Robot‐assisted therapy for upper‐limb rehabilitation in subacute stroke patients: A systematic review and meta‐analysis

Abstract Background Stroke survivors often experience upper‐limb motor deficits and achieve limited motor recovery within six months after the onset of stroke. We aimed to systematically review the effects of robot‐assisted therapy (RT) in comparison to usual care on the functional and health outcomes of subacute stroke survivors. Methods Randomized controlled trials (RCTs) published between January 1, 2000 and December 31, 2019 were identified from six electronic databases. Pooled estimates of standardized mean differences for five outcomes, including motor control (primary outcome), functional independence, upper extremity performance, muscle tone, and quality of life were derived by random effects meta‐analyses. Assessments of risk of bias in the included RCTs and the quality of evidence for every individual outcomes were conducted following the guidelines of the Cochrane Collaboration. Results Eleven RCTs involving 493 participants were included for review. At post‐treatment, the effects of RT when compared to usual care on motor control, functional independence, upper extremity performance, muscle tone, and quality of life were nonsignificant (all ps ranged .16 to .86). The quality of this evidence was generally rated as low‐to‐moderate. Less than three RCTs assessed the treatment effects beyond post‐treatment and the results remained nonsignificant. Conclusion Robot‐assisted therapy produced benefits similar, but not significantly superior, to those from usual care for improving functioning and disability in patients diagnosed with stroke within six months. Apart from using head‐to‐head comparison to determine the effects of RT in subacute stroke survivors, future studies may explore the possibility of conducting noninferiority or equivalence trials, given that the less labor‐intensive RT may offer important advantages over currently available standard care, in terms of improved convenience, better adherence, and lower manpower cost.


| INTRODUC TI ON
Stroke is one of the leading causes of death and disability worldwide (Feigin, Lawes, Bennett, & Anderson, 2003). About 17%-40% of stroke survivors experienced upper extremity spasticity, worsening their abilities in performing activities of daily living (ADL) (Hsieh et al., 2017;Pollock et al., 2014). Upper-limb rehabilitation is crucial during the first six months since the onset of stroke because the motor and ADL recovery of stroke survivors declines afterward (Kwakkel & Kollen, 2013). After the 6-month poststroke period, 33%-66% of patients fail to achieve upper-limb functional recovery (Kwakkel & Kollen, 2013).
Conventional poststroke rehabilitation, including "hands-on" therapy (manual therapy techniques), constraint-induced movement therapy, repetitive task training, and mirror therapy (Pollock et al., 2014), usually requires patients to perform partial or full-assisted movement in arm/hand joints manually under the supervision of therapists. However, the time-consuming and labor-intensive nature of conventional therapies has limited its cost-effectiveness.
Robot-assisted therapy (RT) is a novel approach to poststroke rehabilitation, which utilizes robotic devices to deliver motor or task-oriented training to patients (Brewer, McDowell, & Worthen-Chaudhari, 2007). Apart from providing repetitive and high-intensive training in a cost-effective fashion (Lo, Stephenson, & Lockwood, 2018), stroke survivors can perform independent training with less supervision from therapists, receive timely feedback on their performance from robotic devices, and achieve better adherence to treatment with an introduction of games or interactive upper-limb tasks (Hesse, Hess, Werner, Kabbert, & Buschfort, 2014;Kwakkel, Kollen, & Krebs, 2008).
Despite several advantages of RT suggested in literature, there are no conclusive evidence for the beneficial effects of RT over usual care in stroke patients. Two meta-analyses have indicated that when the dose of RT was matched with that of usual care, no significant between-group differences were found in motor control and abilities in performing basic ADL (Kwakkel et al., 2008;Norouzi-Gheidari, Archambault, & Fung, 2012). However, other meta-analyses have shown that using RT as an adjunct to usual care is more effective than RT alone on improving upper-limb motor function, in terms of motor control (e.g., Fugl-Meyer Assessment of the arm), muscle strength/tone and basic ADL (Bertani et al., 2017;Mehrholz, Pohl, Platz, Kugler, & Elsner, 2018;Veerbeek, Langbroek-Amersfoort, van Wegen, Meskers, & Kwakkel, 2017;Zhang, Li-Tsang, & Au, 2017). Of note, in the aforementioned reviews (Bertani et al., 2017;Mehrholz et al., 2018;Veerbeek et al., 2017;Zhang et al., 2017), pooling outcomes of RT studies to obtain an overall effect irrespective of different phases of poststroke recovery may have produced overgeneralized conclusions. Throughout the stroke trajectory, patients' training needs and progress of recovery change over time and may have masked the pooled effect of RT. Though the optimal stroke recovery occurs in the first few months after a stroke, the impacts of RT within six months poststroke remains unclear in literature. In this review, we aimed to examine the research evidence in the past 20 years regarding the effects of RT on outcomes related to body function, activities, and social participation in patients diagnosed with stroke within six months and assess the methodological quality of the included studies.

| ME THODS
This review adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines for reporting systematic reviews (Moher, Liberati, Tetzlaff, & Altman, 2009).

| Search strategy
We identified studies published between January 1, 2000 and

| Study selection
Two authors (WTC, HYC) independently reviewed the title and abstract of identified studies, examined the full-text reports of all potentially relevant studies according to the predefined eligibility criteria. Disagreements of study selection were resolved through discussion with another author (YYC).
We included randomized controlled trials (RCTs) published in English, which had examined the effects of RT. In these RCTs, the participants were at aged 18-65 years (of both gender) in which at least 60% of them had a primary diagnosis of first-ever stroke with the poststroke period equal or less than six months at the baseline of the study. The included RCTs might adopt RT with different formats (robot-integrated physiotherapy, homebased robotic tele-rehabilitation, robotic training combined with games, and bilateral robotic priming) as the main component of intervention. The RT was reported as a stand-alone therapy or as an adjunct to conventional therapy used in usual care, while the control conditions could be of any types but not RT, such as conventional therapy and physical therapy. The primary outcome was motor control under the "body function" domain based on the International Classification of Functioning, Disability and Health (ICF) (Sivan, O'Connor, Makower, Levesley, & Bhakta, 2011). The secondary outcomes were functional independence and upper extremity performance from the "activities" domain, muscle tone from the "body function" domain and quality of life (QoL) from the "participation" domain of the ICF. We included studies if they measured at least one outcome under the "body function" or "activities" domain (Sivan et al., 2011).
We excluded studies if: (a) using RT as an adjunct component to other therapy/intervention (except for usual care); (b) the diagnosis of participants were not clearly described, and/or >50% of participants were comorbid with other injuries, surgical interventions, and/ or serious upper-limb impairments, thereby being undesirable to perform upper-limb training.

| Data extraction
Two authors (MKT, CWC) independently extracted the following information of each included study: study design, characteristics of the participants, treatment conditions in both arms, outcome measures/instruments, main findings, attrition rates, and safety and cost of RT, using a self-developed data extraction form. Disagreements of data extraction were resolved through discussion with another author (YYC).

| Risk of bias assessments of included studies
Two authors (WTC, MKT) independently assessed the overall risk of bias of the included studies by using the Revised Cochrane risk of bias tool (RoB2) (Sterne et al., 2019). The RoB2 covers five domains of bias, including bias arising from randomization; bias due to deviations from intended interventions; bias due to missing outcome data; bias in measurement of outcome; and bias in selection of the reported result. An overall risk of bias judgement was rated for each included study, ranking from low, some concerns to high risk of bias.
The authors resolved the disagreements of assessments with another author (YYC) by discussion.

| Data analysis
For each included study reported with continuous data, we calculated the between-group effect sizes (ESs) by comparing the means between groups at postintervention (<3, 3-7 and >7 months postintervention) using the following formula (Cohen, 1988): d = M 1 -M 2 /SD pooled (M 1 and M 2 refer to the means of both groups). The pooled standard deviation (SD pooled ) was calculated by using the F I G U R E 1 PRISMA flow chart of study selection    (Continues) following formula (Cohen, 1988): SD pooled = √ (SD 1 2 + SD 2 2 )/2. Next, for the outcome being assessed by more than two RCTs, we calculated the pooled effects using SMDs with 95% confidence intervals (CIs) through random effects models accounting for the variations in the use of instruments. When standard errors (SEs), but not SDs, were available, we computed the missing SDs by using the following formula: SD = SE√n (Higgins & Green, 2011). To avoid unit of analysis error in three-arm RCTs, the continuous outcomes of two intervention groups were combined into a single intervention group and compared this with the results of the control group. The magnitude of SMDs can be interpreted as small (0.2), medium (0.5), or large (0.8) (Cohen, 1988). In addition, we assessed the heterogeneity between

TA B L E 1 (Continued)
RCTs by using the I 2 statistic, an I 2 larger than 50% with p < .05 indicates a large and significant heterogeneity across studies. We used the RevMan 5.3 software to conduct the aforementioned analyses.
If the studies reported with skewed data as medians and interquartile ranges, or if there is a lack of RCTs (≤2 per outcome), we excluded them from meta-analyses.

| Quality of evidence
We assessed the quality of evidence as high, moderate, low, or

F I G U R E 4
Forest plot: Comparison of the effect of robotic-assisted therapy and usual care on motor control at post-treatment As percentage (intenƟon-to-treat) Low risk Some concerns High risk

| Ethical statement
This article does not contain any studies with human participants or animals performed by any of the authors; thus, ethical approval is not required.

| RE SULTS
Our search yielded a total of 321 records. After removing duplicates, non-English and brief reports, 186 abstracts were screened.

| Robot-assisted therapy
Seven studies integrated RT with usual care and the time spent  (Hesse et al., 2014;Volpe et al., 2000). The mean attrition rate was 10.5% (range = 0%-28.9%). All studies reported with no serious adverse events, but one study reported that nine patients experienced discomfort and two patients had blisters in the fingertips after RT (Hesse et al., 2014). The costs per patient were 4.15 € for RT and 10.00 € for conventional therapy, respectively (Hesse et al., 2014).
Of the included RCTs, three applied the Action Research Arm Test

| Risk of bias
Four included RCTs were considered as "high risk" of overall bias (Daunoraviciene et al., 2018;Hesse et al., 2014;Masiero et al., 2014;Orihuela-Espina et al., 2016), four were rated as "some concerns" (Volpe et al., 2000;Wolf et al., 2015), and three were judged as "low risk" (Barker et al., 2017;Sale et al., 2014;Stinear et al., 2014) (see Figures 2 and 3). Specifically, six RCTs described random sequence generation with insufficient information on how the allocation was concealed (Daunoraviciene et al., 2018;Hesse et al., 2014;Masiero et al., 2014;Volpe et al., 2000;Wolf et al., 2015); and one study did not implement allocation concealment and was judged as at high risk of bias (Orihuela-Espina et al., 2016). Bias due to deviations from the assigned interventions and missing outcome data were relatively low across studies, given that the attrition mainly occurred in the conventional therapy or usual care (e.g., hospital discharge). Attrition rates at post-treatment (0%-28.9%) and follow-up (0%-18%) were low across studies.

| Effects of robotic-assisted therapy
The effect sizes of RT on the primary and secondary outcomes of each included RCT were tabulated in Appendix S2.

| Motor control
At post-treatment, the overall effect of RT for improving motor control was insignificant when compared to usual care (SMD = 0.18, 95%CI −0.16, 0.51, p = .31, 5 RCTs, 274 participants) (see Figure 4). As shown in the SoF, the quality of evidence using the GRADE approach was downgraded from high to low (2 points) because two of the pooled studies were considered at high risk of bias (Daunoraviciene et al., 2018;Hesse et al., 2014), and the pooled effect was based on wide 95% confidence interval (see Appendix S3) (Daunoraviciene et al., 2018;Dehem et al., 2019;Hesse et al., 2014;Sale et al., 2014;Wolf et al., 2015). Hesse et al. (2014) assessed the treatment effect at 3-month post-intervention, but no significant effect was found.

| Functional independence
At post-treatment, the overall effect of RT for improving functional independence was not significant when compared to usual care (SMD = 0.40,0.95,p = .16,4 RCTs,183 participants) (see Figure 5) (Dehem et al., 2019;Hesse et al., 2014;Villafane et al., 2018;Volpe et al., 2000). In the SoF, we downgraded the quality of this evidence using the GRADE approach from high to low (2 points) because one of the pooled studies (Hesse et al., 2014)  but no significant effect was found.
The quality of this evidence was downgraded from high to low (2 points) because one of the pooled studies (Hesse et al., 2014) was at high risk of bias and the sample size per arm in each included study was small (16-26 per arm; see Appendix S3) (Hesse et al., 2014;Sale et al., 2014;Villafane et al., 2018). Hesse et al. (2014) measured the treatment effect at 6-month poststroke, but no significant effect was shown.

| Quality of life
The overall effect of RT for improving quality of life was not signifi-

| D ISCUSS I ON
We systematically reviewed a total of 11 RCTs to examine whether the effects of RT outweighed usual care for improving motor control, functional independence, upper extremity performance, muscle tone, and quality of life in patients experienced in early stage of poststroke rehabilitation. Nonsignificant effects were found in all our selected outcomes at post-treatment up to 12-month posttreatment and the evidence was generally rated as low-to-moderate quality.
The aforementioned findings were in line with a previous systematic review published in 2012, suggesting the effects of RT was not different from that of dose-matched conventional therapy or usual care regardless the poststroke rehabilitation phase (i.e. acute, subacute, or chronic) (Norouzi-Gheidari et al., 2012). In the present review, all participants of the included RCTs were enrolled within six months after their first episode of stroke. For those who were allocated to the control conditions, the standard care that they received from physiotherapists and/or occupational therapists were often well-evidenced and recommended as clinical treatments (Barker et al., 2017;Hesse et al., 2014;Masiero et al., 2014;Orihuela-Espina et al., 2016;Sale et al., 2014;Stinear et al., 2014;Villafane et al., 2018;Wolf et al., 2015). Hence, when the dose of RT was matched with conventional therapy or usual care or even acted as an adjunct therapy to usual care, the additional benefit of RT for producing high-intensity movement no longer existed. In other words, the gains in motor and functional outcomes in stroke patients at post-treatment appeared to be attributed to highly intensive and repetitive movements, regardless of whether it was delivered by therapists or robotic devices. We also found that adverse events were uncommon and the mean attrition rate at postintervention was low (approximately 10%), indicating that RT is generally safe and acceptable to most participants at the subacute phase of stroke.
As only one study assessed the cost of RT (Hesse et al., 2014), the cost-effectiveness of RT remains uncertain.

| Study limitations and implications
The small number of included RCTs per outcome and large clinical heterogeneity across trials confined us from meta-analyzing the effects of RT on our selected outcomes at post-treatment and follow ups. In addition, most of the included studies were at uncertain or high risk of bias due to insufficient information on allocation concealment and the lack of blinding of outcome assessors. We only included RCTs that were written in English, which may have inadvertently omitted other relevant studies that were published in other languages. Our findings might have limited generalizability to the Asian populations as the reviewed RCTs were conducted in Western countries.
Previous review has indicated the lack of evidence supporting the effects of RT for people affected by stroke within the first three months (Veerbeek et al., 2017). Our findings address this knowledge gap, given that all the participants of the included RCTs were enrolled within six months after their first episode of stroke. Nevertheless, there is still a need for well-designed, adequately powered RCTs to evaluate the benefits and harms of RT in subacute stroke survivors. Apart from providing more standardized and detailed information regarding the components of RT, trial authors shall specify whether they are adopting an intensity-matched and/or duration-matched design for a head-to-head comparison with the control arm. In addition, further research may examine the effect of RT tailoring to patients undergoing different phases of stroke recovery/rehabilitation, and standardize the use of parameters for a better comparability between outcomes across studies. For a more comprehensive evaluation of RT, future studies could include more patient-reported and practical (e.g., safety, adherence, and cost) outcomes (Reeves et al., 2018), explore participants' experiences in receiving RT and investigate the long-term effects of RT. Aside from superiority trials to determine whether RT demonstrates better therapeutic effects when compared with usual care, future studies may explore the possibility of conducting noninferiority or equivalence trials, given that the less labor-intensive RT may offer important advantages over currently available standard care, in terms of improved convenience, better adherence, and lower manpower cost.

| CON CLUS IONS
Robot-assisted therapy produced benefits similar, but not significantly superior to those from usual care, for improving motor con- with low methodological quality, should be interpreted with caution.

ACK N OWLED G M ENTS
This study was supported by the seeding money for collaborative

research, The Hong Kong Polytechnic University [G-YJ05] and The
Chinese University of Hong Kong, HKSAR. We are particularly grateful to all the people who have given us help on our review.

CO N FLI C T O F I NTE R E S T
The authors declare that there are no conflicts of interest.

AUTH O R CO NTR I B UTI O N S
Wai-tong Chien designed this review. Wai-tong Chien, Yuen-yu Chong, and Ho-yu Cheng searched literature. Wai-tong Chien, Yuen-yu Chong, Man-kei Tse, and Cheuk-woon Chien analyzed the data. Wai-tong Chien, Yuen-yu Chong, Man-kei Tse, and Ho-yu Cheng wrote the manuscript. All authors approved the final manuscript as submitted.

PEER R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1002/brb3.1742.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available from the corresponding author upon reasonable request.