Adjusted Indirect and Mixed Comparisons of Interventions for the Patient-Reported Outcomes Measures (PROMs) of Disabled Adults: A Systematic Review and Network Meta-Analysis

This systematic review adopted the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Statement (PRISMA) guidelines and used the method of network meta-analysis to compare the effects of different types of interventions from different perspectives which were abilities of daily life activity, psychological health, social functioning, and overall life quality. The eligibility criteria were: (1) Participants were adults above 18 years old with disabilities; (2) Interventions could be classified into active exercise, passive therapy, psychological education, psychosocial support program, multi-disciplinary program, and usual care; (3) Outcomes should be the patient-reported outcome measures (PROMs) that could be classified into abilities of daily life activity, psychological health, social functioning, and overall life quality; (4) Randomized designed and published in English. The keywords and their search field were: (1) “people with disabilities/disability”, “disabled”, “handicapped”, or “disable people” in titles or abstracts; (2) AND “randomized” or “randomised” in titles or abstracts; (3) NOT “design”, “protocol”, or “review” in titles. After searching in databases of Medline (EBSCO), PubMed, CINAHL, and Ovid, 16 studies were included. As a result, active exercise and passive therapy are most likely to be the best interventions for overall life quality, psychological education and passive therapy are most likely to be the best interventions for abilities of daily life activity, and psychosocial support programs are most likely to be the best intervention for psychological health and social functioning.


Introduction
Disability, whose definition is the inability to engage in complex or gainful activity due to physical or mental impairment, is a result of the interactions between the impaired individual and the environment in which they function. The determination of disability includes assessments of impairment, associated functional impairments, and any impact on the ability to perform activities of daily living and work. Evaluation of disability is an important aspect of social and clinical care. Accurate evaluation is significantly meaningful to the wellbeing of both patients and society, giving the impact of disability status on financial remuneration, return to work, personal and workplace productivity, and access to existing and future health care needs [1].
However, assessment of disability is complex, variable, and challenging even among clinicians experienced in disability determination. Many factors give rise to these challenges [2], the most important of which is that the determination of disability requires a

Participants/Population
All studies whose participants were adults above 18 years old with disabilities would be included in this review. The definition of disability was following the description given by the SSA, which was that "the inability to engage in any substantial, gainful activity because of a medically determinable physical or mental impairment(s), which could be expected to result in death or which has lasted or can be expected to last for a continuous period of not less than 12 months" [4].

Intervention(s)
As the principle of the re-classification of interventions, all the interventions in the included studies would be classified into eight classes: (1) Passive therapy (all kinds of physiotherapist-led interventions such as vibration therapy, massage, and passive stretching); (2) Active exercise (resistance training, exercise with or without supervision or assistance); (3) Usual care (treatment as usual, waiting list, and no-treatment); (4) Drug treatment; (5) Placebo; (6) Psychosocial support programs (service dogs, caregiver education programs, and group sessions); (7) Psychological education (cognitive behavior therapy); (8) Multi-disciplinary Programs (long-term group courses, multi-modal therapy, and psychological education combined with physical exercise).

Comparator(s)/Control
Since the network meta-analysis is based upon the Bayes' theorem, it is feasible to make the indirect comparisons of the interventions mentioned above. The comparator(s)/control criteria were the same as the intervention(s) criteria.

Outcomes
The patient-reported outcome measures (PROMs) had been used as the primary outcome measure of this review. There are a lot of scales and questionnaires in PROMs. In this review, the self-reported scales and questionnaires had been re-classified into perspectives as follows: (1) Abilities of daily life activity; (2) Psychological health; (3) Social functioning; (4) Overall life quality. All the outcomes should be presented as scores, and the scales or questionnaire must have a total score. If not, the original scores could not be transferred into a uniform scale.

Study Design(s)
To make sure that the evidence quality of this systematic review would be higher as possible, only studies of randomized controlled trials (RCTs), whether single-armed or multi-armed, would be included in this review.

Exclusion Criteria
Studies would be excluded if: (1) not all the participants in the study meet the SSA's description of disability such as children, teenagers, and people with preclinical disabilities; (2) the study evaluated other treatments that could not be classified into one of the classes mentions in the PICOS, such as surgery and injections; (3) the study was a published abstract or lack of data; (4) the outcomes of the study could not be classified into one of the classes in the PICOS; (5) the scores of the outcome could be transferred into a percentage of a total score, for example, the outcome whose scores presented in exponential form would be excluded in this review; (6) the study was not about randomized controlled trials, such as cross-sectional studies, case reports, cohort studies and cross-over trials in a single group.

Search Strategy
A comprehensive, reproducible search strategy was performed on the following databases from 1 January 1990 to 31 December 2020: PubMed, MedLine, Ovid, and CINAHL. Reference lists of included studies were also searched. Grey literature was searched to identify potential studies. If data were insufficient, authors would be contacted and requested for missing data. The search terms used in each database was as follows: (1) in PubMed, the search term was "((people with disabilities) OR (disabled) OR (people with disability) OR (handicapped) OR (disable people) [Titile]) AND ((randomized) OR (randomised) [Title/Abstract])"; (2) in MedLine, CINAHL, and Ovid, the search term was "(TI people with disabilities OR disabled OR people with a disability OR handicapped OR disable people) AND ( AB randomized OR randomised) NOT (TI design or protocol or review)".
The search term "disability*" was not used in the search strategy to reduce the heterogeneity within studies and increase the consistency of the mathematical model. The reason was that the results of the search term "disability*" would involve a lot of disorders, dysfunctions, and impairments caused by diseases that were completely curable or self-limited. These disabilities might not match the description given by the SSA, which was that "be expected to result in death or which has lasted or can be expected to last for a continuous period of not less than 12 months".
Title, abstract, and full-text screening were made by two independent authors (Yining Xu and Xin Li). Any disagreement would be resolved by a third independent reviewer (Zhihong Sun).

Risk of Bias Assessment
The Cochrane Collaboration Risk of Bias Assessment Tool was used to evaluate the risk of bias. All the included studies were evaluated by two independent authors (Yining Xu and Xin Li). Any disagreement would be discussed and an independent arbitrator (Zhihong Sun) was invited when an agreement could not be met.

Data Extraction and Synthesis
All potential studies were imported into EndNote X9 (Thomson Reuters, Carlsbad, CA, USA) and duplicates would be removed. Data were extracted by two independent authors (Yining Xu and Xin Li). Any discrepancies would be solved by an independent arbitrator (Zhihong Sun).
Findings were summarized and population characteristics such as age, gender, type of disability, information of intervention protocols, and the reclassification of the intervention were collected and put into the extraction sheet of summary of included studies.
Details of the patient-reported outcome measures (PROMs), which included the full name, total score, length of follow-up, and reclassification of each scale, would be shown in another extraction sheet.
The original data of each study, which involved the sample size, the average scores within each group before and after the intervention, and the standard deviations of the scores, would be recorded in an independent extraction sheet for the data pre-processing.

Data Pre-Processings
Data pre-processing and analysis were made by two independent investigators (Yining Xu and Xin Li). The Microsoft Office Excel (Version 16.0, Microsoft Corporation, Redmond, WA, USA) was used to pre-process the original data, transferring all the scores (mean ± SD) of different scales and questionnaires in the included studies into the form of a percentage of the total score (mean% ± SD%). The Aggregate Data Drug Information System (ADDIS V1.16.8 Produced by Drugis.org) was used to analyze all the processed data, calculated effect size, pool data into network meta-analysis, and output all the results and figures. The effect would be presented by the form of Mean differences (MD). The results of the network meta-analysis from each perspective would be presented in the following parts.

Processing of a PROM with Subscales of Different Dimensions
If a PROM is a comprehensive measure of life quality with different subscales of various dimensions, each subscale would be considered separately and re-classified into a certain perspective. For example, the MOS 36-item short-form health survey (SF-36) would be divided into two components, a physical component score (SF-36 PCS) that belonged to a measure for Abilities of daily life activity, and a mental component score (SF-36 MCS) that belonged to a measure for psychological health.

Processing of Different PROMs of the Same Dimension in a Study
If a study reported more than one PROM from the same perspective, the average score and its standard variation would be calculated. For example, the study by Szantan [47] reported scores of Activities of Daily Life Difficulty (ADL) and Instrumental Activities of Daily Life Difficulty, since both of two scales were PROMs for abilities of daily life activity, the average score of the two scales and its standard variation would be calculated. By this method, each study would have only one pair of data from each dimension before being pooled together into the network meta-analysis. Meanwhile, all the scores would be converted into positive (higher is better) by the method of converting the score of a negative scale (lower is better) into its opposite number.

Network Geometry
The network geometries displayed the overall number and type of treatments in comparison, informed indirectly by the Bayesian simulation modeling, and provided key information about the strength of evidence informing each direct link between two different treatments [48]. In each network geometry, every node represented one of the competing interventions, while the lines corresponded to the available direct comparisons between each pair of interventions, and the amount of available information could be presented by "weight" the edge using numbers of arms on them.

Consistency and Inconsistency Analysis
If there are closed loops in the evidence structure, the inconsistency of the evidence should be assessed because in network meta-analysis the evidence structure is more complex. Inconsistency assessment could occur when a treatment C has a different effect when it is compared with A or B, for example, studies comparing A and C are systematically different from studies comparing B and C. Therefore, inconsistency may even occur with normal meta-analysis, but can only be detected using a network meta-analysis [49].
If there is no relevant inconsistency in the evidence, or there is no closed loop in the evidence structure, a consistency model could be used to conclude the relative effect of the included treatments. Network meta-analysis gives a consistent, integrated picture of the relative effects. However, given such a consistent set of relative effect estimates, it may still be difficult to conclude a potentially large set of treatments. Fortunately, the Bayesian approach makes it possible to process complex data, to estimate the probability that given by the priors and the data. The results would be shown in the rank probability plot. The sum of all rank probabilities is 1, both within a rank over treatments and within a treatment over ranks [49].
The valid results from network meta-analysis depended on the evidence network being internally consistent: direct and various sources of indirect evidence should be in agreement. Inconsistency referred to differences between direct and indirect effect estimates for the same comparison and significant inconsistency threatened the validity of the results of a network meta-analysis. Therefore, if presented, further exploration of inconsistency would be needed to identify possible sources of disagreement [50]. The random-effects standard deviations would be calculated under both consistency and inconsistency models and compared with each other to identify if there was inconsistency within interventions. If random effects standard deviations calculated under both consistency and inconsistency models were fully identical, it meant that there was a good consistency with the interventions. If not, the p-values from the analysis of the node splitting would be checked to determine which modal would be used [51].

Network Meta-Analysis
A league table would be after the model of data analysis had been determined, reporting results that represented the mean difference with 95% confidence intervals in the column-defining treatment compared with the row-defining treatment [52].
If the included studies had a good consistency, the ranking of measures and probability would be made to facilitate simultaneous inference regarding all treatments. A table showing the ranking of treatments would be made, based on the probability of each treatment being the most effective or the least effective. The overall sum of the percentage in each row or column should be 1.00 (100%) [53]. Probabilities are estimated for a treatment to be ranked at a specific place (first, second, and so on) according to each outcome. However, a ranking of treatments based solely on the probability for each treatment of being the best should be avoided. This is because the probability of being the best does not account for the uncertainty in the relative treatment effects and can spuriously give higher ranks to treatments for which little evidence is available. The probability of being the best has the disadvantage that it does not reflect the spread of rankings for the treatments and to consider just the crude figures may be misleading [54,55].

Pair-Wised Meta-Analysis
If two interventions were appearing separately, an additional pair-wise meta-analysis should be made. The result would be shown in forest-plot and the heterogeneity within studies would be assessed by the statistic I 2 [56].

The Split Note Calculation
While the results are easier to interpret, it requires a separate model to be run for each node to be split [49]. The node-splitting analysis is an alternative method to assess inconsistency in network meta-analysis. It assesses whether direct and indirect evidence on a specific node (the split node) is in agreement.
Node splitting has been proposed by Dias et al. [49] and essentially involves distinguishing between the direct and indirect evidence. It aims to identify consistency discrepancies associated with specific nodes. It is performed within a Bayesian framework and is computationally more intensive than other approaches. Whether the identified discrepancy is statistically significant could be determined by examining the calculating a respective Bayesian p-value [49].

Search Strategy and Information Extraction
The search yielded 954 titles and abstracts for screening. 94 full texts were screened and 14 were excluded. Sixteen studies were included in the final analysis [47,[57][58][59][60][61][62][63][64][65][66][67][68][69][70][71]. The identification process is shown by a flow diagram [72] (Figure 1). The information of all included studies is shown in Table 1, and all the information about the PROMs was provided in Table 2. All the original data is shown in the Supplementary file.

Risk of Bias
The result of the risk of bias assessment is shown in Figure 2. After discussion, a consensus was obtained for all items. Overall results were shown in Figure 2a. It could be seen that four studies had a high risk of bias, four studies had a moderate risk of bias, and eight studies had a low risk of bias. The overall bias was presented in Figure 2b: (1) the risk of performance bias (blinding of participants and personnel) was high (high in 12 studies ); (2) the risk of detection bias (blinding of outcome assessors) was low (high in six studies); (3) the risk of attrition bias (incomplete outcome data) was low (high in two studies); (4) the risk of selection bias (random sequence generation and allocation concealment) was low (high in three studies); (5) the risk of reporting bias (selective reporting of outcomes) was low (low in all studies).  The scale of independent behavior-revised (SIB-R) Social interaction  eight studies had a low risk of bias. The overall bias was presented in Figure 2b: (1) the risk of performance bias (blinding of participants and personnel) was high (high in 12 studies ); (2) the risk of detection bias (blinding of outcome assessors) was low (high in six studies); (3) the risk of attrition bias (incomplete outcome data) was low (high in two studies); (4) the risk of selection bias (random sequence generation and allocation concealment) was low (high in three studies); (5) the risk of reporting bias (selective reporting of outcomes) was low (low in all studies).

Overall Life Quality
The network geometry of the interventions for the overall life quality of the disabled was presented in  In the mixed treatments comparison of AE, MP, PS, and UC, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.02 (0.00, 0.07), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.03 (0.00, 0.07), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were 0.04 (0.00, 0.07). Moreover, the inconsistency factors with the 95% confidence intervals in the cycle of MP, PS, and UC were −0.00 (−0.10, 0.08). The mean value of the inconsistency factors was closed to 0. Therefore, there might be consistency discrepancies associated with specific nodes, and that it was necessary to make a node splitting analysis. In the mixed treatments comparison of AE, MP, PS, and UC, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.02 (0.00, 0.07), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.03 (0.00, 0.07), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were 0.04 (0.00, 0.07). Moreover, the inconsistency factors with the 95% confidence intervals in the cycle of MP, PS, and UC were −0.00 (−0.10, 0.08). The mean value of the inconsistency factors was closed to 0. Therefore, there might be consistency discrepancies associated with specific nodes, and that it was necessary to make a node splitting analysis.
In the adjusted indirect comparison of DT, PT, and Placebo, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.55 (0.24, 0.80), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.55 (0.24, 0.80), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were also 0.42 (0.02, 0.79). Since the random effects standard deviations of the consistency modal and the inconsistency modal were almost the same. It means that the analysis under consistency modal had a good validity. Table 3 shows the league tables of the network geometries in Figure 3a,b. Bold characters indicate that the data was statistically significant (0 was not included in the 95% confidence intervals).
The ranking of measures and probabilities is provided in Table 4 and shown as a bar graph ( Figure 4). What should be paid attention to was the fact that since the smaller hallux valgus angle indicated a better condition, in the figure of rank probability, Rank 1 was the best one, and Rank N was the worst one. According to the results, Active Exercise and Passive Therapy might have the highest probability of being the best intervention for the overall life quality of the disabled.
Since there was no pair of two interventions appearing separately, it was unnecessary to perform a pair-wise meta-analysis. The results of the node splitting analysis would be provided in Table 5, which showed the estimated quantiles for the direct evidence, the indirect evidence, the combined evidence, as well as the p-value. A large p-value indicates no significant inconsistency was found. According to Table 5, all the p-values were greater than 0.05, meaning that the consistency model should be used. In the adjusted indirect comparison of DT, PT, and Placebo, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.55 (0.24, 0.80), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.55 (0.24, 0.80), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were also 0.42 (0.02, 0.79). Since the random effects standard deviations of the consistency modal and the inconsistency modal were almost the same. It means that the analysis under consistency modal had a good validity. Table 3 shows the league tables of the network geometries in Figure 3a,b. Bold characters indicate that the data was statistically significant (0 was not included in the 95% confidence intervals).  The ranking of measures and probabilities is provided in Table 4 and shown as a bar graph (Figure 4). What should be paid attention to was the fact that since the smaller hallux valgus angle indicated a better condition, in the figure of rank probability, Rank 1 was the best one, and Rank N was the worst one. According to the results, Active Exercise and Passive Therapy might have the highest probability of being the best intervention for the overall life quality of the disabled.

Abilities of Daily Life Activity
The network geometry of the interventions for the abilities of daily life activity of the disabled is presented in Figure 5.  Since there was no pair of two interventions appearing separately, it was unnecessary to perform a pair-wise meta-analysis. The results of the node splitting analysis would be provided in Table 5, which showed the estimated quantiles for the direct evidence, the indirect evidence, the combined evidence, as well as the p-value. A large p-value indicates no significant inconsistency was found. According to Table 5, all the p-values were greater than 0.05, meaning that the consistency model should be used.

Abilities of Daily Life Activity
The network geometry of the interventions for the abilities of daily life activity of the disabled is presented in Figure 5. From this perspective, there was a mixed interventions comparison of AE, MP, PE, PS, and UC (Figure 5a), and there was an adjusted indirect comparison of interventions of DT, PT, and Placebo (Figure 5b). Since there was a closed loop in the evidence structure, the inconsistency of the evidence should be assessed.
In the mixed treatments comparison of AE, MP, PE, PS, and UC, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.01 (0.00, 0.03), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.01 (0.00, 0.03), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were 0.07 (0.00, 0.18). Moreover, the inconsistency factors with the 95% confidence intervals in the cycle of MP, PS, and UC were −0.01 (−0.20, 0.10), and the inconsistency factors with the 95% confidence intervals in the cycle of AE, MP, PS, and UC were −0.01 (−0.16, 0.10). The mean value of the inconsistency factors of the two cycles were both closed to 0. Therefore, there might be consistency discrepancies associated with specific nodes, and that it was necessary to make a node splitting analysis. In the mixed treatments comparison of AE, MP, PE, PS, and UC, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.01 (0.00, 0.03), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.01 (0.00, 0.03), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were 0.07 (0.00, 0.18). Moreover, the inconsistency factors with the 95% confidence intervals in the cycle of MP, PS, and UC were −0.01 (−0.20, 0.10), and the inconsistency factors with the 95% confidence intervals in the cycle of AE, MP, PS, and UC were −0.01 (−0.16, 0.10). The mean value of the inconsistency factors of the two cycles were both closed to 0. Therefore, there might be consistency discrepancies associated with specific nodes, and that it was necessary to make a node splitting analysis.
In the adjusted indirect comparison of DT, PT, and Placebo, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.01 (0.00, 0.03), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.02 (0.00, 0.03), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were also 0.01 (0.00, 0.03). Since the random effects standard deviations of the consistency modal and the inconsistency modal were almost the same. It means that the analysis under consistency modal had a good validity. Table 6 shows the league tables of the network geometries (Figure 5a,b). Bold characters indicated that the data was statistically significant (0 was not included in the 95% confidence intervals).
The ranking of measures and probabilities would be provided in Table 7 and shown in the bar graph ( Figure 6). What should be paid attention to was that since the smaller hallux valgus angle indicated a better condition, in the figure of rank probability, Rank 1 was the best one, and Rank N was the worst one. According to the results, Psychological Education and Passive Therapy might have the highest probability of being the best intervention for the abilities of daily life activity of the disabled.
Since there was no pair of two interventions appearing separately, it was unnecessary to make a pair-wise meta-analysis. The results of the node splitting analysis would be provided in Table 8, which showed the estimated quantiles for the direct evidence, the indirect evidence, the combined evidence, as well as the p-value. A large p-value indicates no significant inconsistency was found. According to Table 8, all the p-values were greater than 0.05, meaning that the consistency model should be used. In the adjusted indirect comparison of DT, PT, and Placebo, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.01 (0.00, 0.03), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.02 (0.00, 0.03), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were also 0.01 (0.00, 0.03). Since the random effects standard deviations of the consistency modal and the inconsistency modal were almost the same. It means that the analysis under consistency modal had a good validity. Table 6 shows the league tables of the network geometries (Figure 5a,b). Bold characters indicated that the data was statistically significant (0 was not included in the 95% confidence intervals).
The ranking of measures and probabilities would be provided in Table 7 and shown in the bar graph ( Figure 6). What should be paid attention to was that since the smaller hallux valgus angle indicated a better condition, in the figure of rank probability, Rank 1 was the best one, and Rank N was the worst one. According to the results, Psychological Education and Passive Therapy might have the highest probability of being the best intervention for the abilities of daily life activity of the disabled.

Psychological Health
The network geometry of the interventions for the psychological health of the disabled was presented in Figure 7. From this perspective, there was a mixed interventions Since there was no pair of two interventions appearing separately, it was unnecessary to make a pair-wise meta-analysis. The results of the node splitting analysis would be provided in Table 8, which showed the estimated quantiles for the direct evidence, the indirect evidence, the combined evidence, as well as the p-value. A large p-value indicates no significant inconsistency was found. According to Table 8, all the p-values were greater than 0.05, meaning that the consistency model should be used.

Psychological Health
The network geometry of the interventions for the psychological health of the disabled was presented in Figure 7. From this perspective, there was a mixed interventions comparison of AE, MP, PE, PS, and UC (Figure 7a), and there was an adjusted indirect comparison of interventions of DT, PT, and Placebo (Figure 7b). Since there was a closed loop in the evidence structure, the inconsistency of the evidence should be assessed.
In the mixed treatments comparison of AE, MP, PE, PS, and UC, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.31 (0.23, 0.47), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.32 (0.23, 0.48), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were 0.24 (0.01, 0.72). Moreover, the inconsistency factors with the 95% confidence intervals in the cycle of MP, PS, and UC were 0.00 (−0.45, 0.47), and the inconsistency factors with the 95% confidence intervals in the cycle of AE, MP, PS, and UC were −0.01 (−0.58, 0.53). The mean value of the inconsistency factors of the two cycles were both closed to 0. Therefore, there might be consistency discrepancies associated with specific nodes, and that it was necessary to make a node splitting analysis. In the mixed treatments comparison of AE, MP, PE, PS, and UC, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.31 (0.23, 0.47), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.32 (0.23, 0.48), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were 0.24 (0.01, 0.72). Moreover, the inconsistency factors with the 95% confidence intervals in the cycle of MP, PS, and UC were 0.00 (−0.45, 0.47), and the inconsistency factors with the 95% confidence intervals in the cycle of AE, MP, PS, and UC were −0.01 (−0.58, 0.53). The mean value of the inconsistency factors of the two cycles were both closed to 0. Therefore, there might be consistency discrepancies associated with specific nodes, and that it was necessary to make a node splitting analysis.
In the adjusted indirect comparison of DT, PT, and Placebo, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.01 (0.00, 0.03), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.01 (0.00, 0.03), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were also 0.02 (0.00, 0.03). Since the random effects standard deviations of the consistency modal and the inconsistency modal were almost the same. It means that the analysis under consistency modal had a good validity. Table 9 shows the league tables of the network geometries (Figure 7a,b). Bold characters indicated that the data was statistically significant (0 was not included in the 95% confidence intervals).
The ranking of measures and probabilities is provided in Table 10 and shown as a bar graph (Figure 8). What should be paid attention to was that, since the smaller hallux valgus angle indicated a better condition, in the figure of rank probability, Rank 1 was the In the adjusted indirect comparison of DT, PT, and Placebo, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.01 (0.00, 0.03), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.01 (0.00, 0.03), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were also 0.02 (0.00, 0.03). Since the random effects standard deviations of the consistency modal and the inconsistency modal were almost the same. It means that the analysis under consistency modal had a good validity. Table 9 shows the league tables of the network geometries (Figure 7a,b). Bold characters indicated that the data was statistically significant (0 was not included in the 95% confidence intervals).
The ranking of measures and probabilities is provided in Table 10 and shown as a bar graph (Figure 8). What should be paid attention to was that, since the smaller hallux valgus angle indicated a better condition, in the figure of rank probability, Rank 1 was the best one, and Rank N was the worst one. According to the results, the Psychosocial Support Program and Passive Therapy might have the highest probability of being the best intervention for the psychological health of the disabled.   indirect evidence, the combined evidence, as well as the p-value. A large p-value indicates no significant inconsistency was found. According to Table 11, all the p-values were greater than 0.05, meaning that the consistency model should be used.     Since there was no pair of two interventions appearing separately, it was unnecessary to make a pair-wise meta-analysis. The results of the node splitting analysis would be provided in Table 11, which showed the estimated quantiles for the direct evidence, the indirect evidence, the combined evidence, as well as the p-value. A large p-value indicates no significant inconsistency was found. According to Table 11, all the p-values were greater than 0.05, meaning that the consistency model should be used.

Social Functioning
The network geometry of the interventions for the social functioning of the disabled was presented in Figure 9. There was an adjusted indirect comparison of interventions ( Figure 9a) and a directed comparison of interventions (Figure 9b) from this perspective. There was no closed loop in the evidence structure, so a consistency model would be used to conclude the relative effect of the included treatments.

Social Functioning
The network geometry of the interventions for the social functioning of the disabled was presented in Figure 9. There was an adjusted indirect comparison of interventions ( Figure 9a) and a directed comparison of interventions (Figure 9b) from this perspective. There was no closed loop in the evidence structure, so a consistency model would be used to conclude the relative effect of the included treatments. Before using the consistency modal, the vilification of the modal would be done. In the adjusted indirect comparison of AE, PS, MP, PE, and UC, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.36 (0.23, 0.64), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.36 (0.24, 0.63), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were also 0.41 (0.02, 0.80). Since the random effects standard deviations of the consistency modal and the inconsistency modal were almost the same. It means that the analysis under consistency modal had a good validity. Table 12 shows the league tables of the network geometries (Figure 9a). Bold characters indicated that the data was statistically significant (0 was not included in the 95% confidence intervals).
The ranking of measures and probabilities is provided in Table 13 and shown as a bar graph (Figure 10). What should be paid attention to was that, since the smaller hallux valgus angle indicated a better condition, in the figure of rank probability, Rank 1 was the best one, and Rank N was the worst one. According to the results, the Psychosocial Support Program might have the highest probability of being the best intervention for the social functioning of the disabled.
The results of the direct comparison of interventions would be provided by forest plots as shown in Figure 11. Moreover, since the results could not be interpreted, it wasn't required a separate model to be run for each node to be split. Before using the consistency modal, the vilification of the modal would be done. In the adjusted indirect comparison of AE, PS, MP, PE, and UC, the random effects standard deviations of the consistency modal and its 95% confidence intervals were 0.36 (0.23, 0.64), the random effects standard deviations of the inconsistency modal and its 95% confidence intervals were 0.36 (0.24, 0.63), and the inconsistency standard deviation of the inconsistency modal and its 95% confidence intervals were also 0.41 (0.02, 0.80). Since the random effects standard deviations of the consistency modal and the inconsistency modal were almost the same. It means that the analysis under consistency modal had a good validity. Table 12 shows the league tables of the network geometries (Figure 9a). Bold characters indicated that the data was statistically significant (0 was not included in the 95% confidence intervals).  The bold means the result is statistically significant.
The ranking of measures and probabilities is provided in Table 13 and shown as a bar graph (Figure 10). What should be paid attention to was that, since the smaller hallux valgus angle indicated a better condition, in the figure of rank probability, Rank 1 was the best one, and Rank N was the worst one. According to the results, the Psychosocial Support Program might have the highest probability of being the best intervention for the social functioning of the disabled.

Discussion
PROMs have an extremely relevant role in practice. Managing to implement a systematic collection of PROMs would be one of the hardest challenges at a system level. The collection of PROMs may become part of clinicians' daily practice and may lead to a The results of the direct comparison of interventions would be provided by forest plots as shown in Figure 11. Moreover, since the results could not be interpreted, it wasn't required a separate model to be run for each node to be split.

Discussion
PROMs have an extremely relevant role in practice. Managing to implement a systematic collection of PROMs would be one of the hardest challenges at a system level. The collection of PROMs may become part of clinicians' daily practice and may lead to a

Discussion
PROMs have an extremely relevant role in practice. Managing to implement a systematic collection of PROMs would be one of the hardest challenges at a system level.
The collection of PROMs may become part of clinicians' daily practice and may lead to a change in the relationship and communication between clinicians and their patients. By this say, clinicians could accept to have their job reviewed and not be afraid to be evaluated by their patients [73]. Introducing a successful systematic collection of PROMs would be beneficial for the performance of clinicians, improve the patients' satisfaction, and provide more valuable information for the development of disabled-care programs. Further research should be helpful for the managers of the medical system and the social security to formulate an official guide for collecting PROMs of the disabled systematically.
The results show that, to overall life quality, active exercise and passive therapy might have the most potential to becoming the best choice of intervention., Disability would cause inability to engage in any substantial, gainful activity and then decrease the daily life activities of the disabled because of a medically determinable physical or mental impairment. With the decrease of daily activity, the self-efficacy and activity willingness of the disabled also decrease gradually [74]. Therefore, it would be most important for the disabled to preserve their remaining physical functions as much as possible so that they could keep their physical activities as more as possible. However, disabilities might increase the risk of injury or re-injury when people with disabilities doing active exercises [75]. The fear of injury and re-injury would make disabled people suffering from psychological setbacks and fear-avoidance beliefs during training. It indicated that future research needs to explore the best active exercise or passive therapy scheme for the disabled.
As to the abilities of daily life activity, there are no statistically significant differences in all the head-to-head comparisons. However, the strengths of the network meta-analysis, which based on the Bayesian method, lies in that the significance of difference and the pvalue would no longer be the main factor affecting the conclusion, and the intervention that is most likely to be the best choice could be selected by probability judgment. According to the result presented in the table of the ranking of measures and probabilities, psychological education and passive therapy are the most potential interventions. The difference between the effects of psychological education and usual care is almost statistically significant. The reason might be that the outcomes in this review are Patient-reported scales or questionnaires which might show some kind of subjectivity, and the psychological education could increase the self-efficacy and self-esteem of people with disabilities [76][77][78] so that the patients might report positively. In a conclusion, it is undeniable that psychological interventions could successfully increase the subjective feelings of the disabled and make them more active in daily life and healthier [79].
When considering psychological health, the most potential intervention is a psychosocial support program, whose effect is statistically significant when compared with that of usual care. The result indicated that, since all the disabled are part of the society and improving the overall life quality of people with disabilities is essential to complete their socialization, psychosocial factors should be taken into serious consideration when designing a care program for the disabled and all sectors of the society should be involved [80][81][82][83]. Moreover, when designing a disabled-care program, the focus should not only be on the subjective psychological health of the disabled but also their psychosocial health. Different severity of disabilities might represent different residual body function, people with different disabilities would have different abilities of daily life activity. However, it doesn't mean that people with poorer physical abilities necessarily have poorer psychosocial health. For example, completely paralyzed patients might have the same psychological health as patients with mild disabilities. In further researches, the correlation of psychological well-being, the willingness of activities, and abilities of daily life activity should be studied [84,85].
When it comes to social functioning, a psychosocial support program, whose effect is significantly different from that of usual care, is the most potential intervention. Seco's team concluded that the effect of passive therapy was better than placebo [57]. However, there was no trial compared the effects of psychosocial support program and passive therapy directly, meaning that the result that the multi-disciplinary program which included psychosocial support is less potential than psychosocial support program in this network meta-analysis is possible because of the lack of data. Besides, the cost-efficiency of disabled-care programs is always the focus of social concern [86][87][88]. However, few studies are comparing the comprehensive cost-efficiency of disabled-care programs. Future studies should focus on the optimal application of resources, especially in multi-disciplinary programs.
The strengths of this systematic review are that, first of all, different outcomes of PROMs and interventions were reclassified so that, as showed in the results, the heterogeneity within studies is reduced, the consistency of the calculation model is increased, and the inconsistency indicators of each model are very low. Secondly, the scores of scales and questionnaires with different total scores are normalized into the same scale, making it feasible to pool the original data together and compare the results. Finally, the use of network meta-analysis realizes to make adjusted indirect and mixed comparisons of different types of interventions.
The main limitation of this systematic review is that all the disabilities caused by fully curable or self-limited diseases were excluded in this review. However, some diseases, both acute and chronic, could cause irreversible disabilities that meet the criteria of the SSA. Meanwhile, not every potential study reported the detail of the participants' disabilities, making it infeasible to judge every disability in each potential study in the library with the criteria of the SSA. This limitation also illustrates the necessity to introduce a successful systematic collection of PROMs.

Conclusions
This systematic review reclassified the interventions for the disabled into active exercise, passive therapy, psychological education, psychosocial support program, multidisciplinary program, and usual care, using the method of network meta-analysis to compare the effects of these interventions from the perspective of abilities of daily life activity, psychological health, social functioning, and overall life quality. Consistency modal was used in the network meta-analysis and had been verified a good consistency. In conclusion, active exercise and passive therapy are most likely to be the best choices for overall life quality, psychological education and passive therapy are most likely to be the best choices for abilities of daily life activity, and psychosocial support programs are most likely to be the best choice for psychological health and social functioning. The results remind us that the disabled are also an important part of society, intervention programs for the disabled should not only focus on their physical health but also their psychological health and socialization.

Data Availability Statement:
The data that support the findings of this study are available from the corresponding author, upon reasonable request.