Mild cognitive impairment (MCI), the intermediate state between normal age-matched cognition and dementia, is characterized by self- or informant-reported cognitive complaints, objective memory impairment, preserved independence in functional abilities, and no dementia [1, 2]. The average prevalence is 16% [3], and it has been reported that individuals with MCI are more likely to develop dementia, with annual conversion rates ranging from 10 to 15%, which is higher than 1–2% in healthy controls [4]. Due to the high morbidity [5] and costs [6] of dementia, effective preventive strategies are becoming increasingly important. Hence, it is of great significance to find an effective way to delay or even reverse cognitive decline in those with MCI.

Pharmacological and nonpharmacological interventions are two promising options for MCI. Because of the uncertain benefits of pharmacological interventions, patients’ preferences and safety, nonpharmacological interventions are receiving increasing advocacy [7]. Among them, exercises have always been the focus. Different types of exercises have been tried to delay the cognitive decline associated with MCI, such as aerobic exercises, resistance exercises, mind–body exercises, and exergames [8,9,10,11]. Additionally, meta-analyses have indicated that aerobic exercises and resistance exercises are both effective [12, 13]. However, the efficacy of mind–body exercises and exergames has not been pooled and quantified, and there is no conclusive evidence for which form of exercise is the most effective. To our knowledge, there are no meta-analyses evaluating the relative efficacy of different types of exercises at present, and the relevant randomized controlled trials (RCTs) are very limited. A meta-analysis showed that compared with resistance exercises, aerobic exercises may work better on the improvement of cognition in MCI patients through the subgroup analysis [12]; while a RCT indicated that resistance exercises could better improve selective attention, associative memory, and functional brain plasticity [14].

Therefore, it is necessary to conduct a network meta-analysis to compare the efficacy of different types of exercises on global cognition in adults with MCI.


Inclusion/exclusion criteria

We included RCTs that evaluated the efficacy of exercises on global cognition in adults with MCI. Participants must have been adults aged 50 years or older, and had MCI as defined by the original authors. Interventions must have included at least two of the following: aerobic exercise, resistance exercise, mind–body exercise, exergame, or control (stretching, education, usual lifestyle, and social recreational activities). Global cognition must have been measured as an outcome and provided as the mean and the standard deviation (SD). We included only published full-text papers written in English or Chinese; study protocols, conference abstracts, theses, and letters were all excluded because peer review had not been conducted, or there was insufficient information for methodological quality assessment in these publications. Papers with data contained in other included studies were also excluded.

Search strategy, study selection, and data extraction

We systematically searched PubMed, EMBASE, the Cochrane Library, Web of Science, PsycINFO, and the China National Knowledge Infrastructure (CNKI) from their inception to December 14, 2018. We conducted searches using MeSH terms and free terms. The keywords were “mild cognitive impairment”, “aerobic exercise”, “resistance training”, “mind–body exercise”, “tai chi”, “yoga”, “qigong”, “dancing”, “exergames”, “video games”, “virtual reality”, and “cognition”. The search strategy for PubMed is presented in Supplementary Table 1. Additionally, relevant reviews and meta-analyses were reviewed to prevent omission.

Study selection (Wang and Wang) and data extraction (Wang and Yin) were both conducted by two separate reviewers. Disagreements were resolved through discussions. Titles/abstracts and full texts of the studies retrieved through database searches were successively screened for eligible studies. A standardized form containing the first author, publication year, country, sample size, patient characteristics, interventions, and outcome measures was used to extract data.

Methodological quality assessment

The Cochrane Collaboration Risk of Bias tool [15] was used to assess the methodological quality of the studies eventually included. Six domains were assessed, including selection bias, performance bias, detection bias, attrition bias, reporting bias and other bias, and they were rated as low, unclear, and high. Two reviewers (Wang and Yin) independently performed the assessment. Disagreements were resolved by consulting a third reviewer (Chen).

Statistical analysis

Initial, pairwise meta-analyses were performed using Review Manager 5.3. The mean change and SD from baseline to endpoint of the outcome measures were used to calculate the “effect size”, which was expressed as a standardized mean difference (SMD) due to the different scales. And it was rated as small (0.2), moderate (0.5), or large (0.8) [15, 16]. The I2 test was used to assess heterogeneity when two or more studies were pooled, and level of heterogeneity was rated as none (0%), low (25%), moderate (50%), or high (75%) [17]. We used a random-effects model because it is more conservative [18].

Network meta-analysis was performed using Stata 12.0 and GeMTC 0.14.3. For the network meta-analysis, the number of chains was 4; initial values scaling, 2.5; tuning iterations, 20,000; simulation iterations, 50,000; and thinning interval, 10. Inconsistency was evaluated by inconsistency factors and node-splitting analysis [19]. If the 95% credible intervals (CI) for an inconsistency factor contained the neutral value (0) and p ≥ 0.05 in the node-splitting analysis, it indicated that there was no significant inconsistency, and the consistency model would be used. Otherwise, the inconsistency model would be used.


Search process

The process of search and selection is shown in Fig. 1. Initially, 8673 records were retrieved from six databases. After removing 3238 duplicates, 5435 records remained. Subsequently, 5305 records were excluded during the title/abstract screening, and 130 records had the full-text screened; from these, 113 records were excluded, and 17 records remained. Additionally, one paper was obtained through relevant reviews, and 18 papers were finally included.

Fig. 1
figure 1

Flow diagram of search and selection

Characteristics of the included studies

Table 1 summarizes the characteristics of the 18 included studies. The studies were conducted in China (n = 7), Korea (n = 3), Australia (n = 2), America (n = 1), Iran (n = 1), Spain (n = 1), Japan (n = 1), Greece (n = 1), and Croatia (n = 1). Sample sizes ranged from 20 to 389. Participants were all diagnosed with MCI and were aged 50 or older. Overall, women accounted for a greater proportion than men. For the experimental groups, five studies [8, 20,21,22,23] conducted aerobic exercises; four studies [9, 24,25,26] conducted resistance exercises; seven studies [10, 27,28,29,30,31,32] conducted mind–body exercises; and one study [11] conducted exergame. Usual lifestyle (n = 9), health education (n = 3), or placebo (n = 5) were conducted in the control groups. One study [33] compared the efficacy of exergame and mind–body exercise. The intervention periods were 6–48 weeks, the frequencies were 1–6 sessions/week, and the duration of each session was 21–90 min. The instruments used for measuring global cognition included the Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA) and its Korean version (K-MoCA), Alzheimer’s Disease Assessment Scale-Cognitive subscale (ADAS-Cog), and Computerized Assessment of Mild Cognitive Impairment (CAMCI).

Table 1 Characteristics of the included studies

Risk of bias of included studies

The risk of bias of the included studies is presented in Supplementary Fig. 1. In terms of selection bias, nine studies [8, 10, 11, 20, 24, 25, 30,31,32] used random numbers generated by computers, one study [28] conducted the randomization by a technician using an algorithm, the other eight studies [9, 21,22,23, 26, 27, 29, 33] just mentioned randomization, and did not describe the specific method; six studies [10, 20, 25, 28,29,30] reported the information about allocation concealment, and the other twelve studies [8, 9, 11, 21,22,23,24, 26, 27, 31,32,33] did not. Regarding performance bias, as the outcome measures were objective and could not be influenced by the awareness of participants, all studies had a low risk of bias. Regarding detection bias, nine studies [10, 20, 21, 24, 25, 27,28,29,30] blinded the assessors, one study [11] conducted assessment by computers, one study [23] did not blind the assessors, and the other seven studies [8, 9, 22, 26, 31,32,33] did not report the relevant information. Regarding attrition bias, nobody dropped out in two studies [11, 23], seven studies [8, 10, 20, 21, 24, 25, 33] conducted intent-to-treat analyses, two studies [22, 31] did not report the information, and the other seven studies [9, 26,27,28,29,30, 32] were considered high-risk. In terms of reporting bias, just one study [28] did not report all outcome measures listed in the methods. Regarding other bias, five studies [9, 11, 22, 26, 33] were considered high-risk because of small sample sizes.

Analyses of outcomes

Pairwise meta-analyses

The direct comparisons of different types of exercises and the controls are presented in Fig. 2. For studies conducting aerobic exercises, the interventions in the control groups included stretching, health education, usual lifestyle, and social recreational activities. For studies conducting resistance exercises, the interventions in the control groups included stretching, usual lifestyle, balance and toning exercises. For studies conducting mind–body exercises, the interventions in the control groups included usual lifestyle, health education, stretching and toning exercises. For the study conducting exergame, the intervention in the control group was health education. Aerobic exercises [n = 5, I2 = 5%, SMD = 0.80, 95% CI (0.55, 1.05), p < 0.00001], resistance exercises [n = 4, I2 = 50%, SMD = 0.80, 95% CI (0.29, 1.32), p = 0.002], and mind–body exercises [n = 7, I2 = 43%, SMD = 0.41, 95% CI (0.20, 0.62), p = 0.0001] all had positive effects on global cognition in individuals with MCI. Exergame [n = 1, SMD = 0.57, 95% CI (− 0.32, 1.47), p = 0.21] did not show significant effects on global cognition. There was no significant difference between exergame and mind–body exercise, while exergame seemed to be better [n = 1, SMD = 0.44, 95% CI (− 0.31, 1.19), p = 0.25].

Fig. 2
figure 2

The direct comparisons of different types of exercises and the control

Network meta-analysis

As shown in Fig. 3, five studies compared aerobic exercises with control, four studies compared resistance exercises with control, seven studies compared mind–body exercises with control, one study compared exergame with control, and one study compared exergame with mind–body exercise. One cycle containing mind–body exercise, exergame, and control was formed. Overall, 171 persons participated in aerobic exercises, 77 persons participated in resistance exercises, 340 persons participated in mind–body exercises, 24 persons participated in exergames, and 622 persons participated in the control groups.

Fig. 3
figure 3

Network map for the comparison of different interventions

The consistency model was used to conduct the network meta-analysis because the median (95% CI) of the inconsistency factor was 0.45 (− 2.42, 3.98), and the three p values in the node-splitting analyses were 0.64, 0.58, and 0.53.

The relative effects of the different interventions are presented in Table 2. The analysis revealed that aerobic exercises (SMD = 1.89, 95% CI 1.21, 2.69), resistance exercises (SMD = 3.01, 95% CI 1.85, 4.17), mind–body exercises (SMD = 1.04, 95% CI 0.45, 1.57), and exergames (SMD = 1.97, 95% CI 0.39, 3.63) all had positive effects on global cognition in participants with MCI relative to the controls; resistance exercises worked better than mind–body exercises (SMD = 2.01, 95% CI 0.70, 3.24); and there were no significant differences among the other comparisons.

Table 2 Relative effects of different interventions

The rank probability of the efficacy of different interventions is shown in Table 3. Resistance exercises were most likely to be ranked 1 (83%); exergames were most likely ranked 2 (40%); aerobic exercises were most likely ranked 3 (49%); mind–body exercises were most likely ranked 4 (86%); and the controls were most likely ranked 5 (99%).

Table 3 Rank probability of the efficacy of different interventions


Main findings and interpretation

The meta-analysis compared the efficacy of aerobic exercises, resistance exercises, mind–body exercises, and exergames on global cognition in adults with MCI. We found that the four types of exercises all had significant positive effects, and resistance exercises showed better efficacy than mind–body exercises; resistance exercises may be the most efficient, followed by exergames, aerobic exercises, and mind–body exercises; the efficacy of aerobic exercises and exergames seemed to be roughly equal, but exergames may work better.

Our finding that aerobic and resistance exercises were both effective in improving global cognition in adults with MCI was consistent with a previous review [12]. The finding that there was no statistically significant difference between the efficacy of aerobic and resistance exercises was inconsistent with the review of Song et al. [12], which revealed that aerobic exercises were more effective. The possible reason for this difference in results was that we included more studies [8, 22,23,24, 26], and we conducted a network meta-analysis, while the previous review conducted a subgroup analysis. The finding that resistance exercises were likely to outperform aerobic exercises was partially consistent with the study of Nagamatsu et al. [14], which showed that resistance exercise was more effective. One possible explanation for the nonsignificant difference in our finding was that the intensities of resistance exercises conducted in the included studies were different, while Nagamatsu et al. conducted high-intensity resistance exercise. It has been reported that resistance exercises with high intensity and frequent practice had better efficacy [9, 26, 34].

Mavros et al. [35] indicated that strength gains were the intermediary factor between resistance exercises and cognitive benefits. High-intensity resistance exercises improved strength more effectively, and strength gains mediated cognitive benefits. It has also been reported that higher muscle/lean mass may be related to better cognition and brain size [36, 37]. However, the specific mechanism is unclear, and further exploration is needed. Possible mechanisms for the effectiveness of resistance exercises include the positive restructuring of the brain [14], the regulation of insulin-like growth factor 1 (IGF-1), brain-derived neurotrophic factor (BDNF), insulin sensitivity, cortisol levels, serum homocysteine, and inflammation [34, 35].

We found that mind–body exercises could significantly improve global cognition in MCI patients, and there have also been reviews [38, 39] showing positive efficacy in healthy populations or populations with dementia. We also found that resistance exercises outperformed mind–body exercises. This may be because mind–body exercises have less impact on strength. While they have additional positive effects on mood and memory [39], and there have not been studies directly comparing the two types of exercises. Hence, it is necessary to conduct direct comparative studies and further explore the mechanism.

We included only one study [11] examining the efficacy of exergame relative to the control group, and no significant effect, but a positive trend, was shown. This may be due to the small sample size (ten persons/group). This hypothesis was supported by the fact that the effect became significant in the network meta-analysis where the study was pooled with a study of an indirect comparison between exergame and control (exergame vs. mind–body exercise) [33]. A review [40] revealed that exergames had a positive effect on global cognition in healthy populations or populations with neurocognitive impairment, and it further indicated that exergames outperformed aerobic exercises alone, which was partially consistent with our finding. We found that their efficacy was roughly equal, but exergames may work better. Therefore, more relevant RCTs should be conducted in populations with MCI to confirm the relative efficacy.


First, the number of included studies was small, especially studies on exergames. Second, publication bias may exist because only studies published in Chinese or English were included.


Exercises are recommended for MCI to delay or even reverse cognitive decline, and high-intensity and frequent resistance exercises seem to be the optimal choice. Additional high-quality RCTs that examine the efficacy of exergames on global cognition in adults with MCI are needed. Multi-arm designs that compare the efficacy of different types of exercises are suggested. The mechanisms of different types of exercises need to be explored further.


The results of this meta-analysis suggest that four types of exercises all have significant positive effects on global cognition in adults with MCI. High-intensity and frequent resistance exercises may be the most effective, followed by exergames, aerobic exercises, and mind–body exercises. More high-quality RCTs are needed to examine the efficacy of exergames. Multi-arm RCTs are needed to evaluate the relative efficacy of different exercise types.