This experimental study is part of the registered randomized clinical trial (ClinicalTrials.gov NCT07103122) and therefore uses an exploratory, single-arm pilot design focused on user experience and feasibility outcomes, without a control group. It was conducted in full compliance with the Declaration of Helsinki for research involving humans. The local ethical committee of the Fondazione Santa Lucia approved this research on December 2, 2019, (protocol CE/Prog. 790). Patients and caregivers were informed about the study procedures prior to enrollment into the study, and patients provided written informed consent.

Participant privacy and confidentiality were ensured in accordance with standard data protection procedures. All data were anonymized prior to analysis, stored on secure institutional servers, and access was restricted to authorized research personnel only. No financial compensation or reimbursement was provided to participants for their participation in the study.

Based on our pilot study [31], we decided to enroll only patients with chronic stroke, who typically exhibit greater clinical stability, allowing for clearer assessment of treatment effects without the fluctuations seen in acute or subacute patients. Therefore, patients who had experienced a stroke more than a year before the first evaluation were enrolled for this study.

This study was designed as an exploratory pilot focusing on feasibility and user-experience outcomes rather than hypothesis testing. Therefore, the aim was to estimate usability/workload and feasibility parameters rather than to test efficacy; thus, a priori power analysis was not performed. Sample size was determined to provide adequate precision of key continuous user experience measures and to inform the design of a future adequately powered trial.

Criteria for enrollment were (1) upper-limb deficits with sufficient level of muscle power such that movement is possible with gravity eliminated (Medical Research Council [MRC] grade 2) or against gravity (MRC grade 3), as assessed using the MRC scale (grades ≥1) [40]; (2) scores ranging from 18/66 (27%) to 54/66 (80%) of upper-limb function, as assessed using the upper-limb extremity section of the Fugl-Meyer Assessment (FMA) scale [40]; (3) absence of severe linguistic impairments that may limit understanding of instructions; and (4) absence of cognitive impairments, as assessed with the Addenbrooke’s Cognitive Examination-Revised version for the Italian population [41], visual deficit, or other neurologic disease in comorbidity that may affect the patient’s ability to interact with the VR environment.

Clinical motor scales were collected at baseline to characterize the sample and stratify participants by impairment level; however, they were not analyzed as outcome measures in this manuscript. This choice reflects the exploratory nature of the study and the limited sample size, which suggested promising trends, although further investigation is needed to confirm rehabilitation efficacy.

Since age-related [42] and severity-related [7,26] differences were found in motor outcomes following VR upper-limb treatments, we created a blocked randomization list with 2 levels of stratification (age and severity), which guaranteed a balanced within-group split throughout the data collection process [43]. Therefore, participants aged between 18 and 55 years were considered young participants according to previous studies [44-46], while those aged between 56 and 80 years were considered older. The level of impairment was computed using 2 FMA-Upper Extremity–based cutoffs selected for functional relevance rather than normative severity thresholds. Patients scoring between 18/66 (27%) and 36/66 (54%) were classified as having severe impairment of the motor performance. This range reflects a markedly reduced voluntary motor control, typically associated with limited functional use of the upper limb.

Conversely, scores between 37 (56%) and 54 (80%) were classified as moderate impairment, as an indicative range of partial motor recovery and greater residual motor capacity, allowing for more consistent execution of goal-directed movements. The distinction between these 2 categories was chosen to ensure a clinically meaningful separation between patients with profoundly compromised motor function and those with a moderate, yet functionally relevant, level of recovery.

Given the limited sample size within each impairment subgroup, stratification was not intended to enable powered statistical comparisons, but rather to ensure representation of multiple functional levels of motor impairment, thus supporting an exploratory evaluation of feasibility-related outcomes.

We used the sealed envelope tool, whose implementation is based on Pocock [44], to create the randomization list of participants as a function of the stratification number of randomization blocks (age and level of impairment), which revealed the need to enroll at least 12 patients. This sample size was also used in previous work [16] about the effectiveness of the VR system on the rehabilitation of the upper limb in patients with stroke.

Although the rehabilitation protocol and technological platform are shared with the previous study [31], this analysis focuses on a different set of outcome measures, enrolled patients, and research questions. Specifically, the data reported here concern user experience, participation, usability, and perceived workload, which were not the primary outcomes of the earlier publication.

A detailed description of the equipment is described in previous publications [31,33]. Briefly, the system consists of an HTC Vive VR System, including a head-mounted display, 2 motion trackers, 2 controllers, 2 base stations, which emit infrared light and synchronization signals to enable accurate spatial tracking of the head-mounted display, controllers, and trackers, and 2 Myo armbands (Thalmic Labs). Each tracker of HTC and each Myo armband were positioned respectively on the hand dorsum and on the forearm of each participant. The tracker system provides the position and orientation of the patient’s hand. The Myo armband, consisting of 8 sEMG sensors, measures the forearm electric muscle activity of the user on the skin surface. The movement of the virtual hands is directly inferred from the kinematic data captured by the trackers, whereas the motion of the fingers is estimated based on forearm muscle activity. This estimation relies on a model trained using a machine learning approach consisting of an iterative version of random Fourier features ridge regression [47]. The acquired signals (tracker kinematic and sEMG) are transmitted wirelessly via Bluetooth. All processing, which includes signal acquisition, filtering, machine learning model training, and prediction, is carried out on an off-the-shelf laptop (Dell Alienware 15) with a dedicated GPU (NVIDIA GeForce GTX 1070).

The rehabilitation software runs in Unity, a game engine platform commonly used for the creation of interactive 3D applications and VR environments. It consists of 3 different virtual environments in which a table in front of the patients matches the position and dimensions of a rehabilitation table. For each patient, an anonymous form is created and filled with the participant’s ID, date of birth, more affected side, and dominant hand. Before starting the training session, the program requires the acquisition of the positions of the table corners with 1 of the 2 controllers in use for the VR session for calibration, so that the virtual model of the table is spatially aligned with the corresponding physical table where the patient is performing the exercise. This was necessary to provide correct visuo-haptic feedback during the VR exposure.

After the clinical evaluation, each patient was enrolled for experimental training with the VR-MT. Before starting, a researcher explained the task to facilitate the patient’s interaction with the VR system. For all the patients, this was the first experience with a VR system in which myoelectric control must be exploited to achieve a task. The patient performed VR treatment while sitting on a chair with the possibility of resting their hands on a table. At the start of each session, the researcher carefully positioned the 2 Myo armbands over the brachioradialis muscles on each arm and attached the trackers to the back of the patient’s hands, aligning the z-axis between the index and middle fingers. As the system supports bimanual rehabilitation tasks by leveraging VR capabilities to provide mirrored assistance, assessment of proximal (limb movement) and distal control (hand gestures).

Prior to each session, to personalize the training exercises by calibrating the reachable workspace the patient was asked to (1) assume a resting pose, with a comfortable position of the hand on the table, (2) extend the less impaired arm to the maximum range in the anteroposterior and mediolateral directions, as shown in Figure 1, and (3) extend the less affected limb in the same directions with the more affected limb, as much as they could. For each arm, the software evaluates the reachable area.

Although the calibration procedure visually emphasizes the movements of the less affected limb, the reachable space of the more affected limb is estimated based on the patient’s attempted movements with the more affected arm during the calibration phase. Specifically, patients were instructed to extend, first the less affected, and then the more affected limb in the anteroposterior, vertical, and mediolateral directions, with the latter moving as far as their residual motor capabilities allowed. The system records the maximum spatial extent achieved by the more affected limb during these attempts, which is then used as an estimate of its reachable workspace.

A secondary assessment was dedicated to calibrating the visualization of the opening and closing of a virtual hand (gesture) according to the electromyography signals recorded by the Myo armband. The patient is asked to (1) assume a resting pose with a comfortable position of the hand on the table and minimal muscle activity, (2) reproduce the gesture of holding a ball with the fingers extended, and (3) reproduce the gesture of holding the handles of a concertina with finger flexed.

After acquiring these 3 independent recordings, the researcher verifies the quality of gesture recognition and decides if the neural network needs more training or not. Otherwise, to improve the gesture reconstruction, it is possible to add new gesture repetitions, discard all the recordings of a gesture, or discard them all and redo the training gesture from scratch.

The training comprised 24 sessions, each lasting about 20 minutes in a quiet room using a height-adjustable table and ergonomic seating to allow participants to perform reaching tasks in a physiologically appropriate posture, independently of body size. In the first session, the researcher encouraged the patient to visually explore the environment to facilitate familiarization with the VR scene, also avoiding possible distractions during the training. Then, the patient receives verbal instructions to perform the first tasks. The training involves 2 reaching virtual objects (a ball or a concertina) presented alternately to elicit different hand gestures. Specifically, interacting with the concertina requires performing a fist gesture (closing the fingers), while the ball requires a hand extension gesture (opening the fingers). Both the concertina and the ball feature 2 handles, positioned on opposite sides, to guide correct hand placement during the task, as depicted in Figure 2.

A reaching movement is considered successful when both virtual hands are positioned within 5 cm of the respective target handles. Real-time feedback on hand placement and gesture execution is provided within the VR environment. When both virtual hands are within the required tolerance, the handles change color from orange to green. Additionally, an on-screen gesture icon indicates the required hand posture and disappears once the correct gesture is performed. After both the position and gesture requirements are met, the target object gradually turns green, visually signaling the required holding time before task completion. Auditory feedback is also provided to enhance user awareness. A positive sound indicates successful task completion, while a negative sound is played if the trial exceeds the maximum allowed duration of 20 seconds. The position of each target varies in height and lateral deviation with respect to the initial resting position of the hands, which is recorded at the beginning of the session with the hands placed on the table. The maximum distance at which the target can appear is tailored to the maximum arm extension, while the minimum matches the resting position, both recorded by the patient prior to each session, for a total of 12 possible target positions. Targets were presented in a randomized order to prevent the repeated presentation of highly demanding trials, such as targets positioned near the limits of the participant’s reachable workspace or requiring more effortful hand gestures, which could otherwise induce premature fatigue. All training sessions were supervised by a physiotherapist, with specific attention to movement quality and prevention of compensatory strategies, such as excessive shoulder elevation or trunk displacement during reaching. Verbal feedback was provided to encourage controlled arm extension while maintaining proper posture. The patient is guided by the therapist to follow the instructions in the VR environment on how to place their hands inside the 2 spheres that appear next to the target object, with open hands in the case of the ball, with closed fists in the case of the accordion.

The first training session is dedicated to finding the right match among parameters according to the level of ability of the patient; therefore, a baseline level, which also serves as a familiarization phase to the experimental training. It consists of alternatively setting α and β parameters to 1 (independence by the less affected limb) and 0 (totally dependent). The baseline is intended to identify the optimal level of performance at the VR task for each patient; it must be difficult enough to create a challenging exercise but not excessively engaging, so that it can be sustainable for the duration of the training session without frustration. At the end of each block, consisting of 12 reaching movements, the therapist adjusts the proximal and distal agency and capability parameters based on the patient’s performance score. This score, calculated from the success rate within the block, guides the modulation of task difficulty for the subsequent block. When the patient achieves a success rate of >90%, corresponding to at least 11 successful trials, task difficulty is increased to promote greater motor engagement. If the success rate falls within 70%-90% (9-10 successful trials), task difficulty remains unchanged, ensuring the patient continues to perform within an optimal challenge level. Conversely, if the success rate is below 70% (≤8 successful trials), task difficulty is reduced to keep the task within the patient’s current capabilities and to prevent frustration or disengagement. This adaptive strategy was adopted to ensure that task performance remained within a sustainable range (ie, not falling below approximately 80%), thereby operationally defining feasibility as the ability to maintain engagement and performance over repeated sessions rather than complete task autonomy. Indeed, this framework is intended to sustain motivation and maximize motor improvement by continuously tailoring task difficulty to the patient’s performance, with the goal of maintaining a success rate within the target range of 70%-90%. The specific parameter to be adjusted, as well as the direction of adjustment, is determined by the therapist, who makes these decisions based on clinical expertise and real-time observation of the patient’s motor performance, fatigue, and overall condition. A complete description of how to set these 2 parameters to obtain virtual assistance is described in Multimedia Appendix 1.

A total of 3 assessments with clinical scales and questionnaires were provided. A motor baseline at T0 (enrollment) was performed by a physiotherapist, which evaluated the patient’s muscle power through the MRC classification, and the residual motor proficiency with the FMA scale. Assessment was further performed at the intermediate evaluation (T1, after the completion of 12 sessions), and at the end of the therapy (T2, after the completion of 24 sessions). If the patients were discharged or expressed a willingness to interrupt the trial, the T2 evaluation was anticipated. Since acceptability reflects participants’ subjective evaluation of the intervention after experiencing it, a meaningful baseline assessment at T0 was not considered conceptually appropriate. At T0, participants had not yet interacted with the VR system and therefore could not provide an informed evaluation of satisfaction or motivation related to the intervention. For this reason, changes in acceptability were evaluated between the intermediate (T1) and final (T2) assessments, when participants had already gained direct experience with the therapy.

Acceptability is a multifaceted construct that reflects the extent to which people delivering or receiving a health care intervention consider it to be appropriate [48]. Here, we assessed this construct according to the emotional responses the patients exhibited during the intervention using 2 visual analog scales for motivation and satisfaction. As in a previous study [16], patients rated their level of motivation to attend the therapy and the satisfaction of doing the therapy for each session on a visual analog scale ranging from 0 to 10 (with a higher score indicating a higher level of motivation/satisfaction). In addition, the therapist/researcher assessed patient adherence to therapy, reporting on a Likert scale, ranging from 1 to 6 (where 6 stands for the highest level), the patient’s participation in the exercise through the Pittsburgh Participation to Rehabilitation Scale [49].

Usability, intended as the users’ potential to achieve their goals with effectiveness, efficiency, and satisfaction in a specified context of use of technology [50], was assessed using specific tools. Precisely, the User Satisfaction Evaluation Questionnaire (USEQ) [51] is a questionnaire properly designed to evaluate the satisfaction of the user in virtual rehabilitation systems. It is composed of 6 items assessing (1) the user’s experience with a VR system with respect to the enjoyment associated with the experience, (2) the success rate in using it, (3) the perceived control of the system, (4) the clarity of information provided by the system, (5) the level of comfort experienced, (6) and the evaluation of its impact for the patient’s rehabilitation. We performed this assessment at the first (T0) and last (T2) sessions of each patient journey, to evaluate their first impression of the VR system versus the final considerations acquired after multiple uses of the system.

Finally, in the framework of the person-based approach, we assessed the feasibility of our VR intervention as a workload [52] associated with the experimental VR therapy. For this assessment, we used the NASA (National Aeronautics and Space Administration) Task Load Index (NASA-TLX) [53] for the multidimensional subjective assessment rates of perceived workload associated with the task. Patients compiled this scale at the end of each training session, making an evaluation of the task according to mental, physical, temporal, effort, and frustration demands. Each domain is scored on a straight line that goes from 0 to 10 points (with 10 as the highest perceived workload), and the patients mark the point that corresponds to their evaluation.

The statistical analysis was performed using IBM SPSS statistical software (version 23), while customized algorithms were implemented in MATLAB (version 2020b; MathWorks) environments for data visualization. Data are reported in terms of means and SDs for continuous measures, as medians and IQRs for ordinal measures, and percentages for reporting variables expressed as frequency. The Kolmogorov-Smirnov test revealed a deviation from a normal distribution of all variables, thus nonparametric tests were adopted for comparing performance and ratings across the different times of evaluation (T0, T1, and T2). The Wilcoxon signed-rank test for related samples was used to assess the between-group analysis of differences in the self-assessed evaluation of patients after T1 and T2 sessions (after 12 and 24 treatment sessions, respectively). As a post hoc analysis, to assess differences within group (age and impairment), the Mann-Whitney U test was performed. The Spearman coefficient was further used to evaluate effect size in post hoc analysis to further test the contribution of impairment to the self-assessed cognitive workload domains. Analyses involving impairment level were conducted with an exploratory intent, aimed at identifying descriptive trends and potential patterns relevant to feasibility and user experience, rather than at testing formal hypotheses between subgroups.

Longitudinal usability outcomes were analyzed using linear mixed-effects models to account for repeated measurements and incomplete observations across sessions. Separate models were fitted for motivation, satisfaction, and participation after normalization to their respective theoretical maxima (motivation and satisfaction: 10; participation: 6), resulting in bounded outcomes in the range 0-1.

For each outcome y_ij, measured for subject i at session j, the following model was specified:

where Session_c,ij denotes the session index centered around the sample mean, Group_i represents the impairment group, b_0i ~ N (0, σ²_b) is a subject-specific random intercept, and ε_ij ~ N (0, σ²) is the residual error term.

Models were estimated using maximum likelihood. Fixed effects were evaluated to characterize overall temporal trends, baseline group differences, and exploratory group-by-session interactions. Subject-specific random intercepts were included to account for within-subject correlation induced by repeated measures. Missing data were handled implicitly by the model under a missing-at-random assumption, without imputation. Given the limited sample size within impairment subgroups, all linear mixed-effects model analyses were considered exploratory and hypothesis-generating.

Only data from participants who completed a sufficient number of training sessions (≥10 sessions) were included in the usability analyses. This threshold was selected as a pragmatic criterion corresponding to approximately 3 weeks of repeated use, considered sufficient to move beyond initial familiarization effects and to obtain stable user experience measures, in line with usability research indicating that reliable usability estimates emerge after repeated interactions rather than during early exposure [54].

For all the tests, the significance was set at 0.05, except for post hoc analyses for which the α level of significance, according to Bonferroni correction, was divided by the number of possible comparisons (equal to 8), thus set at 0.006.

A total of 15 community-dwelling stroke survivors were initially recruited for the study. Of these, 12 community-dwelling stroke survivors (4/12, 33% females) aged between 27 and 72 (mean 52.9, SD 16.0) years completed a sufficient number of training sessions and were therefore included in the usability analyses. Three participants were excluded from the usability analyses due to insufficient exposure to the intervention, having completed fewer than 10 sessions. A detailed overview of participant flow is provided in the participant flow diagram (Figure 3).

Stroke chronicity was between 1 and 28 (median 2.9, IQR 1.8-8.4) years, and participants were classified as having moderate to severe motor impairment according to the FMA-Upper Extremity score. A total of 4/12 (33%) patients reported weakness following the stroke, on the right side of the body, and only 3/12 (25%) had a mild linguistic disorder (eg, anomia and dysarthria) that did not compromise either the patient’s ability to communicate or comprehend the verbal instructions given during the intervention sessions. Demographic and clinical information of all participants are presented in detail in Table 1.

All participants completed the whole experiment except for Sub09, who dropped out prior to the intermediate evaluation (T1); therefore, this participant was not included in the statistical analysis for T2 but was included for analysis at the other assessment points. Linear mixed-effects models showed consistently high levels of motivation, satisfaction, and participation across participants throughout the training period. Motivation exhibited a high baseline level and remained stable over sessions, with no significant effects of impairment group or session, and a nonsignificant trend toward a group-by-session interaction (group: P=.55; session: P=.31; group × session: P=.07). Satisfaction similarly showed high baseline values and remained stable over time, with no significant main effects or interactions (group: P=.72; session: P=.41; group × session: P=.86). In contrast, participation showed a significant positive effect of session, indicating a progressive increase over time, while no significant effects of impairment group or group-by-session interaction were observed (session: P=.01; group: P=.96; group × session: P=.37).

In line with the overall longitudinal trends, acceptability outcomes were further explored by comparing intermediate (T1) and final (T2) assessments to characterize clinically meaningful changes over time and across impairment levels. Comparisons of reported scores at T1 and T2 revealed a significant difference (P=.01, Wilcoxon signed-rank test) in the mean level of satisfaction declared by patients between the T1 (72%) and T2 (85%), as well as a significant difference (P=.01, Wilcoxon signed-rank test) in the mean level of patient participation assessed by the therapist at T1 (93%) and T2 (89%), but not in the level of motivation (P=.38, Wilcoxon signed-rank test) which for moderate patients always remained very high as shown in Figure 4. Within-group analyses did not highlight differences explained by age for any of the parameters investigated. Indeed, adherence to therapy, assessed as the level of motivation expressed by patients and participation assessed by the researcher/therapist, was different for patients according to the level of impairment (motivation: P=.02; r=0.142; participation: P=.045, r=–0.124; Wilcoxon signed-rank test and Spearman coefficient of correlation, respectively).

We asked patients to evaluate the device used for VR motor training according to the item of the USEQ scale reported in Figure 5. The patient performed this evaluation after the completion of the first training session (baseline, T0) and after the last one (end line, T2). All domains assessed with this scale scored at least more than 80% up to 95% in the first session, with small variation in the last one. Therefore, we did not find any difference between the two assessments (P=.56, Wilcoxon signed-rank test).

Importantly, the increase in satisfaction over time reflects patients’ growing engagement and acceptance of the therapy, whereas the stable usability scores indicate that the system did not require a prolonged learning phase and was perceived as usable from the outset.

Patients’ evaluation of cognitive load associated with the VR-MT training was assessed through the NASA-TLX cognitive index as shown in Figure 6, in which differences between the intermediate (T1) and the final evaluation (T2) were found for cognitive (P=.04, Wilcoxon signed-rank test), performance (P=.007, Wilcoxon signed-rank test), and effort domains (P<.001, Wilcoxon signed-rank test). As for acceptability measures, within-group analyses showed that age did not contribute to the variance of NASA domains in the longitudinal evaluation of the treatment.

On the contrary, post hoc analyses with Mann-Whitney U test for independent samples and Spearman correlation coefficient as depicted in Figure 7, clearly show that cognitive workload is perceived differently according to the level of impairment for cognitive (P<.001; r=–0.462), physical (P<.001; r=–0.422), temporal (P<.001; r=–0.408) and effort domains (P<.001; r=–0.317).

Notably, differences between impairment subgroups were more evident in measures related to perceived workload and participation, particularly in patients with severe impairment. While these observations cannot support inferential conclusions, they are consistent with clinical expectations and reinforce the relevance of impairment severity as a key factor in feasibility evaluations.