Five electronic databases (PubMed, Scopus, PEDro, ScienceDirect and Web of Science) were searched for relevant articles published in English or French from their inception until December 2022. An update was made to extend search to June 2023. To initiate the search, general keywords were first designed using core concepts: population (stroke), intervention, comparator and outcomes. A more detailed search strategy using combinations of key terms related to core concepts and their synonyms was also carried out. The search strategy was adapted to each database with combinations of keywords and Medical Subject Headings (MeSH) terms used as applicable. Published reviews and the reference lists of retrieved publications were searched manually in databases.

After duplicates were removed, two reviewers independently examined the titles and abstracts of identified studies for relevance using EndNote X9 software. Full-text copies of potentially eligible studies were assessed and determined according to the following inclusion and exclusion criteria. Study selection was determined by consensus between reviewers, and rating was performed. Differences in scores were discussed until consensus was reached. When necessary, disagreements were resolved by consensus involving a third author.

The inclusion critera were randomised controlled trials (RCTs) published in English and French, involving acute and subacute (<6 months) stroke survivors (18) aged >18 years; investigating bilateral UL training like sensorimotor training, active and non-active movements; task-oriented training, strengthening and BT with or without a device, compared to unilateral training with or without a device; conventional therapy; neurodevelopmental therapy; conventional occupational or physiotherapy; electrical stimulation, to establish the effects of the interventions on UL function measured by the Fugl-Meyer Assessment for upper extremity (FMA-UE), the Wolf Motor Function Test (WMFT), Action Research Arm Test (ARAT), Box and Block Test (BBT) and the Functional Independence Measure (FIM).

Systematic reviews or meta-analyses, uncontrolled trials, clinical trials, quasi-randomised trials, case studies, stroke duration ≥ 6 months post-stroke, other neurological conditions apart from stroke and those with participants under 18 years of age were excluded. In addition, studies that did not provide data as mean scores and standard deviation (SD) of outcomes were excluded from the meta-analysis, and those not in line with the definition of the World Health Organization (WHO) pertaining to rehabilitation (WHO 2011) such as invasive and pharmacological interventions.

Two authors used the Cochrane risk of bias tool to assess the risk of bias in studies. This tool assesses the risk of bias in seven areas: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting and any other bias (Higgins et al. 2011).

Two types of BT were considered in our review. The first category was bilateral UL training with device-assisted BT and non-device-assisted BT. The second category was symmetrical or non-symmetrical bilateral UL training. In symmetrical training, both ULs perform identical movements to manage a task. In non-symmetrical BT or tasks functional bilateral training, both ULs perform typically functional tasks, e.g. closing a box.

The most common definitions reported for dose dimensions of motor intervention include the duration of practice reflected by time spent in a therapy, schedule of therapy, for example, frequency of sessions and intensity level of task (Dalton et al. 2022; Hayward et al. 2021). The included studies were also classified into two subgroups according to the training dose: (group 1) total duration of training ≥ 20 h or a session length ≥ 5 h per week considered as a high dose; (group 2) total duration of training of < 19 h or a session length < 5 h per week considered as a low dose.

Upper limb impairment severity in the included studies was assigned based on the FMA-UE scores or ARAT at baseline. Fugl-Meyer Assessment for upper extremity scores of 0–22 or ARAT scores of 0–10 represent no capacity; FMA-UE scores of 23–31 represent poor capacity and match ARAT scores of 11–21; FMA-UE scores of 32–47 represent limited capacity and match ARAT scores of 22–42; FMA-UE scores of 48–52 represent notable capacity and match ARAT scores of 43–54; FMA-UE scores of 53 through 66 or ARAT scores 55–57 represent full capacity (Hoonhorst et al. 2015).

A meta-analysis was performed for the data synthesis using Review Manager Version 5.3 software, with a random effects model in which a p-value < 0.05 was considered significant. Effect size (ES) was estimated by calculating the standardised mean difference (SMD). The SMD reflects the intervention ES in each study relative to the variability observed in that study. An SMD of 0 means that the treatment and control have equivalent effects. Improvement is associated with higher scores on the outcome measure. Standardised mean differences >0 or <0 indicate the degree to which the treatment is more or less effective, respectively, compared to the control. Effect size was calculated based on means and standard deviations and on the sizes of the intervention and control groups. Heterogeneity was assessed using the results of the chi-squared test (significance level: p = 0.05) and the I² statistic to quantify consistency. An I² value of 50% or higher indicated the presence of substantial heterogeneity.

This systematic review and meta-analysis did not require formal ethical clearance because all data were obtained from publicly available sources and were analysed anonymously.

Figure 1 shows an overview of our selection strategy process. A total of 558 studies were selected through electronic databases, while four additional records were identified through other bibliographic sources. After removal of duplicates, screening of titles and abstracts and reviewing of full texts, 15 RCTs met the inclusion criteria for the qualitative analysis and 14 studies were included in the meta-analysis. One trial was excluded for meta-analysis because data were not available either publicly or from the authors (Burgar et al. 2011).

The main characteristics of the studies and participants in the 14 studies are shown in Table 1. A total of 706 participants were included with an age range from 49.3 to 74.3 years (SD from 2.0 to 13.22). Seven studies recruited patients with limited motor capacity with a FMA-UE score mean from 33 to 43 (Desrosiers et al. 2005; Dhakate & Bhattad 2020; Lum et al. 2006; Meng et al. 2017; Van Delden et al. 2015) or ARAT score mean 30 and 26 (Brunner, Skouen & Strand 2012; Easow & Chippala 2019); three other studies had participants with no motor capacity with a FMA-UE score mean from 6.6 to 17 (Hesse et al. 2005; Lee et al. 2017; Renner, Brendel & Hummelsheim 2020); in three studies, patients had poor motor capacity (Hsieh et al. 2017; Ma et al. 2022; Morris et al. 2008), and in one study, patients had notable capacity according to the aforementioned classification of severity above (Kumagai et al. 2022).

Table 2 gives the type of BT described in each included study. Most studies were symmetrical biliteral training without device assistance.

The details of interventions and the control groups are provided in Table 2. Six studies explored the effects of bilateral device-assisted training on UL motor function after stroke. The control group was unilateral UL training, including conventional training, for example, neurodevelopmental therapy. Eight studies investigated BT without device-assisted and these were functional tasks training. Among these eight studies, six studies performed symmetrical BT and two studies performed non-symmetrical BT.

The total duration of the interventions was 10–30 h of BT for 2–6 weeks while the duration of a session varied from 20 min to 2 h, 3–6 times per week.

Overall, 100% of studies presented low risk of bias with respect to random sequence generation, 64.29% with allocation and concealment, 14.29% with blinding of participants and personnel, 71.43% with blinding of outcome assessment, 28.57% with incomplete outcome data, 50% with selective reporting and 64.29% with other biases.

The results of the FMA-UE scores from 10 studies revealed a significantly improvement in favour of BT compared to the unilateral training group (SMD = 0.48; 95% CI: 0.08–0.88; p = 0.02) (Figure 2). However, high heterogeneity was present (I² = 78%, p < 0.0001).

An analysis was performed according to the types of bilateral intervention on FMA-UE scores: device-assisted training and non-device-assisted training (Figure 2) and symmetrical with or without device training and non-symmetrical bimanual training (Figure 3). No significant improvements between bilateral UL training and unilateral UL training were observed in the analysis in terms of subgroups of interventions: bilateral device-assisted training (SMD = 0.36; 95% CI: −0.01 to 0.74) and bilateral non-device-assisted training (SMD = 0.64; 95% CI: −0.14 to 1.42). However, the results demonstrated significant improvement in favour of non-symmetrical BT (SMD = 1.36; 95% CI: −0.98, 1.75), but only one study was considered (Meng et al. 2017).

A subgroup analysis of high dose of FMA-UE scores showed significant effect (SMD = 0.64; 95% CI: 0.08–1.20; p = 0.03) but the low dose demonstrated no significant effect (SMD = 0.31; 95% CI: −0.22 to 0.85; p = 0.25) (Figure 4).

Three trials included in the subgroup of ‘no motor capacity’ reported significant results in favour of bilateral UL training in improving the UL impairments compared to unilateral training (SMD = 0.66; 95% CI: 0.16–1.15; p = 0.009). Two studies (Hsieh et al. 2017; Ma et al. 2022) with ‘poor motor capacity’ participants showed non-significant results (SMD = 0.06; 95% CI: −0.50 to 0.61). Five trials included in the subgroup of ‘limited motor capacity’ have also demontrated no significant improvement (SMD = 0.49, 95% CI: −0.23 to 1.20). The comparative effectiveness on impairment (FMA- UE) according to the severity of UL paresis is represented in Figure 5.

The effects of bilateral UL training compared to unilateral UL training on activities analysed by the WMFT, ARAT and BBT scores of seven studies did not demonstrate significant improvement in overall effect of activities following a group of training (SMD = −0.09 points; 95% CI: −0.15 to 0.32). We observed homogeneity of studies (I² = 34%, p = 0.17). The comparison revealed no significant difference in the analysis in terms of types of bilateral UL intervention compared to unilateral training. Based on dose of intervention and severity of UL paresis, the comparison between bilateral UL training and unilateral UL training showed no significant difference in the WMFT, ARAT and BBT scores.

However, a significant difference was observed in favour of bilateral UL training compared to unilateral UL training in terms of improvement of daily activities measured by FIM (SMD = 0.45; 0.13–0.78; p = 0.006) (Figures 6, 7). There was homogeneity in the studies (I² = 6%, p = 0.37).

A strength of our systematic review and meta-analysis is that it includes only studies with high-quality scientific evidence, namely RCTs. In addition, to the best of our knowledge, our review may be the first meta-analysis looking at the effect of types of bilateral UL training in acute and subacute stroke. A limitation of our systematic review and meta-analysis is the heterogeneity encountered among the studies and limited studies in some analysed subgroups. In addition, the search strategy was limited to full publications in English or French; therefore, relevant publications in other languages may have been missed.