This review adhered to the updated criteria of Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 (PRISMA 2020) (20). The protocol was pre-registered on PROSPERO on 21st October 2024 (registration number CRD42024604104).

A pilot search was initially conducted on Medline and EMBASE to assess the feasibility of this review by confirming the availability of potentially relevant studies. The search also yielded no results for recent systematic reviews exploring the specific aims of this review. Therefore, an electronic search was conducted on Medline, EMBASE, EMCARE and PsycINFO via Ovid from inception of databases to 17th December 2024, for relevant studies available in English. Validity of the search was confirmed by using several pre-identified relevant articles. Full search strategy can be found in Supplementary Appendix A1.

Included studies were required to be randomised controlled trials of upper-limb rehabilitation in adults (≥18 years) with a diagnosis of stroke, in the late subacute (>3–6 months) or chronic (>6 months) phase of recovery (21). Interventions were required to have a minimum of two groups delivering different therapy durations; a low-dose group (considered the control) and a high-dose group (>50% increase in duration). Dose was defined as the total amount of practice, operationalised as the time spent in upper-limb therapy (hours). Studies were required to have the same therapy content between groups. Included studies were required to have a measure for upper-limb outcomes, such as the Fugl-Meyer Assessment (FMA), the Action Research Arm Test (ARAT) and the Wolf Motor Function Test (WMFT). The measures were required to be performed pre- and post-interventions.

Theses, non-traditional outputs (portfolios, creative works and research projects from external bodies), conference abstracts and studies where full text could not be located were excluded. Additionally, studies were excluded if they compared usual care with usual care plus another therapy without controlling for therapy content, compared therapy with no therapy or with a sham control and did not report the therapy duration (e.g., total hours).

Following the search, all references were imported to EndNote where duplicates were removed. The remaining records were imported to Covidence, where further duplicates were automatically removed by the automation tool. Two independent reviewers screened titles and abstracts for relevance, with the screening distributed among three reviewers. All studies excluded by the automation tool were cross-checked by one reviewer to confirm ineligibility. Two reviewers then independently reviewed full texts of the remaining studies to screen for potentially eligible studies. At all stages, any conflicts and discrepancies were resolved through a discussion and involved a third reviewer where necessary.

One reviewer extracted all the relevant data from each study which was then cross-checked by a second reviewer. Extracted information included, first author's name and publication year, study design, participant's—age, sex, sample size enrolled and completed, time since stroke, description and duration of interventions and upper-limb outcome data. Where key data such as sample size, mean values, or standard deviations for outcome measures was missing, study authors were contacted via email. Where required, contact was attempted on up to two occasions, separated by at least two weeks.

Two independent reviewers used the Cochrane risk of bias tool—two (RoB2) to assess bias of the included studies (22). The tool measures bias across five areas: bias from the randomisation process, deviations from intended interventions, missing outcome data, measuring of outcome and the selection of reported results. Each study was assigned an overall risk of bias rating (low risk, some concerns, high risk) based on these domains. Summary graphs and tables were showcased with the robvis tool.

A narrative synthesis was conducted for all the included studies. Meta-analyses were conducted to summarise and compare results where outcomes could be pooled. The meta-analyses examined the effect of increased upper-limb rehabilitation duration on upper-limb recovery post-stroke to calculate pooled effect sizes. To address the primary objective, the pre-and post-intervention values from the primary (FMA) and secondary (ARAT, WMFT) outcome measures were used to explore the effect of increased rehabilitation duration. Effect sizes were reported as mean difference (MD) with a 95% confidence interval (CI). A sensitivity analysis was conducted if significant heterogeneity was present. Certainty of evidence was rated as high, moderate, low or very low using GRADE methodology (23). GRADE considers risk of bias (informed by RoB2), inconsistency, indirectness, imprecision and publication bias. To investigate secondary objectives, secondary analyses were conducted involving subgroups. The subgroup analyses were done by groupings based on percentage increase (50%–100%, 100%–150% and ≥200%) of therapy dose.

The electronic search identified 2,600 studies across all the databases, after screening, four studies were included in the review (Figure 1).

Four studies included a total of 151 participants (24–27). The mean age across studies ranged from 57 to 61 years and included roughly 57% males and 43% females. Most participants (n = 148) were in the chronic phase of stroke recovery (>6 months), with only three participants in the late subacute phase (>3 months) from one study (24). All studies delivered upper-limb therapy, with a minimum of one group receiving a higher duration of therapy and a lower duration group of the same therapy. Total time spent in rehabilitation ranged from 15 to 80 h, delivered across three to five sessions per week for six to ten weeks. Three studies included multiple treatment-dosage groups that met the inclusion criteria (25–27). All studies used the ARAT to assess upper-limb function, with two also reporting on the FMA and one on the WMFT. One study could not be included in the meta-analyses as essential data (standard deviations) were not available and could not be retrieved from the author (25). A full summary of study characteristics is provided (Table 1).

Ross et al. included two groups receiving 15 h and 45 h of task-specific upper-limb therapy (24). The 45-hour group had a mean change of 12.4 ± 20.7 on the ARAT and 13.4 ± 19 on the WMFT compared to the 15-hour, with mean change of 2.6 ± 5.8 (ARAT) and 0.4 ± 1 (WMFT) post-intervention. Page et al. (26) explored three groups receiving 20, 40, 80 h of repetitive task-specific upper-limb therapy (26). The 80-hour group showed a mean ARAT change of 3.7 ± 3.0 and FMA change of 4.1 ± 2.9; the 40-hour group changed by 1.4 ± 1.8 (ARAT) and 1.3 ± 2.2 (FMA); the 20-hour group changed 1.9 ± 3.9 (ARAT) and 1.9 ± 1.6 (FMA). Lang et al. compared three groups 13.6, 26.3 and 32.8 h of task-specific upper-limb therapy (27). The ARAT scores changed from 33.7 ± 7.9 to 37.8 ± 8.8 in the 13.6- hour group, from 32.1 ± 12.3 to 35.7 ± 14.3 in the 26.3-hour group and from 31.6 ± 10.3 to 36.9 ± 12.6 in the 32.8-hour group. Page et al. included two groups receiving 25 and 45 h of repetitive task-specific upper-limb therapy (25). The 25-hour group improved scores from 24.22 to 26.78 on the ARAT and FMA scores from 30.86 to 34.20. The 45-hour group improved from 25.17 to 26.78 on the ARAT and 28.33 to 34.00 on the FMA.

Two studies had low risk in the randomisation process, while the other two studies showed “some concerns” due to inadequate allocation concealment (24–27). Bias due to deviations from intended intervention was low in two studies and raised “some concerns” in the others due to lack of blinding. Three studies had low risk for missing outcome data. All studies had low risk for bias in outcome measurement. Overall, all studies were rated as having “some concerns” (Figure 2).

This review is limited by the small number of included studies, with only three eligible for meta-analyses. Furthermore, intervention protocols, outcome measures and participant characteristics varied substantially across studies. Although upper-limb activity was consistently measured using the ARAT across all studies, only one study reported impairment-level outcomes using the FMA and another using the WMFT for activity measure. The limited number of studies reporting these impairment-level measures precluded pooled analysis at that level. While the overall findings suggest that more therapy is associated with better outcomes on the FMA, and that a ≥ 200% increase is needed to see larger results on the ARAT, more studies are needed to provide robust evidence to verify this. Furthermore, timing of intervention should be considered. While all included studies included people in the sub-acute and chronic phase, we note time since stroke varied between studies. It may be that even greater amounts of rehabilitation are needed in those who are very chronic. Another limitation was that some potentially relevant studies used different intervention types across dose groups. They could not be included as the type of intervention may influence the outcome. To rigorously explore the effect of dose, more studies that control the type of intervention across different doses are needed.

Across the four included studies, the low-dose (considered the control) group consistently provided upper-limb therapy between 13.6 and 25 h over six to ten weeks. This relative consistency can allow a practical extrapolation: a minimum 200% increase would equate to roughly 40–75 h of therapy over six to ten weeks. The two highest dose arms that met or exceeded this dose showed clinically meaningful results in upper-limb recovery post-stroke (24, 26). This suggests that future research and clinical guidelines could consider this as a potential benchmark for achieving functional gains in the late subacute and chronic stroke population.

There is a need for large-scale, high-quality RCTs with multiple upper-limb outcome measures, transparent trial protocols in the late subacute and chronic phase of stroke recovery and specifically looking at upper-limb recovery. This will assist in investigating threshold effects and identify patient subgroups most likely to benefit from an increased dose of rehabilitation.