Immersive virtual reality with synchronous neurostimulation for upper-limb recovery after stroke — a promising feasibility signal

Clinical Question (PICO)

In adults at least three months after ischemic or hemorrhagic stroke, with residual upper-limb impairment and a Fugl–Meyer upper-extremity score (FMA-UE) of 10–60, does MultiSensy—immersive virtual reality with task-synchronized, transcutaneous median-nerve sensory stimulation—outperform time- and movement-matched conventional upper-limb rehabilitation for arm impairment, task performance and perceived body dimensions over 33 days? The principal comparison is the randomized feasibility cohort, not the separate nine-person usability pilot. No prespecified treatment-effect subgroup estimates are presented in the main report; sex-specific analyses were not performed. Methods and eligibility.

Bottom Line

The authors conclude that combining immersive VR and targeted sensory stimulation can improve motor function, tactile discrimination and body representation, justifying larger studies. In the 25-person randomized cohort, mean FMA-UE improvement at day 33 was 13.17 versus 7.54 points, a 5.63-point incremental gain calculated from the reported means; P = 0.01. This is an encouraging, assessor-blinded feasibility result—not proof of durable independence gains, a specific stimulation mechanism, or effective home delivery. Primary results.

Design

Trial type: Single-center, parallel-group randomized feasibility study; therapists performing clinical assessments were blinded. Participants and treating personnel could not realistically be blinded to the intervention.
N randomized: 25.
- MultiSensy: 12.
- Conventional rehabilitation: 13.
Separate pilot: Nine participants; do not add them to the randomized treatment-effect denominator.
Randomization: Allocation generated by a custom MATLAB script supplied by an independent third party; the authors report allocation concealment. No stratification is described in the main text.
Setting: Clinic for Rehabilitation Dr. Miroslav Zotovic, Belgrade, Serbia.
Enrollment: April 2024–April 2025 for the randomized study.
Follow-up: Scheduled assessments on days 1, 10, 19 and 33; the final assessment was two weeks after treatment ended. A mean follow-up duration is not reported.
Analysis: Linear mixed-effects models for FMA-UE, Action Research Arm Test (ARAT) and Barthel Index, with group, time and interaction terms plus participant-level random intercepts. Follow-up comparisons used baseline-corrected changes. Other outcomes used repeated-measures or nonparametric tests as appropriate. The main report does not identify the clinical analysis as a formal intention-to-treat analysis.
Primary outcomes: FMA-UE motor impairment (0–66), ARAT upper-limb task performance (0–57), and body representation measured with a VR adaptation of the Body Landmark Test; higher motor scores are better. Body-dimension ratios approaching 1 indicate more accurate perception.

Design and statistical methods.

Population

Inclusion Criteria

Ischemic or hemorrhagic stroke at least three months previously.
FMA-UE 10–60 in the final eligibility criteria.
Ability to provide informed consent.

Exclusion Criteria

Previous neurological or psychiatric disorders, as specified by the investigators.
Severe cognitive impairment, defined as Montreal Cognitive Assessment score below 10.
Epilepsy.
Pacemaker or another electronic implant.
Inability to provide informed consent.

Baseline Characteristics

Randomized cohort: mean age 59 ± 10 years; nine women and 16 men.
Table 1 includes substantial variation in impairment and time since stroke, including participants many years after the event.
Four of 13 controls started at the ARAT maximum of 57, creating a ceiling effect and a smaller denominator for improvement analyses.
The FMA-UE lower eligibility bound was originally 16. Following an enrollment deviation involving a more severely impaired participant, it was broadened to 10 after ethics review. The investigators report a post hoc sensitivity analysis supporting the robustness of their findings; this does not make the amendment prespecified.

Participant characteristics and eligibility amendment.

Interventions

MultiSensy: Twelve one-hour sessions over three weeks, four sessions weekly. Each included individualized calibration and 6–8 VR games, lasting at least four minutes each, targeting movements of the affected shoulder, elbow, wrist and hand. A Meta Quest 2 headset tracked movement; median-nerve sensory stimulation at the affected arm used 50-Hz, charge-balanced biphasic pulses, with individually calibrated intensity and task-dependent pulse-width modulation. This was sensory feedback, not an implanted brain stimulator or a trial of motor-evoking functional electrical stimulation.
Conventional rehabilitation: Twelve one-hour therapist-guided sessions over the same period, combining individually tailored physiotherapy and occupational therapy, with duration and movement content matched to the experimental program. Controls briefly performed a VR task at the first and last treatment sessions for kinematic measurement, without the experimental sensory stimulation.
Delivery context: Treatment occurred in a clinical setting. Calibration took approximately 5–10 minutes initially and 2–5 minutes subsequently; reduced staffing or autonomous home treatment was not established by the randomized comparison.

Intervention details.

Outcomes

Primary Outcomes

The figures below are mean changes ± standard error of the mean (SEM), not final absolute scores or standard deviations. The primary comparison is a continuous-score difference, so HRs, ORs and RRs are not the appropriate effect measures. The main text does not report 95% confidence intervals for the follow-up between-group change differences; none are invented here.

Outcome and time	MultiSensy	Conventional rehabilitation	Incremental change, calculated	Reported P
FMA-UE, day 33	+13.17 ± 1.30	+7.54 ± 1.48	+5.63 points	0.01
ARAT, day 19	+8.25 ± 1.96	+2.44 ± 1.08	+5.81 points	0.029
ARAT, day 33	+9.17 ± 2.37	+3.11 ± 0.98	+6.06 points	0.049

Longitudinal models: Group-by-time interaction P = 0.0035 for FMA-UE and P = 0.013 for ARAT. The timepoint-specific change comparisons were post hoc comparisons within the primary-outcome analysis; they are not independent confirmatory trials.
Threshold-based improvement: At day 33, 12/12 MultiSensy participants versus 8/13 controls exceeded the paper's FMA-UE improvement threshold of 5.25 points. For ARAT, 8/12 versus 3/9 exceeded 5.7 points; four ceiling-score controls were excluded from that ARAT comparison. These are descriptive responder counts, not a published NNT or prespecified subgroup treatment effect.
Body representation: In the MultiSensy arm, perceived-to-real arm-length ratio improved from 0.709 ± 0.039 at baseline to 0.965 ± 0.027 at day 33; the paper reports P < 0.001 versus baseline. This is a within-group result. Controls did not show corresponding arm-representation improvement; hand perception did not show a consistent benefit. A within-arm P value alone does not establish a between-arm causal effect.
Subgroups: No prespecified subgroup effect estimates are supplied in the main paper. A secondary chronicity ANCOVA and a post hoc eligibility sensitivity analysis are described in supplementary material; this review does not invent numerical subgroup estimates from their mention.

Primary-outcome results.

Secondary Outcomes

Functional independence: Four of 12 MultiSensy participants and two of 13 controls reached the paper's clinically meaningful Barthel improvement threshold by day 33, but the group-by-time interaction was not significant. Better arm scores did not establish superior overall ADL independence.
Tactile discrimination: At day 33, improvement in correct-response proportion on the two-point discrimination task was 0.208 ± 0.046 versus 0.030 ± 0.028, P = 0.006. At day 19 it was 0.162 ± 0.044 versus −0.006 ± 0.038, P = 0.008.
Peripersonal-space testing: No significant distance-by-time interaction in either arm: P = 0.23 for MultiSensy and P = 0.37 for controls. This does not support a broad claim that all aspects of body-space perception normalized.
Exploratory kinematics: A principal component correlated with FMA-UE (r = 0.66) and ARAT (r = 0.60), both P < 0.001. These repeated-session associations were not an independently validated replacement for clinical assessment.

Secondary and exploratory results.

Adverse Events / Safety

No adverse events were reported in either the nine-person pilot or the 25-person randomized study. The small, selected sample and short observation period cannot establish uncommon-harm rates or safety in patients excluded because of epilepsy or electronic implants. Mortality, symptomatic intracranial hemorrhage and recurrent stroke were not the efficacy endpoints of this rehabilitation study. Safety report.

Figures

Figures are reproduced from Aurucci et al. (2026), with editorially paraphrased captions. The article is available under CC BY 4.0, subject to any separately credited third-party exceptions. Images are unchanged; captions below are adapted explanations. Click a thumbnail to enlarge.

Figure 1 — Participant flow through the randomized feasibility study — *Figure 1.* Participant flow through the randomized feasibility study. The diagram tracks screening, allocation to MultiSensy or conventional rehabilitation, follow-up and the observations available for analysis. The randomized comparison comprised 12 MultiSensy and 13 control participants; the separate nine-person pilot is not part of that comparison.

Figure 2 — The platform and its pilot evaluation — *Figure 2.* The platform and its pilot evaluation. Panel a depicts immersive rehabilitation tasks and median-nerve sensory stimulation synchronized to virtual contact; in the pinch task, stimulation builds while an object is held and stops at release. Panel b illustrates the movement-monitoring interface. Panel c shows the body-landmark task: the arm is hidden from view and the participant directs a pointer toward perceived anatomical landmarks. Panel d outlines the pilot comparison of two four-minute VR-plus-stimulation rounds with matched movements guided by physical posters. Panel e presents embodiment and usability ratings; panel f compares movement speed, smoothness, forearm rotation and trajectory error. Panel g presents immediate changes in perceived arm dimensions. These pilot observations concern usability and immediate responses, not the randomized longitudinal efficacy contrast.

Figure 3 — Changes in arm task performance and motor impairment — *Figure 3.* Changes in arm task performance and motor impairment. Panels a and b show individual examples of improved pinching and unsupported movement during training. Panels c and d plot changes in ARAT and FMA-UE at the intermediate, end-of-treatment and follow-up assessments. Boxes show medians and interquartile ranges, with individual observations displayed. Four controls whose baseline ARAT was already 57 were excluded from the ARAT improvement analysis. The reported between-group comparisons favored MultiSensy for ARAT at day 19 (P = 0.029) and day 33 (P = 0.049), and for FMA-UE at day 33 (P = 0.01). These are follow-up comparisons of change scores, not comparisons of the final absolute scores.

Figure 4 — Body-dimension perception and tactile discrimination — *Figure 4.* Body-dimension perception and tactile discrimination. Panel a contrasts actual and perceived landmarks and plots perceived-to-actual dimension ratios across the four assessments; a ratio of 1 represents accurate perception. Arm-length perception moved toward that value with MultiSensy, whereas controls did not show comparable arm-representation change. Panel b illustrates the one-versus-two-pin task and plots changes in correct-response proportion. Sensory comparisons used 12 participants per arm and favored MultiSensy at the end of treatment (P = 0.008) and follow-up (P = 0.006). Bars represent means with SEM. Within-group body-perception tests should not be mistaken for direct between-group treatment-effect estimates.

Figure 5 — Movement-derived assessment and exploratory clustering — *Figure 5.* Movement-derived assessment and exploratory clustering. Panel a illustrates frontal- and transverse-plane task trajectories and indices describing speed, smoothness, accuracy and task performance. Panel b follows these measures through 12 MultiSensy sessions; controls contributed recordings at the first and last sessions rather than equivalent repeated VR exposure. Panel c shows principal-component variance, feature contributions and the association of the leading component with clinical motor scores. Panel d groups 143 patient-session observations into three clusters and compares their clinical-score distributions. Those observations are repeated sessions, not 143 independent patients; imputation and interpolated clinical scores were used in parts of the exploratory analysis. The clusters are development-sample findings, not externally validated severity predictions.

Extended Data Figure 1 — Treatment and assessment schedule — *Extended Data Figure 1.* Treatment and assessment schedule. Both randomized groups received 12 sessions over three weeks, matched for time and movement content. Each MultiSensy session included individualized stimulation calibration followed by six to eight selected games of at least four minutes each. The shared assessment schedule was baseline on day 1, the intermediate evaluation on day 10 after six sessions, treatment-end testing on day 19 and follow-up on day 33. Measures covered motor impairment, arm-task performance, independence, tactile discrimination and body-space perception.

Extended Data Figure 2 — Validation of the virtual body-landmark measurements — *Extended Data Figure 2.* Validation of the virtual body-landmark measurements. Panel A contrasts physical landmark measurements taken with a grid against virtual measurements in healthy volunteers. Panel B summarizes systematic dimension-estimation errors for the Meta Quest 2 and HTC Vive Pro systems, using nine and seven healthy participants respectively. Panels C and D compare these system errors with the size of body-representation changes in the longitudinal and pilot studies. Values are summarized with SEM and evaluated with independent-sample comparisons. This technical validation supports measurement plausibility; it does not independently establish the clinical mechanism or efficacy of the rehabilitation package.

Criticisms

Small proof-of-concept sample: Only 25 people were randomized. The sample-size calculation assumed a very large standardized effect of 1.38, leaving considerable uncertainty about more modest effects.
Very short durability window: Follow-up stopped two weeks after treatment; sustained real-world arm use and long-term independence remain open questions.
Multiple outcomes and timepoints: Several primary domains, repeated comparisons and least-significant-difference post hoc tests make isolated nominal P values—especially ARAT P = 0.049—less secure than a single prespecified, multiplicity-controlled endpoint.
Ceiling-related exclusions: ARAT improvement analyses omit four controls already at the maximum; the resulting comparison is not the full randomized sample. Baseline balance tests cannot eliminate the consequences of a small sample or outcome-dependent ceilings.
Protocol deviation and missing observations: The lower impairment threshold changed after enrollment began; supporting sensitivity analysis was post hoc. Outcome sample sizes vary, and the paper says missing observations were handled according to the relevant model without a single, fully specified clinical missing-data strategy in the main text.
Bundled treatment, unresolved mechanism: No VR-only, stimulation-only or sham-stimulation arm isolates synchrony or sensory stimulation from immersive practice, motivation or expectation.
Not a home-delivery or workforce trial: Neither unsupervised home effectiveness nor fewer therapist visits was demonstrated. Superiority on Barthel independence was not established.
Exploratory algorithm limitations: The clustering analysis uses repeated observations from the same small patient group, interpolated clinical scores and imputation for some kinematic data; correlation is not external predictive validation.

Limitations and analysis methods.

Funding

Swiss National Science Foundation, MOVEIT grant 197271; open-access funding from the Medical University of Vienna. The funding section does not specify a separate sponsor role in study design, conduct or analysis. The acknowledgements state that the authors had full data access and that the corresponding author made the final submission decision. The authors declare no competing interests. Registration: NCT06400823. Funding and disclosures.

The paper

Authors. Giuseppe Valerio Aurucci et al.
Title. Immersive virtual reality with synchronous neurostimulation for upper-limb recovery after stroke: a randomized feasibility trial.
Journal. Nature Medicine, volume 32, pages 2790–2802.
Year. 2026; first published online June 26, 2026. This is a look-back review, not a claim that the paper appeared today.
DOI. 10.1038/s41591-026-04486-4
PMCID. PMC13472961
PMID. 42362865

Deep Dive — click to expand

What this is

A rehabilitation proof of concept published in June 2026 asks whether practicing with a virtual arm that both moves and feels like one's own can improve recovery after the early stroke period. The randomized evidence comes from 12 experimental and 13 control patients, not all 34 people described in the abstract. A calculated 5.63-point incremental FMA-UE gain at day 33 is worth pursuing. Its proper destination is a larger, carefully controlled trial—not a claim that a headset replaces a therapist. The commentary below is editorial interpretation of the linked primary paper, not a current guideline recommendation. Trial results.

1. Shadow Audit

The hidden competitor is renewed access to structured rehabilitation, not doing nothing. Controls improved by 7.54 FMA-UE points, and eight of 13 exceeded the paper's clinically meaningful threshold. The experimental group's 13.17-point gain therefore overstates the device-specific opportunity if presented without its comparator: the incremental difference is 5.63 points. Chronic impairment is not synonymous with exhausted rehabilitation potential in this selected cohort.

The other shadow is the mismatch between motor capacity and independence. Barthel's group-by-time interaction was not significant despite positive motor-score results. That does not prove no ADL effect, but it prevents marketing the experiment as established independence restoration. Similarly, within-arm normalization of perceived arm length is interesting mechanistic evidence, not proof that sensory remapping caused the motor gains. Motor and independence results.

2. Inversion Engine

For the opposite efficacy conclusion to hold at the level of the reported means, the 5.63-point FMA-UE advantage would have to disappear or reverse. A future well-controlled trial could do that if novelty, expectation, baseline imbalance, therapist interaction or analysis choices account for a difference of that magnitude. This is an arithmetic tipping-point description—not a formal sensitivity analysis or a calculated probability of bias.

The mechanistic conclusion is easier to invert. Equivalent improvement with VR plus sham or asynchronous stimulation would undermine the claim that synchronized sensory feedback adds value even if the complete package remains helpful. A trial should therefore preserve dose and game exposure while randomizing the feedback component. For clinical importance, one should also distinguish a between-group mean difference from the paper's 5.25-point within-person responder threshold: those are different estimands, not interchangeable pass/fail rules. Design and thresholds.

3. Second-Order Catalyst

The first plausible adopter is a specialist neurorehabilitation research service, with therapists able to calibrate electrodes, select tasks and independently score outcomes. The next operational experiment is not an unsupervised headset shipment: it is a supervised replication for patients resembling those enrolled, retaining the same 12-session schedule and blinded assessment.

If efficacy survives replication, rehabilitation departments could test whether one clinician safely supervises more than one patient, or whether part of the program can move home. Those are separate implementation hypotheses. Initial calibration took 5–10 minutes and later calibration 2–5 minutes; those measurements identify workflow work to do, not verified staffing savings. The likely early change is trial enrollment and measurement discipline, not reduced therapist coverage. Clinical delivery and limitations.

4. Asymmetric Leverage

The favorable asymmetry is recoverable function in people whom a service might otherwise stop reassessing. Even the conventional arm's improvement suggests value in identifying persistent deficits and offering another structured, goal-directed course rather than assuming an immutable plateau. That inference is bounded by selection: the study excluded major cognitive impairment, epilepsy and electronic implants, and cannot represent all stroke survivors.

The technological leverage is potentially a treatment-and-measurement package: repeated practice plus continuous movement data without a separate motion-capture visit. But the principal-component correlations—0.66 with FMA-UE and 0.60 with ARAT—are associations within this development dataset, not evidence that an algorithm can safely replace the examination. Before scaling, demand patient-level external validation, sustained functional benefit, actual staff-time measurement and total cost. The potentially large service denominator does not make a small feasibility trial more certain. Exploratory analysis.

5. Paradigm Destroyer

The reflex most challenged is “past three months means beyond useful rehabilitation,” not “conventional therapy is obsolete.” Both groups improved; the device may add something, but the study cannot assign that addition to immersion, sensation or their synchrony separately.

A cautious two-sentence service update for discussion would be: “Reassess persistent upper-limb limitations after the early recovery period using standardized motor and task-performance measures, and consider renewed individualized rehabilitation where appropriate. Offer combined immersive VR and sensory stimulation through a supervised research or evaluation pathway rather than substituting it for established therapy or assuming home-delivery equivalence.” This is an editorial proposal informed by feasibility evidence, not a new guideline or an instruction to change an individual patient's treatment. Interpretation and limitations.

MVP — Minimum Viable Proof

The minimum convincing next result is a prespecified, blinded-assessor, between-group benefit that persists beyond treatment and translates into meaningful arm use or daily activity, with confidence intervals that rule out a trivial effect. The present numerical starting point is a calculated +5.63 FMA-UE points at day 33, with no reported adverse events—not a proven long-term effect. A replication should retain an active, dose-matched comparator, specify missing-data and multiplicity handling in advance, and show that outcomes are not driven by ceiling exclusions or a few participants. A component-control arm is required if the claim is specifically about synchronized sensory stimulation.

Best Combination

Read this as a small, differentiated contribution to the heterogeneous VR and stimulation literature—not a head-to-head victory over every rehabilitation technology. The paper cites a 2017 Cochrane review of VR, a 2023 immersive-versus-nonimmersive VR review, and earlier hybrid stimulation/robotic work; those are its historical comparisons, not an independently updated evidence synthesis for October 2026. Differences in chronicity, therapy dose, comparator intensity and outcome timing prevent ranking modalities by cross-study raw gains alone. The paper's background and references.

The most defensible combination is task-oriented rehabilitation with standardized outcome assessment, using the multisensory package as an investigational adjunct. The next study should ask both whether the package adds durable patient-important benefit and whether synchronous feedback contributes beyond the same VR practice. Neither question is answered by kinematic correlation or by the absence of reported adverse events in this small cohort.

Overvalue Warning

Do not sell the experimental arm's entire improvement as the device effect. The FMA-UE gain was 13.17 points, but controls gained 7.54; the calculated incremental contrast was 5.63. ARAT improvement analyses also used unequal denominators because four controls began at the ceiling.
Do not translate a supervised, 33-day feasibility study into proven autonomous home rehabilitation or staffing savings. Follow-up ended two weeks after treatment, independence superiority was not established, and rare harms cannot be ruled out by zero reported events in 34 participants across two different study phases.

Educational use notice. The Time Is Brain Review is dedicated to medical education. Figures from published articles are embedded for teaching purposes, with attribution to the original publication. We do not claim copyright over any embedded figure; all rights remain with the original publisher and authors under the article's published license. If you are a rights holder and would like a figure removed, please contact the editorial team. See our editorial notice for the full policy.