning and execution of movement involve the activation of the cerebellum, putamen, inferior frontal gyrus, and supplementary motor area [35]. Thus, the relevance of MI practice lies in reinforcing the motor schemas involved in actual movement through the mental visualization of motor actions [26]. MI can involve explicit or implicit tasks, where the first involves active and conscious imagination of a particular movement, and the second subconsciously induced processes for task-imagery problem solving, such as mental rotation of an image [31].

The use of neurofeedback associated with BCI seems to promote changes not only at the neurophysiological level but also at the clinical level [36]. This can be provided through different modalities, such as visual, auditory, and tactile feedback [37]. Virtual reality (VR) is one of the feedback possibilities to combine with BCI [7,36], which allows users to experience being represented by an avatar body rendered through a 3D environment and interact with it in a realistic way [38]. Multiple meta-analyses have described the positive effects of using VR in ULR, enhancing not only motor function but also activities of daily living (ADLs) [39,40], and practice guidelines recommend its usage to improve stroke rehabilitation outcomes [30]. Immersive VR, when used as a complement to conventional therapy, can provide a suitable environment for practicing the repetitive and specific tasks needed to improve motor skills after stroke [41].

Starting from the premise that one of the main goals of neurorehabilitation is to promote neuroplasticity and restore function, MI, VR, and BCI form an interconnected triad that mutually reinforce one another, representing a powerful approach to achieve this objective. The effectiveness of this approach relies on several interdependent dimensions that are dynamically interconnected and mutually influential, such as: (1) the user’s ability to elicit and modulate neural oscillations associated with sensorimotor processes during MI [42]; (2) the system’s capacity to accurately detect, process, and classify these signals [19]; and (3) the user’s motivation, engagement, and adherence throughout the training process [43]. Each component of the intervention triad influences one or more of these dimensions, creating a reinforcing cycle that amplifies the overall effectiveness of MI-VR-BCI interventions.

MI-based BCIs typically have lower transfer rates compared to other BCI paradigms [44], and achieving effective control requires prolonged training protocols involving simultaneous learning by both the user and the system. While the user must draw on and further develop existing MI skills, the system simultaneously enhances classification accuracy through optimized algorithms and machine learning techniques [19]. However, a considerable proportion of users (approximately 15%-30%) [45] might face difficulties in effectively controlling BCI systems using MI alone, a phenomenon known as BCI inefficiency. Despite this, some authors argue that BCI inefficiency may not solely stem from biological limitations but may also be influenced by inadequate training protocols, likely due to MI being cognitively demanding, unfamiliar to many users, and often associated with fatigue or boredom during prolonged training sessions [46]. VR can counteract these challenges by increasing engagement and facilitating more vivid and effective MI. For example, immersive VR environments have been shown to elicit stronger ERD, facilitate sensorimotor rhythm modulation, and improve BCI performance [47,48]. A key mechanism underlying these effects may be embodiment—the sense of owning and controlling the avatar —which arises from the interplay between body ownership, agency, and self-location [49], with Forster et al [50] also considering the close intertwining with presence. In addition, VR supports action observation, potentially activating the mirror neuron system, both of which are known contributors to neuroplasticity [51]. Closing the loop by providing feedback during MI observation enhances neuroplasticity by engaging pre- and postsynaptic summation mechanisms, thereby promoting motor learning [21]. Furthermore, VR can enhance motor learning, improve body awareness, and elicit kinesthetic sensations even in the absence of physical movement [49,52]. Clinically, this is highly relevant for stroke survivors, who often struggle with body perception and a reduced sense of control over their affected limbs.

VR is also able to leverage certain principles of motor learning, like the specificity of the task, intensity, repetition, and relevance, necessary factors for the development and strengthening of synapses and, consequently, for motor recovery [53,54]. Another aspect that favors the association of MI, BCI, and VR is the multisensory stimulation potential associated with the latter. This combination has the potential to enhance hand function by delivering targeted sensory feedback across different systems, such as somatosensory, visual, and auditory pathways, thereby promoting neuroplasticity [55]. This feedback may help reduce the discrepancy between motor intention and MI commonly observed in BCI applications, potentially increasing the overall effectiveness of the MI-VR-BCI approach. Although it presents multiple theoretical benefits, this paradigm remains recent and innovative, raising important questions regarding its design and implementation.

BCI-based technologies have a broad range of applications and are increasingly being researched, but there remains a gap between the technology and the end user [56]. The development of BCI systems for rehabilitation must consider patients’ needs while encompassing clinical applicability and technical feasibility [57]. Research states that practitioners and researchers need to involve users in creating solutions that consider factors such as convenience, ease of use, privacy, security, safety and quality standards, prioritizing users, and ergonomics [58]. Given these considerations, it is necessary to establish recommendations for the use of MI-VR-BCI and adjust the technology to meet the real needs of users and clinical practice [59]. To address the gaps in existing evidence, this study aimed to explore the insights of (1) neurorehabilitation experts regarding the most effective factors and principles of neurorehabilitation and neurophysiology relevant to adapting for BCI intervention; (2) stroke survivors to inform recommendations based on their needs, preferences, and motivation; and (3) biomedical engineers on the technical knowledge of the technology.

Using a collaborative methodology, the study sought to develop realistic recommendations for designing more effective interventions for the ULR in stroke survivors using BCI-based interventions combined with VR and MI.

A qualitative design grounded in reflexive thematic analysis [60,61] was used within a constructionist epistemology and interpretivist theoretical perspective to draw from the tacit knowledge of experts and experience from stroke survivors through a reflexive thematic analysis. Thus, a multiperspective collaborative design process was used through workshops, including clinical neurorehabilitation experts, stroke survivors, and biomedical engineers [62]. For this study, both web-based and in-person workshops were planned. The first allowed the participation of geographically dispersed participants [63], and the second helped fully leverage interpersonal dynamics for enhanced collaboration and permitted experiencing the MI-VR paradigm. Technology experimentation was made through NeuRow developed by Vourvopoulos et al [64], a first-person perspective upper-limb MI training paradigm involving a bimanual boat-rowing task, with feedback delivered through a head-mounted-display VR and haptic feedback (Figure 1). Logistical and time constraints, particularly the extensive preparation and calibration required for the EEG, limited opportunities for technology experimentation including direct BCI control, only involving MI and VR. The COREQ (Consolidated Criteria for Reporting Qualitative Research) checklist was followed for this study (Multimedia Appendix 1) [65].

Workshop planning was based on the literature review conducted to familiarize with the topic and identify the evidence gaps. This search focused on meta-analyses in the population with stroke, including the MeSH (Medical Subject Headings) terms “stroke” AND “rehabilitation OR intervention OR therapy” and “upper extremity,” published in the last 5 years, and guidelines for rehabilitation and in meta-analysis on “Brain Computer Interface,” “Virtual Reality,” and “Motor Imagery.”

The recruitment took place between January 2023 and July 2024, using a purposive sample. Participants were selected based on the researcher’s judgment [66], through the national stroke patient association’s email list, and the researchers’ professional networks. Considering the study goals, the priority in participant selection was the richness of contributions and good communication skills.

Neurorehabilitation experts were included if they had at least 10 years of experience with stroke rehabilitation. Stroke survivors were included if they (1) had a stroke-caused UL impairment, (2) were more than 18 years old at the time of the stroke and (3) had participated in at least 20 poststroke rehabilitation sessions with preferential previous technological experience either in therapy or daily life. Potential participants with stroke were excluded if they had a diagnosis of cognitive impairment or impaired communication (eg, aphasia) that limited their ability to understand complex instructions and to report their experience with stroke rehabilitation. The inclusion criteria for engineers were based on experience with BCI-based interventions and knowledge of the technical details.

Following an initial informal contact to gauge the willingness to participate, potential participants were sent an invitation letter, a consent form, a demographic characterization questionnaire, and scheduling options.

Of the 25 neurorehabilitation experts who were contacted, 21 expressed interest in participating; however, 8 were not able to participate due to scheduling difficulties. Among the 23 eligible stroke survivors interested in taking part, 6 could not attend due to scheduling issues. Of the 4 engineers who accepted, 1 was unable to attend due to illness. A total of 33 participants were recruited, including 13 neurorehabilitation experts, 17 stroke survivors, and 3 biomedical engineers, and sociodemographic information was collected (Tables 1-3; questionnaires for stroke survivor characterization in the Multimedia Appendix 1).

A total of 6 workshops were conducted. Of them, 3 with neurorehabilitation experts (NEs), 2 with stroke survivors (SSs), and 1 with the participation of neurorehabilitation experts, stroke survivors, and biomedical engineers (BEs). 4 workshops were held remotely, and 2 workshops were conducted in person (workshops 5 and 6). Workshop durations ranged from 90 to 145 minutes. Each session followed a format with four key phases: (1) introduction outlining the workshop’s objectives, structure, and rules, relevant theoretical background, and summary of prior findings; (2) participant introductions and icebreaker activity; (3) core activities tailored to each workshop’s objectives; and (4) completion of a workshop standardized feedback questionnaire.

For each workshop, specific goals were established, and corresponding methodologies were selected based on the recommendation of Ozkaynak et al [62] and Benson et al [67]. Accordingly, multisensory elicitation activities and complementary methods were used to enhance participant engagement, leverage diverse strengths, and maximize workshop productivity and efficacy. The methodologies used included scenarios, task prototyping, World Café, and technology experimentation, implemented either in plenary formats (workshops 1-4) or, where appropriate, also in small groups (workshops 5 and 6)—involving all different stakeholders in the case of the final workshop. The workshops were planned iteratively, with the outcomes of each session informing the next to address any gaps, clarify inconclusive points, or resolve disagreements. Figure 2 presents an overview of the workshops conducted, and a more detailed description is available at Table S1 (Multimedia Appendix 1).

To ensure consistency and comparability across sessions, the research team carefully predefined and pretested the structure and focus of each workshop while also ensuring researcher training. While tailored to each workshop’s specific goals and participant profiles, visual prompts and guiding questions were carefully designed and refined in advance by the research team. Summaries of each workshop were presented in the subsequent session to promote continuity. For workshops 5 and 6, a written summary was also shared with participants for validation. After each workshop, participants had the opportunity to provide feedback on their experience through a form with both closed- and open-ended questions and share their thoughts on the most important topics discussed, suggest any missing topics, and leave any other comments (Workshop Assessment Questionnaire in Multimedia Appendix 1). The analysis of the feedback questionnaires was used to plan the subsequent workshop.

An inductive thematic analysis was conducted using data from the verbatim transcripts [68]. The 6-phase process of thematic analysis was followed: data familiarization, systematic data coding, generating initial themes, developing themes, refining and naming themes, and producing the report [60].

The workshops were initially transcribed using the Microsoft Office Word transcribe feature, followed by manual review. Each transcript was scanned for relevant quotes, which were lifted into a codebook to group them into preliminary themes according to the study’s goals.

The preliminary themes were combined and reassembled into different themes and subthemes until the final thematic map was achieved. The analysis was validated by a second author and discussed with a third research member, followed by frequent debriefing sessions within the research team.

The research team was comprised of 4 physiotherapists and 1 biomedical engineer, with clinical experience and research experience in the field of stroke rehabilitation. They believe in stroke survivors’ potential to improve, the importance of their active role in their rehabilitation, and the importance of neurorehabilitation principles knowledge to support intervention. Although the research team considers the combination of MI-VR-BCI, a promising tool for upper limb stroke rehabilitation, they view it as a complement rather than a replacement for conventional rehabilitation, as supported by high-quality scientific literature. While this is the researchers’ perspective, a reflexive approach was used throughout the study’s conduct and data analysis, with continuous attention to positionality and potential bias. Data interpretation was conducted with transparency and critical awareness, and findings were presented in an open, reflective, and analytically rigorous manner.

Throughout the study, there was an introspective concern to maintain focus on the research question and relevant framework while promoting an environment of free participation by the participants. For the first 4 workshops, there was little to no prior contact and personal relationship with the participants. In the fifth and sixth workshops, a professional association existed between the researcher conducting the sessions and 6 patients, as well as 3 health care professionals, due to their shared involvement at the rehabilitation center. These pre-existing professional relationships did not influence the design, methodology, or conduct of the workshops.

To ensure the quality of the study, meetings were held for discussion of the methodology and analysis processes. This step reduced possible bias from the researchers and increased scientific rigor.

The study was approved by the Ethics Committee of the Polytechnic Institute of Setúbal (CE-IPS 35/2023) and Alcoitão Rehabilitation Medicine Center (CMRA2023/006 and CMRA 2023_012). All participants were informed of the aims of the study, procedure, risks, and confidentiality through an invitation letter. Prior to taking part in this study, the participants signed the informed consent form. Participants did not receive any form of financial or material compensation. Participants were also informed that workshops would be recorded, their identity would be kept confidential, and that they had the right to withdraw from this project at any time without any prejudice.

The thematic analysis revealed 6 themes (Table S2 in Multimedia Appendix 1) regarding recommendations for planning and implementing MI-VR-BCI interventions and were arranged according to 3 dimensions: user profile (themes 1 and 2), intervention planning (themes 3 and 4), and technological development (themes 5 and 6), with all dimensions closely intertwined. The user profile dimension integrates the importance of a patient-centered approach and considerations on clinical evaluation and patient selection. The intervention planning dimension contemplates the recommendations for task design and guidelines for structuring BCI intervention. The technological development dimension considers the key factors influencing motivation and specific recommendations on technological features. A map was created to provide a comprehensive summary of the findings (Figure 3). Identification of the participant’s group (NE, SS, and BE) and number is used in quotations, followed by the identification of the workshop.

By undertaking a multiperspective qualitative study, this study aimed to explore the insights from neurorehabilitation experts, stroke survivors, and biomedical engineers. From the analysis of the results, 6 themes emerged with key recommendations for the design and implementation of a MI-based VR-BCI tool for UL poststroke rehabilitation, which can significantly enhance its effectiveness. By integrating the perspectives of diverse stakeholders, the study informs the development of MI-VR-BCI tools that are not only technically sound but also user-friendly and tailored to the needs of stroke survivors. The individualization of treatment throughout the intervention process, associated with multidisciplinary teamwork and shared decision-making, emerged from the study as important, representing also essential pillars of the patient-centered approach [69]. It is a recommendation in line with international guidelines for UL stroke rehabilitation and provision of health care [11,29,30,70].

This interdisciplinary collaboration has the role of providing comprehensive health care and establishing treatment goals and planning in a shared manner, having in mind the patient’s functionality [11,70]. Goal establishment must consider that individual needs vary, influenced by stroke type, symptoms, and the individual’s cultural and psychosocial circumstances, and must be well-documented, specific, and challenging [70]. While specialized counseling and orientation by a professional when using VR is seen as important [71], the autonomy support (providing individuals the ability to control their own behavior) also enhances UL performance and promotes the use of the more affected UL after stroke [72].

According to recommendations for stroke recovery trials, a holistic, biopsychological assessment must be performed [73,74], encompassing the 3 dimensions of the International Classification System (ICF) [75], namely body functions and structure (impairment), activities (limitations), and participation (restriction). According to international stroke assessment guidelines, the Fugl-Meyer Assessment, which evaluates impairment, should be complemented by the Action Research Arm Test to assess the activity dimension [74]. Integrating instrumental evaluations, such as inertial motion sensors, can further enhance the accuracy of UL impairment assessment [76] and provide complementary insights into patient-reported outcomes related to activity and participation [11]. The assessment of the quality of movement performed (another important aspect in the context of investigation of UL-oriented interventions) might be achieved through performance assays (such as 2D reaching, finger individuation, and grip and pinch strength) and a 3D functional drinking task [77]. These recommendations shall be taken into consideration since, despite them, most research primarily addresses structural factors, with only a portion, including activity [73], a pattern also found in BCI research.

Following the recommendations from our participants, it is essential to establish clear criteria for effectively assessing MI capability [28,29], and neuropsychological aspects like attention and concentration [22]. Considering cognitive functions, Kruger et al [78] emphasize the importance of working memory capacity for MI, as it depends on the ability to retrieve stored sensory information about a specific movement from long-term memory, transform it, and maintain it within working memory. There was a divergence of opinion among participants concerning the inclusion of patients with cognitive impairments, specifically those affecting attention, perception, and mental imagery capacity. In an observational study on influencing factors on MI after stroke, visual-spatial skills were correlated with MI visualization quality, indicating that altered visual-motor skills may impact both MI and movement [26]. Rienzo et al [34] found similar evidence but stated that, according to the evidence found, visuo-spatial and perception impairments seem to have more influence on MI tasks of an implicit and not explicit nature, since implicit MI involves manipulation of mental representations of affected body parts. In agreement, attention and spatial abilities were found to be performance predictors for MI-based BCI, together with the users’ relationship with the technology (such as computer anxiety and sense of agency) [79].

Considering neglect, potentially leading to perceptual impairments and visuomotor deficits [80], there is evidence showing that not only contralesional UL visuomotor MI tasks reduced some neglect symptoms (manifested through copying and drawing tasks) but also enhanced UL sensation [81]. These clinical traits are more commonly associated with right hemisphere lesions, including parietal lobe and fronto-parietal pathways [80], and parietal stroke survivors are often excluded from MI studies [82]. Even so, this lesion location alone, or the presence of neglect, does not necessarily predict MI capacity, as neuroplasticity may allow other brain regions to compensate for the damaged areas [34,83]. These findings should be carefully considered, as factors such as neglect may not justify excluding stroke survivors from studies or clinically implementing MI-VR-BCI that involves explicit MI, and patients can possibly benefit from the intervention. However, these factors may influence the intervention outcomes and design, including the MI training required and the selection of task difficulty.

UL functionality should also be considered in decision-making regarding training protocol, as it directly impacts motor execution that correlates with MI capacity [26]. Initial severe UL impairment (Fugl-Meyer Assessment-Upper Limb ≤30), significant hand spasticity (Modified Ashworth Scale-Hand ≥level I+), and presence of aphasia have been associated with less favorable distal UL recovery after MI-based BCI intervention [84]. Stroke location may also play a role regarding intervention potential, as supratentorial strokes and lesions in the pontine region are associated with poorer UL recovery (with the latter linked to severe disturbances in both sensory and motor functions) [85]. Basal ganglia lesions are found to particularly disrupt frontal lobe connectivity and motor planning [85], possibly negatively influencing rehabilitation outcomes. As corticospinal tract integrity is a strong biomarker of sensorimotor function and motor recovery [86], it may be considered for establishing patient profiles and intervention conditions, with higher corticospinal tract integrity and some degree of residual muscle control possibly particularly benefiting from combined EEG and electromyography systems, enhancing BCI performance [87]. Interestingly, although sensorimotor function may be a predictor for success in intervention, in cases of severe compromise of the corticospinal tract, MI may be particularly beneficial by optimizing effector-independent learning mechanisms, relying on additional projection systems (ie, the corticoreticular tract) [88], and enhancing connectivity between the sensorimotor network and other brain networks [89]. To tailor clinical decisions regarding rehabilitation options and enhance prediction of therapy response, it will be important to assess white matter disconnection patterns (disconnectome), besides clinical presentation and stroke location [88]. In addition, stroke stage must also be considered for decision-making on optimal training protocol. While effective in both subacute and chronic phases, BCI shows greater effect sizes in the subacute stage, and efficacy may change according to BCI feature combination (for a meta-analysis [90]). Chronic phases with smaller effects might require adjustment of intervention dosage, with Salvalaggio et al [91] suggesting that effective UL interventions in the chronic phase must be delivered in a minimum of 10 hours if augmentative therapies are used, while priming interventions require higher doses (between 10 and 30 hours).

When considering task selection and execution, rehabilitation principles to augment motor control and restore sensorimotor function recommend the training to be task- and context-specific, goal-oriented, repetitive, meaningful, and function-based [29,30]. Accordingly, participants in our study mentioned the importance of the activities being based on ADLs as they represent more meaningful and useful activities when transferred to real needs. These results are in line with those collected by Isbel et al [71] in a study of a similar nature, with participants stating that the use of daily encountered activities is not only more relatable but also more motivating. The recommendation of context-specific training is also a perspective that VR can leverage by allowing the representation of different scenarios and an adequate selection according to stroke survivors’ needs. Although the MI-VR-BCI paradigm does not involve motor practice, as it shares principles with scientifically supported task-practice interventions [29,30,92], its principles may be useful for task design [93].

This recommendation is significant because, although current VR development guidelines consider the importance of task relevance [94], a meta-analysis conducted [7] on the design characteristics of BCI paradigms revealed a gap regarding the tasks performed. Most common BCI tasks involved motor attempts or imagination of grasping and extending the fingers, multijoint movements, or individual finger movements, with only a few including ADLs. Aprigio et al [35] provide an overview of tasks used in conventional MI, noting that the most commonly performed activities are rooted in daily life functions. Interestingly, these align closely with the activities our participants identified as significant. Thus, it is crucial to steer BCI task development toward activities that mirror daily life functions.

When considering the practice of bilateral or unilateral activities, the effects of each activity modality remain unclear [8,30,32,95], and evidence shows that bilateral activities appear to have more impact on function-related outcome measures and unilateral activities have more impact on activity dimension [95-97]. In line with the participants’ recommendations, clinical guidelines [29] recommend that both bilateral and unilateral arm training must be considered, taking into account the task and level of impairment.

Regarding intervention structuring, and in line with the recommendation to grade activity difficulty to match the stroke survivor’s abilities, Proulx et al [55] emphasize the importance of tailoring tasks to individual progress. The authors advocate for adaptable difficulty levels and appropriate task demands, noting that intense active practice requires continuous attention and focus from the stroke survivor. Keeping task demands appropriately challenging requires continuous adaptation and progression [30], which in turn supports both motivation and motor learning [54] and may facilitate the experience of flow state during MI-based BCI tasks [98]. Accordingly, Di Rienzo et al [99] found that low-demand tasks can induce fatigue and boredom, reinforcing the need to optimize task difficulty. Poveda-García et al [26] suggest aligning task difficulty with motor function, as motor function correlates with MI capacity. To support this individualized balance, BCI systems may incorporate performance accommodation mechanisms (PAMs), which dynamically adjust task difficulty based on the user’s performance, helping to maintain engagement and motivation. Jochumsen et al [100] compared various PAMs, including augmented success, mitigated failure, and input override. The authors identified that users with low BCI control experienced greater perceived control when assisted by PAMs, while those with high BCI control reported decreased perceived control and reduced performance when assisted. Accordingly, selection and implementation of PAMs should consider user performance in real time but account for individual differences in BCI control.

While task complexification was identified by the participants as a key method for progression, using EEG to capture brain activity poses challenges due to its high temporal resolution but poor spatial resolution [101]. Although BCI systems can record activity from cortical and subcortical networks, the detail and extent of the captured information largely depend on the signal acquisition method. MI activity is most effectively detected in the premotor cortex, primary motor cortex, and supplementary motor area, which are also activated alongside the basal ganglia and thalamus [102]. However, because EEG is a noninvasive method, it can only reliably capture signals from cortical areas, whereas intracortical electrodes are required to access activity from the basal ganglia and thalamus [16]. The increasing complexity of tasks, combined with the varying accuracy of MI, may not be fully captured by the recorded signals. This raises concerns that imagining an activity inaccurately could strengthen neural circuits that hinder effective motor learning. These considerations highlight the critical importance of accurately measuring brain activity and ensuring that mental imagery of tasks is closely aligned with the intended motor objectives.

As briefly mentioned by our participants, using high-density mapping could increase spatial resolution but result in a difficult signal-to-noise ratio across channels [103]. The challenges associated with data acquisition, processing, and pattern recognition in MI-based BCI interventions are increasingly being studied in order to overcome the challenges associated with using EEG and improving accuracy [104]. Saha et al [16] theoretically explore the addition of other source-capturing signals like functional near-infrared spectroscopy or magnetoencephalography in a complementary way to EEG to boost classification performance. The recommendation to create more complex scenarios by increasing cognitive demands (eg, by introducing additional visual or auditory stimuli) aligns with previous recommendations [55] and offers an alternative means of raising task difficulty. This method emphasizes cognitive mechanisms rather than relying solely on motor challenges, also recruiting dual-task mechanisms, as stated by participants.

Using connectivity as a coadjuvant progression criterion to assess MI efficacy aligns with findings from the study by Bagarinao et al [105], which indicate that MI expertise leads to more targeted recruitment of motor networks and higher brain activation intensity. Regarding the integration of MI-VR-BCI into clinical practice, participants stated it should complement conventional therapy, which is in line with the findings of Kruse et al [22] with MI acting as a priming tool with positive influence on impairment and activity outcomes in patients with stroke [106].

Motivation levels directly influence the process of ULR, posing as a fundamental aspect of intervention planning [107]. While repetition is essential for neuroplasticity and learning, requiring substantial practice volume and intensity, it can present challenges for sustaining patient engagement and motivation. In line with our recommendations, participants from Isbel et al [71] also stated that motivation can be enhanced through realism, gamification, scoring, and feedback. The usage of variability, besides having a positive impact on motivation, is a principle associated with optimized motor learning [108]. Moreover, gamification is a strategy with proven results in rehabilitation, associated with higher levels of motivation [109]. Its addition to VR shows improvements in the overall effect in comparison with only VR [36]. Gamification has been increasingly explored as a strategy to optimize MI-based BCI training protocols and outcomes. By enhancing motivation and engagement and reducing the monotony often associated, gamification may improve BCI performance (average classification accuracy of 74.35%) and support adherence and sustained participation in the intervention [110]. A recent systematic review [111] explored gamification in MI-based BCI training, highlighting the positive impact of elements, such as feedback, avatars, assistance, and social interaction, on both performance and user experience. However, most studies lacked rigorous designs, often missing control groups or relying on small samples, making it difficult to draw robust conclusions about the impact of individual game elements on effectiveness.

Regarding feedback customization recommendations, Proulx et al [55] state that treatment design for individuals with stroke must consider feedback delivery parameters, task complexity, and sensory deficits. They argue for combining various modalities of augmented feedback tailored to each individual condition, rather than a one-size-fits-all approach. Accordingly, the addition of multiple types of feedback may enhance performance and motivation and minimize frustration in healthy participants [112]. However, in patients with stroke, symptoms such as sensory hypersensitivity, including hyperesthesia, or potential cybersickness must be carefully considered, and optimal conditions should be established, ensuring adequate levels of immersion and presence to optimize the benefits of VR. When considering multimodal stimulation, feedback adjustment, and instruction modality, it is important to recognize that the relevance of different feedback types may vary depending on the nature of the task, its significance to the intended action, and the distinct brain areas and cortical networks they activate [78].

In the absence of proprioceptive feedback, visual feedback becomes particularly salient, making it a critical factor for MI-VR-BCI designs. Consistent with the statements of participants in our study, who expressed a preference for avatars with humanoid characteristics, several studies have explored the benefits of anthropomorphic design [44,47]. Through an experimental paradigm involving healthy participants, Alimardani et al [47] identified that using humanlike features raises the sense of agency, ownership, and embodiment. These findings are complemented by Bashford and Mehring [113], who showed that observing movements congruent with the imagined task can enhance the sense of control over the rendered body—even in the absence of accurate, real-time BCI performance. Although MI-based BCI performance may not differ between humanoid and abstract avatars, humanoid design can significantly increase users’ sense of ownership and agency [44]. By reproducing a more faithful representation of the imagined movement and aligning with the users’ real-world sensorimotor expectations, higher excitation of motor processes might be induced, which, in turn, may promote more efficient connectivity and aid the learning process [47]. Considering MI as not only motoric but also motivated by multisensorial movement conception and predictions [114,115], BCI researchers and designers might leverage multisensorial nature and VR environments to elicit a MI response.

Feedback in BCI systems is a multidimensional construct that encompasses more than just modality (eg, visual, haptic, or proprioceptive cues). One critical dimension is feedback granularity, which refers to the level of detail and continuity of feedback over time, encompassing aspects of spatial and temporal congruency, which are closely interrelated [116]. In discrete feedback, the system responds only when MI activity surpasses a set threshold, providing a brief, event-based signal. In contrast, continuous feedback offers real-time information about user performance, with the participants of Kjeldsen et al [117] showing a preference over discrete feedback due to a better sense of agency and movement notion. Despite this, in the same study, it was only possible to establish a correlation between higher agency and performance in the discrete feedback condition. Another key dimension is feedback bias, typically described as positive, negative, or neutral. Positively biased feedback, which rewards users regardless of classification accuracy, can temporarily enhance embodiment and MI performance [118]. Conversely, negative feedback within embodied VR, presenting movement with the contralateral side of the intended imagery during unsuccessful attempts, may impair the sense of agency and hinder MI learning [47]. When comparing contingent feedback with independent feedback through BCI-functional electrical stimulation (FES) in patients with chronic stroke, MI-contingent feedback led to significantly greater improvements in wrist extensor strength, range of motion, and functional connectivity in the affected hemisphere [20]. In addition, Maier et al [108] state that rhythmic cues, explicit feedback (knowledge of results), and implicit feedback (knowledge of performance) are neurorehabilitation principles to consider when developing interventions to elicit motor learning in neurorehabilitation. Importantly, the effectiveness of feedback strategies appears to be performance-dependent, underscoring the need for personalized feedback planning that considers both clinical characteristics and individual BCI proficiency.

A key difference between motor execution and MI is that the latter lacks sensory feedback due to the absence of movement [34]. Motor execution–based BCI interventions have demonstrated higher efficacy in UL function rehabilitation [17,23]. Adding proprioceptive feedback through externally facilitated movement can enhance neural engagement and promote more effective motor learning and recovery. When considering the addition of proprioceptive feedback, meta-analyses comparing BCI intervention designs favor the use of FES for improving UL function [17,23]. Usage of neuromuscular electrostimulation is also recommended by stroke rehabilitation guidelines [11], particularly FES to reduce motor disability and improve function in both acute and chronic phases of ULR [30].

Considering technology implementation, several key steps are essential. To support smoother adoption and informed engagement, both health care professionals and patients must receive adequate education about the MI-VR-BCI system [119]. It will also be important to establish what the neural mechanisms underlying BCI effectiveness for stroke rehabilitation are [120]. In addition, standardizing protocols and assessment timings and using validated clinical scales are recommended to ensure consistency in tracking progress, enhance decision-making, and improve comparability across patients and studies [11]. Based on participants’ input, establishing normative data from healthy individuals may further support clinical interpretation by providing benchmarks for MI quality, engagement, and neurofeedback responsiveness. In addition, when considering BCIs in clinical practice, clear ethical guidelines must be established and followed to ensure proper data management, privacy, and security [121].

This study addresses important gaps in the MI-VR-BCI field by shifting the focus from technical performance in controlled research settings with healthy participants to user-centered, clinically meaningful design. Grounded in the lived experiences and expertise of stroke survivors, rehabilitation professionals, and engineers, our aim is to bridge the gap between empirical findings—often based predominantly on healthy participants —and their clinical applicability. Our approach inverts the traditional trajectory for BCI development, from system- to patient-centered, prioritizing design and outcomes that align with clinical effectiveness and established international rehabilitation guidelines. In doing so, this study raises awareness of stroke-specific symptoms often overlooked in BCI development—such as sensory deficits, altered body perception, attentional challenges, and fatigue—and how these symptoms critically influence both user engagement and system performance. While we contribute to identifying profiles of eligible patients with stroke, we also propose multidimensional guidelines for the development of MI-VR-BCI interventions and adaptation strategies for those with more complex clinical presentations. Our recommendations may help accelerate the iterative design of MI-VR-BCI interventions, paving the way for inclusive, individualized, and effective systems that address a wider spectrum of clinical needs and are truly applicable to stroke rehabilitation settings.

Ultimately, this work may lay the foundation for a more unified framework that integrates clinical profiles, BCI performance, and user experience to optimize training protocols and enhance both learning and meaningful rehabilitation outcomes.

A list of recommendations was created for the development and implementation of MI-VR-BCI systems (Textbox 1), grounded on the results of the workshops and informed by international guidelines for poststroke rehabilitation.

Although the sample’s multiperspective nature is a strength, participants’ limited experience with the technology, due to its innovative nature and lack of clinical implementation, represents a limitation. Despite efforts to address this through rich and elucidative activities, our study design and methodologies did not allow for full experimentation of the MI-VR-BCI paradigm, due to the logistical constraints and design methods used. Findings may be context-specific and influenced by individual participant experiences and knowledge. This reality suggests the need to validate the recommendations, using, for example, a nominal group technique. In that case, we suggest a balanced inclusion of diverse clinical neurorehabilitation experts and stakeholder representatives to capture a comprehensive and balanced range of perspectives.

Many of the recommendations suggested were based on a theoretical foundation and construct validity. Thus, it will be important to develop efficacy studies that can adequately isolate influencing variables and mechanisms of action, also helping to establish the validity of the recommendations through empirical paradigms.

Despite promising advances, current MI-VR-BCI technologies face several limitations that hinder their clinical translation. These include low signal-to-noise ratios and variability in EEG signal quality [122], challenges in accurately classifying brain signals and supporting multiple classes [123], lengthy and demanding calibration protocols [110,124], limited real-time adaptability [125], and reliance on fixed protocols that do not allow for progression [79].

To overcome these limitations, future research should focus on developing adaptive, patient-centered MI-VR-BCI systems that are aligned with international clinical guidelines and real-world clinical needs. This includes optimizing training protocols in terms of duration, progression, and cognitive load management to sustain engagement and motivation. Equally important is raising awareness among, and involving, health care professionals and patients early and continuously in the design and implementation process [119], to ensure that technologies meet real-world clinical needs, encourage adoption, and promote meaningful rehabilitation outcomes [126].

Future studies should analyze the interaction between different types and intensities of feedback (eg, immersive and nonimmersive VR and their effects on embodiment) and their relationship with task characteristics (eg, unilateral and bilateral activities), user profiles, and BCI performance. Considering stroke survivors’ characteristics, it will be important to stratify data according to stage of evolution, lesion location and network disconnection, level of impairment of somatosensory function, and perceptual and neuropsychological alterations. It is also important to establish the efficacy of gamifying training protocols and understand its relationship with motivation, engagement, and BCI performance. Studies should also establish the intervention’s effectiveness across ICF dimensions, both short- and long-term, and assess the cost-benefit ratio for different intervention features.

Technological solutions should evolve to better mimic real-life conditions and address stroke survivors’ clinical needs, balancing personalization with broader applicability.

The MI-VR-BCI paradigm was very well received by participants in our study. The recommendations presented in this study for the development and implementation of MI-VR-BCI technologies aim to enhance their impact on ULR. Grounded in comprehensive consultation with all stakeholders, these recommendations will improve the technology’s relevance and potential efficacy. These recommendations should be taken into account in future studies in order to allow the analysis of the relationship between different characteristics of the stroke survivors, the MI task, and the technological features.