We included studies with participants affected by upper limb disability induced by neurological or neuromotor pathologies. The investigated pathologies include muscular dystrophy, spinal muscular atrophy, multiple sclerosis, amyotrophic lateral sclerosis, spinal cord injury, cerebral palsy or in general myopathies that leave patients with muscular weakness. No restrictions have been put in terms of age or sex.

We investigated the use of wearable assistive devices such as orthoses, prostheses, exoskeletons, electrical stimulation devices, and neuroprostheses with the aim to improve upper limbs activities of daily living. An example of an eligible exoskeleton is the Wilmington Robotic Exoskeleton (Wrex). It is a body-powered, four-degrees-of-freedom device that uses linear elastic bands for both balance and assistance of movement against the effects of gravity in three dimensions (Shank, 2017). The considered intervention is the execution of protocol-defined task(s) assisted by the AD.

The comparison has been performed with respect to the same task(s) executed during the intervention assessment, but without the use of the AD. In other words, each participant is him/herself part of both the intervention and comparison groups, given that s/he repeated the same functional task(s) both with (i.e., intervention) and without (i.e., control) the use of arm support (i.e., the AD).

The primary outcome is the measurement of upper limb functional activity level as a comparison between use/non-use of the assistive device, either in terms of externally-assessed improvement (e.g., Fugl-Meyer Assessment scale, ARAT, Performance of the Upper Limb, and other specific scales, as applied by the therapist) or self-perceived functional improvement (e.g., ABILHAND, Canadian Occupational Performance Measure). The secondary outcomes were acceptance measures of the AD. To this aim, we used withdrawal or dropouts from the included studies due to any reason. We investigated the safety of different devices through the incidence of adverse outcomes, such as cardiovascular events, injuries and pain, and any other reported adverse events. Depending on the aforementioned categories and the availability of variables used in the included trials, all review authors discussed and reached consensus on which outcome measures should be included in the analysis.

All designs of studies were accepted.

A comprehensive electronic search was performed to maximize the likelihood of identifying all eligible studies. In order to select all possible variants of keywords, a preliminary backward-forward references searching was performed. Backward references searching involves identifying and examining the references cited in a selected article of interest. Forward references searching, instead, is when a researcher identifies and looks at the articles that cite the identified article of interest after it had been published. A citation tracking and references list screening of all pertinent articles were broadly performed. Keywords were identified for three categories: (i) pathologies; (ii) devices; (iii) section of the body involved in the intervention (Table 1). Medical Subject Headings and free text terms for neuromuscular diseases and ADs for upper limbs were used to capture all research articles in this area. Methodological search filters by designs of study and outcomes were not applied to avoid potential retrieval bias. Multiple searches were performed to check the influence of each keyword on the search and some adjustments were implemented in response to database engines differences. The keywords (Table 1) were combined using Boolean Operators (AND, OR) as in the following search string: OR/1-13 AND OR/14-26 AND OR/27-32 (Lefebvre et al., 2008). We searched in the following bibliographic databases:

In an effort to identify further trials not available in the major databases, we screened references lists of all relevant articles.

The search results were independently screened by two reviewers (MG, VL). Inclusion criteria were as follows: studies which (1) included assistive devices for the upper limbs with aim of supporting activities of daily living, (2) involved participants with neuromuscular or neuromotor diseases, and (3) used valid outcome measure(s) which assessed the same task(s) with and without the use of the device. Exclusion criteria were as follows. Studies which: (1) involved therapy sessions with the aim to permanently restore a lost motor function without further assistance of the device, as opposed to assistive device with the aim to substitute a lost motor function; (2) do not involve the use of patient's own arm (i.e., the device used is not worn on the participant upper limb and do not directly moves patients upper limb to perform a functional movement), (3) make use of implanted devices or (4) use devices that assist only the patients hand. The two authors independently read the titles and the abstracts of identified publications and eliminated obviously irrelevant studies. We obtained the full text for the remaining studies and, according to the predetermined inclusion and exclusion criteria, they where independently ranked by the two authors as relevant, potentially relevant and irrelevant. Discrepancies between reviewers were resolved through discussion between all authors.

Data extraction from the included studies was completed by one of two reviewers (VL), then the second reviewer (MG) checked the accuracy and completeness of extracted data. Any identified discrepancies were discussed by all authors to ensure the accuracy of the extracted data. The data extracted from the included studies were: (i) participants (i.e., number of participants, age, sex, type of disease, inclusion and exclusion criteria); (ii) type of device used; (iii) primary outcome measure(s). Studies' authors were contacted to request more information, clarification, or missing data when needed.

The evaluation of the risk of bias in clinical trials is required to lower the probability to formulate incorrect decisions about treatment effects (Gluud, 2006), since a systematic error or a deviation from the truth in results or inferences (i.e., biases) can lead to either underestimation or overestimation of the true intervention effect (Higgins and Altman, 2008). In this work, the ROBINS-I tool (Risk Of Bias In Non-randomized Studies–of Interventions) has been used, as described in Sterne et al. (2016), particularly effective in evaluating risks of bias of studies that did not use randomization to allocate interventions. ROBINS-I fundamental underlying principle lays on the comparison between the risks of bias associated with the current evaluated study and a target RCT, hypothetically conducted on the same participants group, even if this RCT may not be feasible or ethical (Schünemann et al., 2018). The ROBINS-I tool includes the evaluation of seven domains through which bias might be introduced into a non-randomized study: (1) bias due to confounding; (2) bias in selection of participants into the study; (3) bias in classification of interventions; (4) bias due to deviations from intended interventions; (5) bias due to missing data; (6) bias in measurement of outcomes (or detection bias); (7) bias in selections of the reported results. The ROBINS-I tool includes a series of signaling questions within each domain in order to facilitate judgments about the risk. The categories for risk of bias judgments are Low risk, Moderate risk, Serious risk and Critical risk. The risk of bias is firstly judged for each domain, and then assessed overall across the study. Two authors independently assessed risk of bias of the included studies (VL, AA), with disagreements between reviewers resolved through discussion between all authors. Two authors (MG and AP) were co-authors of one included trial (Ambrosini et al., 2014); and they did not participate to the quality assessment for this study.

The primary outcome variables of interest were treated as continuous data and entered as means (m_i) and standard deviations (σ_i), where i is the index for each included study. Since the studies assessed the same outcome (i.e., upper limbs functional improvement), but measured it with different outcome measures, the standardized mean difference (SMD_i) with 95% confidence interval (CI_i) was used as summary statistic, in order to standardize the results of the studies to a uniform scale before they can be combined. The SMD_i expresses the size of the intervention effect in each study, relatively to the observed variability (Deeks et al., 2008), and it is expressed as the difference in mean outcome between groups divided by the pooled standard deviation of outcome among participants (Lakens, 2013), as follows (Equation 1):

where m_T, σ_T, N_T and m_C, σ_C, N_C are the mean outcome, the standard deviation and the sample size of the Treatment (T) and the Control (C) group, respectively. A SMD_i of zero means that the treatment and the control group have equivalent effects. Since improvements are associated with higher scores on the outcome measures, SMD_i > 0 indicates the degree to which treatment is more effective than control, and SMD_i < 0 indicates the degree to which treatment is less effective than control.

This SMD_i (Equation 1) overestimates population effect sizes when sample sizes are small, and therefore it was corrected with the Hedges' g as follows (Equation 2) (Hedges and Olkin, 1985):

The within-group standard deviation of g has been calculated as standard error (SE_i) as follows (Equation 3) (Hedges and Olkin, 1985):

The confidence interval CI_i for each g_i was given by Equation (4) (Hedges and Olkin, 1985):

where z1-α2 is the 1-α2 quartile of the normal distribution. In this meta-analysis, α was set to 0.05.

In the second stage of the meta-analysis, a summary intervention effect estimate θ was calculated as a weighted average of the intervention effects estimated in the individual studies defined as in Equation (5) (Marín-Martínez and Sánchez-Meca, 2010):

where g_i is the intervention effect estimated in the i^th study, w_i is the weight given to the i^th study, and k indicates the number of studies included in the analysis.

Weights were estimated with the inverse-variance method (Deeks et al., 2008). Thus, studies with larger sample sizes, which have smaller standard errors, are weighted more than studies with smaller sample sizes, which have larger standard errors. This choice minimizes the imprecision (uncertainty) of the estimated effect size, and it is obtained as follows (Equation 6):

where τ² is the between-study variance derived with the method proposed by DerSimonian and Laird (Equation 7) (DerSimonian and Laird, 1986):

Q is the heterogeneity statistic (Equation 12), and c is given by Equation (8):

wiFE is the weighting factor for the i^th study assuming a fixed-effects model (wiFE=1σi2) (Marín-Martínez and Sánchez-Meca, 2010). The standard error (SE) of the summary intervention effect was then computed as (Equation 9):

Finally, the two-sided confidence interval for θ was obtained with the following (Equation 10), with α = 0.05:

If a study included more than an outcome measure that assessed upper limb functionality (e.g., range of motion at shoulder, elbow and wrist levels) or includes two or more groups, sample sizes, means and standard deviations were properly merged, as detailed in Table 2 (Deeks et al., 2008).

Participants' acceptability of the ADs was investigated analyzing dropouts and withdrawal through the risk difference index (RD) along with its corresponding 95% CI. The risk difference is the difference between the proportions of individuals with the effective assessment of the outcome of interest (i.e., observed risks) in the treatment group and in the control group (Equation 11) (Deeks et al., 2008):

where S_T and S_C are the number of dropouts and withdrawal in the treatment (T) and in the control group (C) respectively, while N_T and N_C represent the total number of participants in each group.

Between-studies variability (i.e., differences among the population effect sizes estimated by individual studies) might be a source of heterogeneity. If there is statistically-significance between-studies heterogeneity, moderator variables can be examined to explain this variability (e.g., participants characteristics, outcome measures, treatment conditions, study designs, etc.) (Huedo-Medina et al., 2006). The statistical test usually applied in meta-analysis for determining whether there is true heterogeneity among study effects sizes is the Q test, proposed by Cochran (Cochran, 1954; Hedges and Olkin, 1985) and defined as (Equation 12):

where ES^FE is the effect size assuming a fixed effect model (i.e., Equation 5, with τ = 0). A shortcoming of the Q statistic is that it has poor power to detect true heterogeneity among studies, when the meta-analysis includes a small number of studies (Huedo-Medina et al., 2006; Fletcher, 2007), as in the current case. To solve this issue, Higgins and Thompson (2002) have proposed the I² index, which quantifies the extent of heterogeneity from a collection of effect sizes by comparing the Q value to its expected value assuming homogeneity, that is to its degrees of freedom (df = k − 1), as in Equation (13):

The I² index can be interpreted as the percentage of the total variability in a set of effect sizes due to true heterogeneity (i.e., between-studies variability) rather than due to sampling error (i.e., chance). I² ranges from 0 to 100%, with 0% indicating that statistical homogeneity exists. Higgins and Thompson (Higgins and Thompson, 2002) proposed a tentative classification of I² values with the purpose of helping to interpret its magnitude. Percentages of around 25% (I² = 25), 50% (I² = 50), and 75% (I² = 75) would mean low, medium, and high (Sedgwick, 2013). Therefore, if a cluster of studies presents a characteristic of non-homogeneity, it is worthy to investigate its source.

Given an expected heterogeneity between the studies involved in this research, two sub-groups analyses were hypothesized. As a first hypothesis, the studies were split into two sub-groups according to the type of device involved: (i) studies that involved a passive device [i.e., non-actuated or passively actuated with counterweights, springs, or elastic bands to passively compensate for the impact of gravity on the arm (Van der Heide et al., 2014)] and (ii) studies performed with an active device (i.e., not dependent on pre-stored mechanical energy, equipped with sensors and actuators). If a study involved the use of both active and passive devices, it has been split into two sub-studies: one including the patients that used passive and one with patients that used active devices.

Another possible source of heterogeneity was identified by the type of primary outcome measure employed, and namely: (i) studies characterized by a self-perceived outcome measure (i.e., outcome measures derived by the perception of the patient him/herself according, for example, to the perceived functional benefit or the difficulty in accomplishing an exercise, e.g., Canadian Occupational Performance Measure) and (ii) studies with externally-assessed outcome measures (i.e., all rigorous and methodological scales performed by clinicians or through a system measurement, e.g., Fugl-Meyer scale, range of motion).

Within the sub-groups analyses, treatment effects in sub-groups have been compared by a test of interaction that investigated whether the effect size of the intervention in the primary outcome measure varied between the sub-groups. In particular, for both sub-groups analyses, differences in treatment effects have been evaluated with the I² (Equation 13). In particular, I² = 0% indicates no effect between sub-groups (Sedgwick, 2013).

The results of all eligible studies were pooled in order to present an overall estimate of the effect of ADs in supporting activities of daily living. For all statistical analyses, Cochrane Review Manager software (RevMan 5.3) was used. We calculated the overall effect size using a random-effects model, regardless of the level of heterogeneity (Mehrholz et al., 2015). Clinical diversity and heterogeneity did not contribute to the decision on when to pool trials, but we described clinical diversity, variability in participants, interventions, and outcomes studied in Tables 3, 4. A z-test was applied to overall estimates of the effect sizes with the null hypothesis that there was no statistically significant difference between the treatment (i.e., users using AD) and the control group (i.e., users not using AD). The null hypothesis of no statistically difference was rejected if p-values were <0.05.

The sample sizes in the trials ranged from three participants, in Ambrosini et al. (2014);Kooren et al. (2015), to 55 participants, in Gunn et al. (2016). The median of the sample size is 9, with 13 as interquartile range. Sample sizes for all included studies are detailed in Table 3.

The characteristics of the 184 participants, grouped for each study, are listed in Table 4. The mean age of participants in the included studies ranged from 8 years, in Shank (2017), to 58 years, in Iwamuro et al. (2008). The percentage of involved males was higher (65% with 95% CI= [63.8%-66%]). Participants across all trials had different diagnoses of neuromuscular diseases. The mostly investigated disease was Muscular Dystrophy (MD), included in 9/14 trials, and involving the 24% of total patients. Other included pathologies were Arthrogryposis Multiplex Congenital (AMC), Stroke resulting in hemiparesis (i.e., hemiparesis stroke), Spinal Muscular Atrophy (SMA), Hemiparesis, Cerebral Palsy (CP), Amyotrophic Lateral Sclerosis (ALS), Multiple Sclerosis (MS), and Spinal Cord Injury (SCI). Only eight studies (Rahman et al., 2007; Bastiaens et al., 2011; Ambrosini et al., 2014; Jan Burgers, 2015; Kooren et al., 2015; van der Heide and de Witte, 2016; van der Heide et al., 2017; Estilow et al., 2018) provided information about baseline deficit of arm motor function through the Brooke scale (Brooke et al., 1981), the Manual Muscle Test (Ciesla et al., 2011) or the Motricity Index (Bohannon, 1999). Mean values and standard deviations are reported in Table 4. For inclusion and exclusion criteria of each included study, see Characteristics of included studies in Supplementary Material.

All included studies involved the use of a wearable upper limb AD, as specified in the studies inclusion criteria. The included ADs are detailed for each included study in Table 3. Four studies concerned the use of an active AD (Lund et al., 2009; Bastiaens et al., 2011; Ambrosini et al., 2014; Peters et al., 2017), 7 investigated the use of passive ADs (Sanchez et al., 2004; Rahman et al., 2007; Iwamuro et al., 2008; Kooren et al., 2015; Gunn et al., 2016; Shank, 2017; Estilow et al., 2018), and 3 investigated the use of both active and passive ADs (Jan Burgers, 2015; van der Heide and de Witte, 2016; van der Heide et al., 2017). The most used (passive) AD was Wrex, which was employed in six studies (Sanchez et al., 2004; Rahman et al., 2007; Iwamuro et al., 2008; Gunn et al., 2016; Shank, 2017; Estilow et al., 2018). Participants were able to familiarize or to use the device for a variable period of time, ranging from a single session in a controlled environment such as the clinical setting (Sanchez et al., 2004; Iwamuro et al., 2008; Bastiaens et al., 2011; Ambrosini et al., 2014; Jan Burgers, 2015; Kooren et al., 2015; Peters et al., 2017; Estilow et al., 2018), to regular use at home for a mean range of four months to 25 months (Lund et al., 2009; Gunn et al., 2016; van der Heide and de Witte, 2016; Shank, 2017). However, the outcome measure assessment with and without the device did have a maximum inter-timing of 2 weeks (Shank, 2017).

The included trials compared the ability of the patient to perform activities of daily living with and without AD assistance. The intensity of treatment (in terms of duration of use of the AD) ranged from one single session in one day (Sanchez et al., 2004; Iwamuro et al., 2008; Bastiaens et al., 2011; Ambrosini et al., 2014; Jan Burgers, 2015; Kooren et al., 2015; Peters et al., 2017; Estilow et al., 2018), up to a mean of 25 months of regularly use of the AD (Shank, 2017). A detailed description of comparison is provided in Characteristics of included studies in Supplementary Material.

Primary outcome measures used in the included studies, and analyzed in this work are detailed in Table 3. Possible secondary outcome measure(s) which may have been assessed have not been included in this study. Analyzed outcome measures included self-perceived outcome measures, and externally-assessed outcome measures. We classified as self-perceived the following outcome measures. (i) Five-point Likert scale (Gunn et al., 2016), where patients are asked to answer to ten questions on their functional ability without and with the use of the AD. The answers ranged from “performing the task without any difficulty” to “unable to do it.” (ii) Canadian Occupational Performance Measure (Shank, 2017), where patients are asked to evaluate their ability to perform 5 selected activities of daily living from 1 (“completely unable to perform it”) to 10 (“able to perform it very well”). (iii) Perceived Functional Benefit (van der Heide and de Witte, 2016), computed as the difference in patients self-evaluation of ability to perform activities of daily living with and without the AD in 51 items, and (iv) the Individually Prioritized Problem Assessment (IPPA) (Lund et al., 2009). On the other hand, the included externally-assessed outcome measures were the following. (i) Upper extremity section of Fugl-Meyer scale (Sanchez et al., 2004; Peters et al., 2017), designed to assess motor functioning, balance, sensation and joint functioning in post-stroke patients. (ii) Jebsen Hand Function Test (Rahman et al., 2007), which calculates the completion time of seven specific upper limbs tasks (e.g., writing or moving light and heavy objects). (iii) Fraction Of Reach scale (Iwamuro et al., 2008), computed by comparing the minimum distance to the target achieved by the participant to the Euclidean distance between the starting hand posture and the target, while performing reaching of objects placed in 12 different positions at the limits of the patient-specific reaching space. (iv) Range Of Motion (ROM) of the shoulder and elbow joints (Bastiaens et al., 2011; van der Heide et al., 2017; Estilow et al., 2018), (v) Action Research Arm Test (ARAT) (Jan Burgers, 2015), (vi) the Root Mean Square Error (RMSE) between the actual and the target angle (Ambrosini et al., 2014). (vii) Performance of the Upper Limb scale (PUL) (Kooren et al., 2015), a new tool recently designed for specifically assessing arm functionality in DMD patients through 22 items at the shoulder, elbow and wrist/fingers levels.

All included studies were cohort studies according to the definition proposed by Mathes and colleagues (Mathes and Pieper, 2017), i.e., studies that contain sufficient data to conduct a re-analysis and thus are appropriate for inclusion in systematic reviews.

The sub-groups analysis clustered the studies involving active devices and studies performed with passive devices, therefore investigating if the kind of device modifies the intervention effect (see Supplementary Material). The active devices group obtained an effect size 1.16 (95% CI = [0.73 1.60]) computed over 51 participants. The result of the Z-test of the null hypothesis that there was no effect was Z= 5.25 (p < 0.00001). The passive devices group included 133 subjects and it was characterized by an effect size of 0.99 (95% CI = [0.57 1.42]). The result of the Z-test was Z= 4.56 with an associated p < 0.00001. No differences between the two sub-groups have been detected (I²= 0%). This means that the type of device used (active or passive) does not modify the intervention effect.

The sub-groups analysis clustered the studies involving the use of self-perceived outcome scores and studies performed using externally-assessed outcome measures, therefore investigating if the typology of outcome measure assessment modifies the intervention effect (see Figure S2). The sub-group including studies with self-perceived outcome measures was composed by 106 subjects and it was characterized by an effect size equals to 1.38 (95% CI = [1.08 1.68]). The result of the Z-test of the null hypothesis that there was no effect was Z= 8.92 (p < 0.00001). The sub-group including studies with externally-assessed outcome measures was composed by 78 subjects, and the effect size resulted to be 0.77 (95% CI= [0.41 1.11]). The result of the Z-test was Z= 4.40 with a p < 0.00001. The I² index significantly decreases for both groups with respect to the general analysis: in the group of studies characterized by a self-perceived outcome measure, it was equal to 0%, while in the other group to 3%. These results indicate, respectively, absence and a very low level of heterogeneity. The test for sub-groups differences revealed a significant interaction, with I² equals to 85.6% (95% CI = [42% 96.4%]). Therefore a statistically significant sub-groups effect was detected, and it can be derived that the considered covariate significantly modifies the treatment effect.