Effectiveness of robot therapy on body function and structure in people with limited upper limb function: A systematic review and meta-analysis

Robot-Assisted Therapy (RT) is an innovative approach to neurological rehabilitation that uses intensive, repetitive, interactive, and individualized practice. This systematic review aimed to investigate the effectiveness of RT on the body function and structure of people with upper limb impairments (PROSPERO registration: CRD42017054982). A search strategy conducted on seven databases identified randomized controlled studies. Methodological quality was assessed using the PEDro scale. When possible, the data were pooled, the strength of evidence was assessed using the GRADE system, and the effect sizes were assessed using Cohen coefficient. Subgroup analyses investigated the impact on the estimated effects of the following parameters: methodological quality; portion of the assessed upper limb; duration of stroke; and intervention dose and duration. Thirty-eight studies involving 1174 participants were included. Pooled estimates revealed small effects of RT on motor control and medium effects on strength compared with other intervention (OI) at a short-term follow-up. Standardized differences in means were as follows: 0.3 (95% CI 0.1 to 0.4) and 0.5 (95% CI 0.2 to 0.8). Effects at other time points and on other investigated outcomes related to body function and structure were not found (p>0.05). The strength of the current evidence was usually low quality. Subgroup analyses suggested that the methodological quality, and duration and dose of RT may influence the estimated effects. In conclusion, RT has small effects on motor control and medium effects on strength in people with limited upper limb function. Poor methodological quality, and lower treatment dose and duration may impact negatively the estimated effects.

Introduction Upper limb motor impairments following a neurological disorder are common and may lead to function limitations, dependence and poor quality of life among the affected people [1]. There are many rehabilitation programs aiming to promote the function, independence and social reintegration of these affected people [2]. These programs include constraint-induced movement therapy, electromyographic biofeedback, mental practice with motor imagery, repetitive task training, functional electrical stimulation and Robot-Assisted Therapy (RT) [3,4,5].
RT is an innovative approach to neurological rehabilitation that involves intensive, repetitive, interactive, and individualized practice [6]. The use of RT for upper limb disorders dates to the 1990s. Since then, a number of robotic devices have become commercially available to clinics and hospitals worldwide [7].
Previous reviews have suggested that RT improves upper limb motor control and muscle strength [8,9,10,11,12]. However, these studies drew limited conclusions about the effectiveness of RT on the body function and structure of people with upper limb impairments. It was not possible to specify comparisons; the use of RT alone or combined with other interventions was compared with minimal or other interventions. Other limitations included few investigated outcomes related to body function and structure in individuals with stroke, absence of protocol registration and assessment of the strength of evidence [11], language restriction for the included studies, and absence of medium-and long-term effects [12].
It also remains unknown whether the estimated effects of RT are impacted by the portion of the assessed upper limb (i.e., proximal shoulder-elbow level or distal hand-wrist level) [9,10], treatment dose and/or duration [11,12], or the methodological quality of the studies. Therefore, the aim of this systematic review was to investigate the effectiveness of RT on outcomes related to body function and structure of people with upper limb impairments at short-, medium-and long-term follow-ups. The potential impacts of the portion of the assessed upper limb, duration of stroke, treatment dose and/or duration, and methodological quality were also investigated.

Search strategy and inclusion criteria
The protocol of this review was prospectively registered at PROSPERO (CRD42017054982). The search for relevant studies was conducted in PEDro (Physiotherapy Evidence Database), EMBASE (Excerpta Medica Database), Medline (Medical Literature Analysis and Retrieval System Online), CINAHL (Cumulative Index to Nursing and Allied Health Literature), Cochrane (Cochrane Collaboration), AMED (Allied and Complementary Medicine Database) and Compendex (Compendex Engineering Index) without language or date restrictions. In addition, a hand search was conducted in reference lists of previous reviews in this area. The search terms were related to "Robot-Assisted Therapy" (robotics, orthotic devices, bionic device, exoskeleton, robotic aided therapy, therapy computer-assisted, robot-assisted, robotics-assisted, self-help devices, robotic device, dynamic orthotic device, robot-mediated therapy, robot-supported, computer-assisted instruction, computer aided, computer-aided design, computer assisted, artificial limb, rehabilitation robotics, human-robot interaction, robotaided rehabilitation, robotic rehabilitation, orthosis, taping, splinting, assistive technology devices, assistive device therapy), "upper limb" (upper extremity, arm, arm injuries, hand, hand injuries, shoulder, shoulder injuries, elbow, axilla elbow, forearm injuries, forearm, finger, finger injuries, wrist injuries, wrist) and "randomized controlled trial" (random allocation, double blind method, single blind method, placebo, random, controlled clinical trial, clinical trial, comparative study, evaluation study, follow-up study, prospective study, crossover studies). See S1 Appendix in the Addenda for the detailed search strategy.
This review included prospective randomized or quasi-randomized controlled studies including inpatients and outpatients from any primary, secondary or tertiary care setting and community. Studies were eligible if they included participants of both sexes, regardless of age, with limited upper limb function caused by stroke. The intervention of interest was RT, which was defined as the application of any electronic, computerized control system connected to mechanical devices designed to perform human functions. Studies were eligible if RT was compared with minimal intervention or other intervention (OI). We defined minimal intervention as when the control group received no intervention, received sham or placebo intervention, or was on a waiting list. We considered any other active intervention that was not robotic therapy, such as conventional therapy and physical therapy. Studies investigating additional effects of RT were also included. The outcomes of interest in this review were those related to body function and structure, according to the International Classification of Functioning, Disability and Health [13]. We considered body function as the physiological functions of body systems, including psychological function and body structure, i.e., anatomical parts of the body, such as organs, limbs and their components [14].

Selection of studies
After removing duplicate studies, the relevant retrieved titles and abstracts were selected. Then, we assessed the potential full texts, and studies fulfilling the eligibility criteria were included.

Methodological quality assessment
We assessed the methodological quality of the included studies using the 0 to 10 PEDro scale, with higher scores indicating greater methodological quality. Disagreements were resolved by consensus. When available, we used scores already on the PEDro database (http://www.pedro. org.au/).

Data extraction
We extracted data on the following characteristics of the included studies: number of participants; mean age; percentage of female; cause of the upper limb disorder and its duration; evaluated joints; type of RT; comparison groups; frequency and total duration of intervention; and outcome measures.
The included studies investigated many different outcomes related to body function and structure. For the feasibility of this review, we arbitrarily decided to include the following five most investigated outcomes in the literature: motor control; strength; spasticity; range of motion; and pain. When a given study evaluated these outcomes with more than one instrument, we considered the most consistent instrument among the included studies. When a given study investigated two different RT groups [15][16][17][18][19][20][21][22][23][24][25][26][27], we considered both groups, consistent with previous reviews in this area [12,28]. We extracted data for the complete upper limb, and we separately considered the proximal (i.e., elbow and shoulder) and distal (i.e., wrist and hand) portions of the assessed upper limb, as suggested by previous reviews [9,10]. When more than one measurement was available for the proximal and/or distal upper limb, we considered the elbow and wrist due to their greater consistency among the included studies and clinical implications [29].
The following outcome data were included: sample size, mean and standard deviation (SD) for each group were extracted at the short-, medium-and long-term follow-ups: 3 months after baseline for short-term; > 3 months but < 12 months after baseline for medium-term; and ! 12 months after baseline for long-term. When multiple time points were available within the same follow-up period, the time point closer to the end of the intervention was used for short-term follow-up, that closer to 6 months was used for medium-term follow-up and that closer to 12 months was used for long-term follow-up. SDs were not available in certain included studies, and in those cases, the SDs were imputed from the 95% confidence interval (CI), standard error (SE), p value, interquartile range and average from other included studies with similar sources of participants. See S2 Appendix in the Addenda for the detailed extracted data.

Data analysis
Data for each outcome were pooled when there was sufficient homogeneity among studies. Homogeneity among studies was assessed using I 2 statistics. Low heterogeneity was defined as if I 2 50%, and moderate to high heterogeneity was defined as I 2 > 50% [30]. Pooled effects were estimated using standardized mean differences (SMDs) with 95% confidence intervals (CI). A fixed-effects model was used to conduct the meta-analysis when I 2 50%, and a random-effects model was used to conduct the meta-analysis when I 2 > 50%. To judge the clinical relevance of RT, the effect size was assessed using Cohen's d coefficient according to the following parameters: 0.2 as small effect, 0.5 as medium effect, and 0.8 as large effect [31]. A funnel plot was used to investigated publication bias when at least 10 studies were pooled [30]. The meta-analysis was performed using the software Comprehensive Meta-Analysis, version 3.3.070.
The GRADE (Grading of Recommendations Assessment, Development and Evaluation) system was used to summarize the overall quality of evidence for each outcome [32]. We rated evidence from the high-quality level and downgraded it one point if one of the following prespecified criteria was present: low methodological quality (average PEDro score < 6); inconsistency of estimates among pooled studies (I 2 > 50%) or when its assessment was not possible (no pooling); indirectness of participants (over 50% of the studies did not describe inclusion criteria); and imprecision (pooling < 300 participants for each outcome) [33].
Subgroup analyses were used to investigated the impact of the following on estimated effects: 1) poor methodological quality (i.e., removing studies with score of five or less out of ten on PEDro scale); 2) investigated portion (proximal and distal portions of the upper limb); 3) duration of stroke (duration of the current episode 6 months and > 6 months); and 4) treatment dose and duration. For dose, analyses investigated whether the effects of studies providing the same amount of intervention differed from those providing different amounts of intervention for RT and control groups. For duration, analyses investigated whether the effects of studies providing > 20 sessions differed from those providing 20 sessions. The impact of dose and duration was also investigated using total volume (i.e., number of sessions x time per session in hours), dichotomized into studies with total volume > 20 hours of intervention and those ones with 20 hours. We arbitrarily decided on these cut-offs because they were the most frequently used cut-offs in the included studies. Subgroup analyses were conducted to compare RT and OI at the short-term follow-up because this was the most investigated followup.

Characteristics of studies
The characteristics of the included studies are presented in Table 1. All included studies were prospective randomized controlled studies published in English between 1997 and 2015. The 38 original studies enrolled 1174 participants of both sexes, with a mean age ranging from 51.2 to 57.8 years. The cause of the upper limb disorder was stroke, with 24 of the 38 studies including people with chronic episodes of this health condition.
The duration of RT ranged from 2 [25] to 20 [57] weeks, and the frequency per week varied from 2 [16,18] to 6 [57] days. The time spent per session of intervention ranged from 0.2 [40] to 2 [57] hours. The total volume of intervention per week (i.e., number of sessions per week x duration of each session) ranged from 1 [54] to 12 [57] hours. The total number of sessions ranged from 12 [16,25,45] to 120 [57], with most studies ranging from 20 to 24. The total duration of the intervention ranged from 6 [45] to 240 [57] hours. On average, RT sessions occurred three times per week with a total duration of treatment of 8 weeks. For some studies, primarily those comparing RT to minimal intervention, detailed information on the dose and duration of the intervention was not available [35,36,48,54,56].
Recruited and the informed consent was obtained in compliance with Veterans Affairs and Stanford University.     WAM control program [56,57], and one study [51] used the Motricity Index (MI). To homogenize the data, Newton and kilogram-force values were converted into pounds. We arbitrarily decided to use pounds because this was the most consistent unit among the studies. Five studies evaluated range of motion using three different measures. Of these 5 studies, three studies (60.0%) [16,18,50] assessed range of motion using goniometer, one study [57] used the WAM control program, and one study [41] used the mean distance between a marker placed on the participant's wrist and five targets. Six studies evaluated pain using three different instruments. Of these 6 studies, three studies (50%) [16,18,26] used the Visual Analogue Scale (VAS), two studies (33.3%) [25,56] used the Pain Scale of Fugl-Meyer, and one study [34] used the Chedoke McMaster Stroke Assessment Pain Inventory Scale.

Methodological quality of the included studies
The detailed methodological quality of the included studies is presented in S1 Table on the Addenda. The mean methodological quality of the 38 studies was 6.0 on the 0 to 10 PEDro scale. Most studies included the following: between-group comparisons (n = 37 studies, 97.3%); precision and variability estimates (n = 35 studies, 92%); group similarity at baseline and assessor blinding (n = 32 studies, 84.1%); and outcome measures for at least 85% of participants (n = 27 studies, 71%). Concealed allocation was presented in 12 studies (31.5%), and intention-to-treat analysis was presented in 10 studies (26.3%). The primary methodological quality issues were related to the blinding of participants and therapists, which was included in only three (7.9%) and two (5.2%) studies, respectively.
Robot-assisted therapy versus other intervention. The pooled estimates showed small effects of RT on motor control and medium effects on strength compared with OI at the shortterm follow-up. The SMDs were, respectively, 0.3 (95% CI 0.1 to 0.4); and 0.5 (95% CI 0.2 to 0.8). Detailed analyses are presented in Figs 2 and 4. There is high-and very low-quality evidence showing that RT has effects on motor control and strength, respectively, compared to OI at the short-term follow-up.
Pooled estimates showed no effects of RT on spasticity, range of motion and pain at shortterm follow-up, or on motor control, spasticity and strength at medium-term follow-up, when compared with OI (p> 0.05). Detailed analyses are presented in Figs 2 to 4 and S1 and S2 Figs. The strength of the evidence ranged from low-to very low-quality. Additional effects of robot-assisted therapy over other interventions. Pooled estimates showed no additional effects of RT on motor control, spasticity and pain at short-term followup, or on motor control and spasticity at the medium-term follow-up in stroke compared with stand-alone OI (p> 0.05). Detailed analyses are presented in Figs 2 and 3 and S1 and S2 Figs. The strength of the evidence ranged from low-to very low-quality.

Subgroup analysis
We investigated the impact of methodological quality, portions of the assessed upper limb, duration of stroke, and treatment dose and duration on the estimated short-term effects of RT compared with OI (see detailed subgroup analyses in S3 Fig). Methodological quality, and dose impacted the estimated effects for motor control. Poor methodological quality, and lower treatment dose and duration may impact negatively the estimated effects.

Discussion
This review included 38 studies comparing the efficacy of RT with minimal intervention or OI, and investigating additional effects of RT combined with OI on body function and structure in people with upper limb limitations caused by stroke. RT has small effects on motor control and medium effects on muscle strength. Moreover, the methodological quality, portion of the upper limb, treatment dose, duration and volume may impact the estimated effects. The current low-quality evidence suggests that estimated effects are likely to change with future high-quality studies, and effects are not consistent among outcomes related to body function and structure.
Our findings revealed that compared with OI, RT has statistically significant but small effects on motor control and medium effects on strength. These short-term findings comparing RT with OI are consistent with other reviews on stroke. Veerbeek et al. [12] showed a small improvement on motor control and muscle strength and no effect on spasticity. Prange et al. [8] also found improvement on motor control at the short-term follow-up compared to conventional rehabilitation.
Previous reviews [8,10] did not investigate the medium-and long-term effects. Despite this, Norouzi-Gheidari et al. [10] suggested no effects on motor control at medium-term follow-up when the same doses of RT and OI were used. Our results were consistent with those of the previous study and suggest that upper limb motor control improvement occurs within the short-term ( 3 months after stroke) [59]. Moreover, Prange et al. [8] found long-term effects on motor control when RT was used compared to OI. Their findings were not consistent with our results, and a possible explanation is that the previous review [8] included poor-quality studies (i.e., non-randomized controlled studies). As suggested by Norouzi-Gheidari et al. [10], future high-quality studies should confirm our findings because current evidence for the estimated effect is very low and likely to change.
When comparing the efficacy of RT with minimal intervention at different time points, despite trends favouring treatment, the current low-quality evidence showed no significant effects. Susanto et al. [52] stated that there is an insignificant effect, but there are few studies, and they have small samples.
The average methodological quality of the 39 included studies was 6 points on the 0 to 10 PEDro scale, ranging from 2 to 8 points. This quality was consistent with that reported by Veerbeek et al. [12], with an average quality of 6.0 points. The primary methodological issues were related to blinding, which is expected due to the difficulty of fulfilling these criteria in the area of RT.
A subgroup analysis showed that the portions of the assessed upper limb influence the estimated effects only for range of motion. These findings were not consistent with other studies [9,10,12] and were similar to those reported by Mehrholz et al. [11]. Therefore, there is no consensus on the impact of portions of the assessed upper limb, and current evidence is low but likely to change with further high-quality studies that include larger samples. Subgroup analysis also suggested a greater effect on motor control in chronic stroke, similar to most recent review [28]. Subgroup analysis also suggested that when conventional therapy (CT) is used at the same dose as robot-assisted therapy (RT), there is a significant effect on motor control, unlike the findings reported by Kwakell [9]. Subgroup analysis also suggested an impact of the number of sessions and treatment volume on some estimated effects. Greater number of sessions seems to impact motor control, and greater treatment volume seems to impact motor control. The effect of greater treatment dose was suggested by Lohse et al. [60]; however, time as a dose representation is a rather crude estimate and provides no evidence of the actual amount of movement or types of movement, nor does this representation take into account periods of inactivity or rest [61]. In this regard, a previous review indicated that although there is no consensus, the minimum dose should be at least 16 hours of training [62].
This review and the current literature have some potential limitations. First, there is only a small number of randomized controlled studies that mainly investigate range of motion and pain, and few studies comparing RT with minimal intervention at different time points. Second, studies typically had small sample sizes. Third, subgroup analyses did not investigate impact of types of RT devices, and dichotomization was a potential limitation to get full information regarding the impact of the investigated factors.
Further high-quality randomized controlled studies with larger sample sizes are warranted to elucidate more precise effects of RT on outcomes related to body function and structure, especially the long-term effects. Studies comparing RT with minimal intervention should be conducted, since the current evidence is very low-quality. These studies should report the treatment dose and duration. Future studies should investigate whether RT is effective on psychological factors, and other outcomes related to the activity and participation domains in the ICF. In addition, it is imperative to conduct studies on the cost-effectiveness of RT.
In conclusion, RT has small effects on motor control and medium effects on strength in people with limited upper limb function caused by stroke. Poor methodological quality, and lower treatment dose and duration may impact negatively the estimated effects. Clinicians consider this approach because it has few or no side effects. In addition, there may be longterm financial benefits to employing therapeutic robots. The current low-quality evidence suggests that the estimated effects are likely to change with future high-quality studies and that the effects are not consistent among outcomes related to body function and structure.