The Cochrane review on VR for stroke
The larger, more authoritative synthesis: 190 trials and a reversed verdict.
The VR stroke meta-analysis and its fine print.
Meta-analyses are where a field's small studies come to be counted honestly. In 2021, a team pooled 19 studies of VR rehabilitation for subacute stroke and found what fans of the technology hoped for, until they corrected for publication bias and the advantage over conventional therapy dropped to zero. Both versions of the result are worth understanding.
Is VR rehabilitation effective for motor recovery after stroke?
The 2021 meta-analysis by Peng and colleagues, pooling 19 studies of subacute stroke, found patients improved substantially from baseline with VR rehabilitation. But VR's apparent advantage over conventional therapy showed significant publication bias, and after statistical correction the effect became non-significant. The supportable conclusion: VR is a feasible, engaging way to deliver rehabilitation, not yet a demonstrably superior one.
That distinction sounds deflating and isn't: an engaging route to the same outcome matters enormously when the alternative is patients doing less therapy, or none.
At a glance
- Article analysed
- Peng, Yin & Cao, Frontiers in Neurology, 2021 (PRISMA systematic review with meta-analysis)
- What was pooled
- 19 studies (17 RCTs, 1 cohort, 1 crossover) of subacute stroke, searched across nine databases; 16 in the quantitative synthesis
- The interventions
- Highly varied: Xbox Kinect, Nintendo Wii, Leap Motion, YouGrabber and others; 20–60 min sessions, 3–5 days/week, 2–12 weeks, mostly added to conventional therapy
- Versus baseline
- Standardized mean difference 1.14 (95% CI 0.77–1.52, p < 0.001), no publication bias detected
- Versus conventional therapy
- Initially SMD 0.47 (p < 0.001), but with significant publication bias; after trim-and-fill correction: SMD 0.08 (95% CI −0.16 to 0.33, p = 0.507)
- Engagement signal
- Feasibility trials rated participation excellent to very good; one study logged 77.6 minutes of active practice in VR vs. 67.3 conventional
- Study quality
- 10 studies low risk of bias, 5 with some concerns, 3 high risk, 1 moderate; heterogeneity high (I² 75–83%)
Key takeaways
- The two headline numbers tell different stories: patients on VR programs clearly improved (SMD 1.14 vs. baseline), but improvement-from-baseline in subacute stroke includes natural recovery, and the comparison against conventional therapy is the one that isolates VR's contribution.
- That comparison did not survive publication-bias correction: from SMD 0.47 to 0.08 after trim-and-fill. Small positive trials got published; small null ones apparently didn't.
- Most included designs compared VR added to conventional therapy against conventional therapy alone, so even the uncorrected effect confounds VR with extra therapy time.
- The pooled interventions ranged from Wii games to purpose-built systems, exactly the heterogeneity Perez-Marcos warned about; a null pooled result over mixed interventions is not a null result for each of them.
- What survives the correction is feasibility and engagement, including more active practice time in VR in one trial. In a field whose limiting factor is dose, that is not a consolation prize.
What did the meta-analysis include?
The authors searched nine databases through October 2020 (updated March 2021) for studies of VR-based motor rehabilitation in subacute stroke, the recovery window after the acute phase, when most spontaneous improvement happens. Nineteen studies qualified: 17 randomized controlled trials, one cohort study, one crossover, with individual samples ranging from 4 to 120 participants. Sixteen entered the quantitative pooling.
The interventions were a bestiary: Xbox Kinect setups, Nintendo Wii, Leap Motion hand tracking, the YouGrabber system, and more, delivered 20 to 60 minutes a session, three to five days a week, for two to twelve weeks. Most trials added VR to conventional therapy rather than substituting for it, and most targeted upper-extremity function.
Methodological quality was mixed but not dismal: by Cochrane risk-of-bias assessment, ten studies were low risk, five had some concerns, three were high risk, one moderate.
What were the results, before and after correction?
Versus baseline: large and stable
Pooled across studies, patients receiving VR rehabilitation improved markedly from their own starting point: standardized mean difference 1.14 (95% CI 0.77–1.52, p < 0.001). Statistical checks found no publication bias on this comparison. But in subacute stroke this number has a low ceiling of meaning, patients in this window improve substantially with any care, or none, because spontaneous neurological recovery is doing much of the work.
Versus conventional therapy: the twist
The comparison that matters, VR versus conventional rehabilitation, initially showed a modest significant advantage: SMD 0.47 (95% CI 0.22–0.72, p < 0.001). Then the authors ran Egger's regression and found significant publication bias (p = 0.001), the small-study pattern suggesting null results went unpublished. After trim-and-fill adjustment, which estimates the effect with the apparently missing studies restored, the advantage collapsed: SMD 0.08, confidence interval spanning zero, p = 0.507.
To the authors' credit, they ran the correction and reported it plainly. Many meta-analyses in rehabilitation stop at the uncorrected number, and secondary citations of this literature often quote the 0.47 without the 0.08.
How should these findings be interpreted?
Three readings are consistent with the data, and they are not mutually exclusive. First, the skeptical one: VR delivers no specific benefit beyond conventional therapy, and the appearance of one was a publication artifact. Second, the confound reading: because most trials tested VR plus conventional therapy against conventional therapy alone, even the uncorrected effect measured extra therapy time as much as VR; a clean substitution design might show more, or less.
Third, the heterogeneity reading: pooling Wii bowling with purpose-built rehabilitation systems, with I² values above 75%, averages together interventions that may have genuinely different effects. This is precisely the confound Perez-Marcos identified in 2018, and immersive, embodiment-based systems remain a small minority of the pooled trials. A washed-out average is what you would expect if some interventions work and others merely entertain.
What no reading rescues is a blanket claim that “VR beats standard rehab.” The current pooled evidence in subacute stroke does not support it, and honest commentary, ours included, has to say so.
What can VR rehabilitation legitimately claim, then?
Feasibility, tolerability, and engagement, and these matter more than they sound. The trials that measured participation rated it excellent to very good, and one logged 77.6 minutes of active practice in the VR condition against 67.3 minutes in conventional therapy. Rehabilitation's oldest problem is that recovery demands thousands of repetitions and patients, understandably, do not do them. A modality that gets ten extra minutes of real practice out of a session is moving the lever that actually drives outcomes.
The same logic extends to reach: systems that work at home, as in the clinic-to-home HEAD pilot, address dose by removing the commute. And for chronic pain, where the mechanism is retraining a protective nervous system rather than rebuilding motor pathways, engagement is even more central, the treatment only works if people keep showing up to move. That framing, with the supporting evidence and its limits, is laid out in our VR pain management guide and program overview.
The field's task, meanwhile, is sharper trials: substitution designs rather than add-ons, immersive embodiment systems analysed separately from console games, and follow-up beyond twelve weeks. The meta-analysis, twist included, is a map of exactly where the evidence needs to go.
Since this meta-analysis was published, a much larger synthesis has appeared: the 2025 Cochrane review of 190 trials and 7,188 participants. It reaches a more favourable conclusion, particularly for VR added to usual care, while carrying its own sensitivity-analysis caveat. If you want the single most authoritative answer on VR for stroke, read our commentary on the Cochrane review alongside this page.
Frequently asked questions
Did this meta-analysis show that VR rehabilitation doesn't work?
No. It showed patients improved substantially from baseline with VR programs, and that current pooled evidence does not demonstrate VR is better than conventional therapy in subacute stroke once publication bias is corrected.
“Not demonstrably superior” differs from “ineffective,” especially when most trials tested VR as an add-on and mixed very different technologies together. It also leaves feasibility and engagement findings fully intact.
What is publication bias and trim-and-fill?
Publication bias is the tendency for studies with positive results to reach print while null results stay in file drawers, inflating pooled effects. Egger's regression detects the statistical asymmetry this leaves behind; in this meta-analysis it flagged the VR-versus-conventional comparison (p = 0.001).
Trim-and-fill estimates what the pooled effect would be with the missing studies restored. Here it moved the effect from 0.47 to 0.08, from significant to indistinguishable from zero.
Why does improvement from baseline mean so little in subacute stroke?
Because the subacute period is when the brain does most of its spontaneous recovery. Patients improve over these weeks with almost any care, so a before-after gain in a treated group cannot be attributed to the treatment.
That is why controlled comparisons are the meaningful test, and why this meta-analysis's corrected between-group result carries more interpretive weight than its large within-group one.
Were immersive headset systems included?
The pooled interventions were predominantly screen-and-sensor systems, Kinect, Wii, Leap Motion, YouGrabber, rather than head-mounted displays. Immersive, embodiment-based VR was scarcely represented.
So the corrected null applies most directly to non-immersive game-style interventions in subacute stroke. Whether embodied immersive VR performs differently is an open question the included trials could not answer, which is the heterogeneity problem discussed in our VR-versus-videogames commentary.
Does this result apply to VR for chronic pain?
Not directly. This meta-analysis concerns motor recovery in subacute stroke, a different population, mechanism, and outcome set from chronic pain, where VR aims to retrain threat evaluation and movement confidence rather than rebuild motor pathways.
The transferable lessons are methodological: prefer controlled comparisons, watch for publication bias, and don't let a pooled average of dissimilar interventions stand in for the specific program in front of you, standards we apply to our own evidence in how it works.
Sources & research.
Related guides.
Free consultation
Ready to unlearn your pain?
Talk with our care team about your pain, your history, and whether KVET™ is right for you. Free, and from the comfort of home.