The process of the study screening is presented in Figure 1. In total, 3979 relevant articles were searched from PubMed (1776), Embase (1851), and Cochrane Library (352) databases based on the preliminary search strategy. After removing 1524 duplicates, 2455 articles were subsequently analyzed, and 2328 irrelevant articles were further excluded by title and abstract reviewing. Next, the reminded 125 articles were future filtered through full text reviewing, and 103 articles that didn’t accord with the inclusive criteria were eliminated, including 25 case series/report, 23 letter/ comment, 29 reviews/meta-analysis, 7 studies with duplicated populations and 31 articles without available data. As result, a total of 12 eligible studies were used to conduct following meta-analysis [13-15, 20-28].
Characteristics of eligible studies
The characteristics of 12 included studies are presented in Table 1. These studies were published between 1998 and 2017. The study regions were involved in Chile, Austria, Belgium, Netherlands, United States, Canada, Germany, Finland and Sweden. AET, control trailing (CT), RT and AET+RT were main types of exercise interventions. In total, 462 participants were recruited, of which 123 undergone AET, 214 undergone CT, 54 undergone RT and 71 undergone AET+RT. Majority of participants were obesity based on BMI values, and most of studies designed a 12-week of follow-up. The pre-diabetes subjects in included studies were mainly comprised of of IFG and IGT individuals.
The results of quality assessment showed that all the included studies demonstrated a high risk of performance bias, and most of studies presented an unclear risk of Detection bias, allocation concealment and other bias. However, most of studies had a low risk of random sequence generation, attrition and reporting bias. Overall, the quality of included studies was moderate (Figure 2).
Before pooling the data, heterogeneity test was performed among studies reporting BMI, FBG, HbA1c, Insulin, and HOMAIR indicators, respectively. The results showed a significant heterogeneity was found among studies involving comparisons of AET vs. CT (I2=56.0%) for BMI, AET vs. CT (I2= 83.4%) and AET+RT vs. CT (I2= 80.4%) for FBG, AET vs. CT (I2=96.0%) and RT vs. CT (I2=83.5%) for Insulin, and AET vs. CT (I2=91.9%) and RT vs. CT (I2=75.9%) for HOMAIR, thus the random effect model was used to pool data. Whereas, the other comparisons for each indicator were calculated using fix effect model. The results showed that the HbA1c value in prediabetes individuals undergone AET was significantly reduced than individuals undergone CT (SMD = -0.7179, 95%CI = -1.0421; -0.3937). In addition, the prediabetes in RT group had a lower HOMAIR value than in CT group (SMD =-1.4826, 95%CI = -2.5750; -0.3902) (Table 2). However, there was no statistical significance in other comparisons for each indicator (Table 2). Notably, the comparisons results obtained from only one included studies were unconsidered in our study.
The network construction diagram showed that four comparisons between groups (RT vs. CT, RT vs. AET, AET+RT vs. CT, and AET vs. CT) were only found in included studies (Figure 3). Based on the Q statistic, the random effects model was used for the network meta-analysis. The results of P-scores for network meta-analysis showed there were no statistical difference between four comparisons for BMI, FBG, HbA1c, Insulin and HOMAIR. However, the greater decreases in BMI values could be seen in AET (p score = 0.7525) and AET+RT (p score = 0.5750) groups, while prediabetes individuals in RT groups had better curative effect on FBG (p score = 0.5750) and HbA1c (p score = 0.8550) values. In addition, the HOMA-IR and insulin values were better changed in AET (both p score = 0.6411) than in other groups (Tables 3 and 4).
Sensitivity analysis and consistency test
In order to evaluate whether the results of network meta-analysis was stable, the fixed and random effects models were both used to pool the data. Notably, the results showed the rank of P-scores for each indicator under above two effect models was basically consistent, indicating that our results of network meta-analysis were stable (Table 4). In addition, the result of Node-splitting analysis demonstrated that and the results from direct and indirect comparisons were consistent with all P>0.05 (Table 5).