Obesity is a direct and indirect cause of serious chronic diseases such as hypertension, diabetes, heart disease, and cancer. Many studies during the past few decades have sought solutions to the problem of obesity. Despite those efforts, however, the prevalence of obesity has increased in many countries. In the attempt to fundamentally improve obesity treatment, many new detailed approaches have emerged.

Recent studies have described several interventions for obesity, including diet, physical activity or exercise, behavioral therapy, and medication. Among the various behavioral strategies, exercise interventions can provide effective weight maintenance, weight loss, weight maintenance after loss, and reduction of obesity. However, exercise alone has a limited effect on the body weights of individuals with obesity.¹ Furthermore, the guidelines for how much physical activity and exercise are needed to improve health are widely known to the public as well as researchers, but they are too general to specifically address obesity, and the effects are variable and often inconsistent. Thus, it is necessary to understand the magnitude of the effects that physical activity or exercise interventions have on obesity.

Most reviews that have examined the relationship between exercise and obesity have considered only a partial range of factors. For example, the study of Witham and Avenell² included only participants with a mean age ≥60 years. The meta-analysis of Türk et al.³ considered only randomized controlled trials (RCTs) that used high intensity exercise with obese patients aged 18 to 60 years. Furthermore, they did not consider the influence of exercise mode (intensity, time, frequency, duration) in their mediation analysis.

Because a systematic review and meta-analysis can statistically integrate multiple independent study variables by their effect sizes for a particular issue and infer overall conclusions, it is a useful strategy in this situation. If we know the average effect size of exercise intensity, time, frequency, and duration in improving obesity, we will be able to concretely establish how different interventions influence the dependent variables, especially weight, body composition, and lipid profile. Therefore, we here include studies of adult subjects to assess the relationship between exercise and obesity using mediator variables. Our purpose in this study was to examine the effectiveness of exercise in terms of time, duration, intensity, and the exercise itself in adults with obesity.

Search strategy and eligibility criteria

We conducted a literature search of three databases: Embase (http://www.embase.com/), PubMed (http://www.ncbi.nlm.nih.gov/pubmed/), and EBSCO Academic Search Premier (http://search.ebscohost.com/), to obtain a comprehensive list of studies from 2007 to 2016 that examined the effects of exercise on obesity. The keywords we used were “exercise and obesity” and “exercise and obese.” After the initial online search, we selected only RCTs and removed duplicate studies using the EndNote program (Thomson Reuters Co., Toronto, Canada).

At the screening stage, we used the participants, interventions, comparisons, outcomes, timing of outcome measurement, settings, study design (PICOTS-SD) frame to determine the scope of the literature.⁴ Three reviewers (professors with expertise in the area of physical exercise and exercise physiology) screened the titles and abstracts of 4,106 studies based on the inclusion criteria and PICOTS-SD and selected 726 studies. To increase the accuracy of the screening, two other reviewers (professors with expertise in the area of physical exercise and exercise physiology) screened those 726 studies and selected 109 studies. The screening process is summarized in a flowchart (Supplementary Fig. 1): participants: adults with obesity (body mass index [BMI], ≥30 kg/m²; age, 18–65 years); intervention: exercise; comparison: routine care, diet therapy, etc.; outcome: body composition, lipid profile, etc.; time: pre- and post-test studies; study design: RCT.

Full-text articles were then assessed for eligibility. Articles were eliminated if inadequate intervention conditions (e.g., improper control conditions) or insufficient statistical data (e.g., partial or not present) were reported and if weight loss and maintenance for adults with obesity were not the primary focus of the study (e.g., analysis of functional capacity after gastric bypass surgery, breast cancer research). In the end, 64 studies were selected for this systematic review and meta-analysis, and all of them were published in English.

Data extraction and quality assessment

For the meta-analysis of exercise interventions in adults with obesity, we extracted body composition (weight, BMI, % body fat, waist circumference) and lipid profile (total cholesterol [TC], triglycerides [TGs], low-density lipoprotein [LDL], and high-density lipoprotein [HDL]) data in the form of the mean, standard deviation, and sample sizes for both the treatment and control groups at pre- and post-test. In addition, to analyze the moderator variable effects, we extracted diet control (with or without diet), exercise time (from <60 to 450 minutes per week), duration (from ≤4 to 48 weeks), and intensity (low, moderate, vigorous, and high) data about the exercise interventions.

The quality of all studies was assessed using the Physiotherapy Evidence-based Database scale (PEDro) as rated by two authors in this study. The total PEDro score was derived by adding all scale items except for item 1 and specifying the eligibility criteria.⁴ The quality of the studies was assessed using four PEDro score categories: excellent (9–10), good (6–8), fair (4–5), and fail (≤3).⁵ In this study, we considered only studies that scored more than 6 points. Disagreements were resolved by discussion.

Reviewer agreement

Three reviewers (CK, SK, HJK) independently screened the titles and abstracts for eligibility. Agreement among the reviewers was assessed using the kappa statistic.⁴ The three kappa statistics for the three reviewers (K_{A_B}, K_{B_C}, and K_{C_A}) were 0.765 (95% confidence interval [CI], 0.568–0.883), 0.788 (95% CI, 0.644–0.932), and 0.702 (95% CI, 0.531–0.873), respectively. After the secondary screening by two other reviewers (KBK, YAS), another kappa statistic (K_{D_E}) was obtained and determined to be 0.766 (95% CI, 0.557–0.976). According to Landis & Koch’s guidelines,⁶ all the kappa values indicated substantial agreement, and they were at all acceptable within Fleiss’s guidelines.⁷

Statistical analysis

The meta-analysis and meta-analysis of variance (meta-ANOVA) were performed using R version 3.5.2 (http://www.r-project.org). The random effects model for computation of the mean effect size was assumed because we considered heterogeneous effect sizes among the studies, including this systematic review and meta-analysis. The moderator variable effects were analyzed by meta-ANOVA. Consecutive variables of the moderator were converted into categorical variables for the meta-ANOVA.

Publication bias and heterogeneity

Publication bias, often called the file-drawer effect because unpublished results are imagined to be tucked away in researchers’ file cabinets, is a potentially severe impediment to combining the statistical results of studies collected from the literature. To consider publication bias, we used Begg’s funnel plots and Egger’s regression test with a significance level of 0.10.

Heterogeneity was tested using I² values, and the extent of heterogeneity was estimated as follows: low (25%), moderate (50%), and high (75%) I² values.⁸ For the meta-analysis, the standardized mean difference (SMD) was considered as the mean effect size, and the 95% CI and I² value were obtained using forest plots. The effect sizes were interpreted using Cohen’s criteria: small (0.20), medium (0.50), and large (0.80).⁸

We included 64 articles in this systematic review and meta-analysis study. Twenty-nine articles had only one comparison between the experimental and control groups, and 35 articles had several (2–5) comparisons. Therefore, the total number of cases for comparison was 109. Table 1 presents the characteristics of the studies included in this meta-analysis.^9–72

Characteristics of the included studies

The characteristics of the 64 included studies are presented in Table 1. Overall, 5,025 subjects were included in these studies, and the groups ranged in size from 5 to 118 subjects. All studies were published between 2007 and 2017, though the search period represented 10 years (2007–2016).

There were 29 studies with diet control and 35 studies without it. The exercise time per week for the interventions ranged from 20 to 450 minutes, with an average time of 153.3 minutes. The exercise duration varied between 2 and 48 weeks, with an average time of 16.5 weeks. The exercise intensity was classified into four categories (low, moderate, vigorous, and high) using the American College of Sports Medicine (ACSM) Guidelines for Exercise Testing and Prescription.⁷³ The ACSM’s Foundations of Strength Training and Conditioning classifies the percent 1 repetition maximum (%1RM) as supramaximal (>100%), very heavy (95%–100%), heavy (90%–95%), moderately heavy (80%–90%), moderate (70%–80%), light (60%–70%) and very light (<60%).⁷⁴ In this study, we used low (<70%), moderate (70%–<80%), vigorous (80%–<90%), and high (≥90%) (Supplementary Table 1). The PEDro scores of all the included studies varied between 6 and 9 points, with an average of 6.52 points (standard deviation, 0.69).

Effects of exercise on body composition and lipid profiles

We used the SMD as the effect size of each study and present the 95% CI of the SMD and I² value as a measure of heterogeneity in Table 2. Heterogeneity in the effect of exercise on weight in 89 studies was high (I²=75.9%), with an average SMD of 0.358 in the range from 0.213 to 0.503. Thus, the average effect size was between small and medium. Heterogeneity in the effect of exercise on BMI in 78 studies was high (I²=84.8%), with an average SMD of 0.533 in the range from 0.349 to 0.716. Thus, the average effect size was medium. Heterogeneity in the effect of exercise on % body fat in 47 studies was moderate (I²=44.3%), with an average SMD of 0.379 in the range from 0.246 to 0.512. Thus, the average effect size was between small and medium. Heterogeneity in the effect of exercise on waist circumference in 75 studies was high (I²=84.1%), with an average SMD of 0.666 in the range from 0.478 to 0.854. Thus, the average effect size was between medium and large.

Heterogeneity in the effect of exercise on TC in 35 studies was high (I²=94.5%), with an average SMD of 0.721 in the range from 0.228 to 1.214, for an almost large average effect size. Heterogeneity in the effect of exercise on TG in 56 studies was high (I²=92.2%), with an average SMD of 0.603 in the range from 0.257 to 0.949, for a medium to large average effect size. Heterogeneity in the effect of exercise on LDL in 49 studies was above moderate (I²=65.3%), with an average SMD of 0.406 in the range from 0.238 to 0.573, for a small to medium average effect size. Heterogeneity in the effect of exercise on HDL in 57 studies was high (I²=88.7%), with an average SMD of −0.222 in the range from −0.495 to 0.052, for a small average effect size.

One result (% body fat) from the meta-analysis is presented as a forest plot (Fig. 1) because it was the only result that met both the statistical significance and heterogeneity criteria (I²<50%) (Table 2).

With respect to the Egger’s regression test for publication bias, the P-value was statistically significant for % body fat (t=3.11, df=45, P=0.003), waist circumference (t=1.95, df=73, P=0.055), and HDL (t=−4.58, df=55, P=0.019). The P-values for the remaining variables (weight, BMI, TC, TG, and LDL) were not significant. The funnel plot was somewhat asymmetric with respect to % body fat (Supplementary Fig. 2), though the regression model resulting from Egger’s regression test for % body fat was satisfactory.

Effects of moderators on exercise and obesity

The meta-ANOVA results showing the effects of the moderators on exercise, obesity, and body composition are shown in Table 3. The average SMD in the studies with diet control was higher than in those with no diet control (except for % body fat). Regardless of diet, the average SMDs for BMI and waist circumference were close to or above medium (0.452–0.761).

The average SMD for minutes of exercise per week did not yield a specific pattern, possibly because the small number of studies caused inconsistent results. The average SMD for an exercise duration of 12 weeks was close to or above medium for all 4 variables (weight, BMI, % body fat, waist circumference). This was especially the case for BMI (0.911) and waist circumference (0.910), which had large average SMDs.

The average SMD for exercise intensity was between 0.043 (% body fat) and 0.609 (BMI) for the low intensity case, 0.346 (weight) and 0.874 (waist circumference) for the moderate intensity case, and 0.293 (weight) and 0.613 (waist circumference) for the vigorous intensity case; no studies used high intensity exercise. The average SMD for the moderate intensity case was higher than that for the low and vigorous cases for BMI, % body fat, and waist circumference. Only with respect to weight was the average SMD higher at low intensity.

The meta-ANOVA results for the effects of the moderators on exercise, obesity, and lipid profile are shown in Table 4. The average SMD for diet treatment was above medium for TC (0.825) and TG (0.772). The average SMD for HDL was smaller than that for the other factors. The average SMD for exercise time per week was above medium and ranged from −0.506 (HDL) to 0.709 (TG) for 120 to 149 minutes of exercise per week. Similar to the results for body composition, the average SMD for exercise time per week did not appear to follow a specific pattern as exercise time increased. For this attribute also, the number of studies was small, which might have caused inconsistent results.

The average SMD for the effect of exercise duration on TC and TG increased rapidly from 12 weeks. The average SMD at 36 weeks was very large and ranged from −0.993 (HDL) to 2.585 (TG).

The average SMD for exercise intensity was between −1.675 (HDL) and 1.191 (TG) for the low intensity case, −0.023 (HDL) and 0.506 (LDL) for the moderate intensity case, and −0.285 (HDL) and 0.410 (TG) for the vigorous intensity case. The average SMD for the low intensity case was far greater than that of the other cases. The meta-ANOVA results for one variable (% body fat) are presented in the form of a forest plot (Fig. 2).

For this study, we performed a systematic review and meta-ANOVA to examine the effect of exercise on both the body composition and lipid profiles of adults with obesity. According to Table 2, the average effect size of exercise on BMI (SMD, 0.533), waist circumference (SMD, 0.666), TC (SMD, 0.721), and TG (SMD, 0.603) was above medium. The average effect size on weight (SMD, 0.358), % body fat (SMD, 0.379), LDL (SMD, 0.406), and HDL (SMD, −0.222) was below medium. The heterogeneity of most of the dependent variables was high (I²=75.9%–94.5%), except for LDL (I²=65.3%) and % body fat (I²=44.3%). The heterogeneity of % body fat was the only variable that was below moderate. These results indicate that the effects of exercise interventions on outward appearance, such as BMI and waist circumference, are larger than the effects on practical factors of obesity such as weight and % body fat. Moreover, the effect of exercise on TG (SMD, 0.603) was larger than that on LDL (SMD, 0.406) and HDL (SMD, −0.222).

The results of the publication bias analysis were statistically significant for % body fat, waist circumference, and HDL. In other words, the studies used in this meta-analysis adequately represent the population only for those three factors. With respect to the effects of the moderators on diet control, the SMDs for the cases with diet treatment were larger than those without diet control, as indicated by the weight, BMI, waist circumference, TC, and TG variables. Conversely, the SMDs with diet treatment were lower than those without diet control with respect to % body fat, LDL, and HDL. In a meta-analysis by Vissers et al.,⁷⁵ supervised exercise-only interventions had a greater effect than combined diet and exercise interventions. In a meta-analysis by Wu et al.,⁷⁶ a combined diet and exercise program provided greater long-term weight loss than a diet-only program. However, they concluded that both the diet-only and diet and exercise programs were associated with partial weight regain and that future studies should seek strategies to limit weight regain and achieve greater long-term weight loss.

In studies that considered the effects of moderators on exercise time per week, a medium or large SMD emerged for both body composition and lipid profile in programs that used more than 120 minutes of exercise per week. The ACSM’s guidelines for exercise testing and prescription recommend a minimum of 150 min·wk⁻¹ (30 min·day⁻¹) progressing to 300 min·wk⁻¹ (60 min·day⁻¹) of moderate intensity exercise for overweight and obese individuals.⁷³ Every study we considered here used a different combination of exercise intensity and duration.

An exercise duration of more than 8 weeks produced a medium or large SMD for body composition, as did a duration of 12 or more weeks for lipid profile. The ACSM’s guidelines for exercise testing and prescription recommend a minimum reduction in initial body weight of 5%–10% over the course of 3–6 months for overweight and obese individuals.⁷³ Furthermore, the “2018 Korean Society for the Study of Obesity Guideline for the Management of Obesity in Korea”⁷⁴ states, “Physical activity is necessary for weight loss and maintenance. At the start, more than moderate physical activity is recommended. Moderate levels of physical activity include exercising between 30 minutes and 60 minutes five times per week. When including resistance exercise, it is recommended to engage in physical activity twice per week.”

In this study, the average SMDs for weight loss by exercise duration were 0.237, 0.455, and 0.576 for 8–11 weeks, 12–15 weeks, and 16–19 weeks, respectively. In other words, the effect size was above medium after exercising for 16 weeks (4 months). Future studies should analyze both the rate of change in body weight and exercise duration together.

By exercise intensity, a medium or large SMD was observed in most cases with low and moderate intensity. Only waist circumference had an SMD above medium for vigorous intensity exercise. In a meta-analysis by Türk et al.,³ training at high intensity was found to be superior to moderate exercise in reducing % body fat in obese adults. That result is supported by some recent studies, but not by others. For example, a study by Keating et al.⁷⁷ showed that high intensity interval training (HIIT) could improve fitness levels with only 50%–60% of the time commitment required by continuous aerobic exercise training (CONT). However, the CONT group showed a reduction in total body fat, whereas the HIIT group did not. In addition, Kemmler et al.⁷⁸ found that HIIT provided more weight loss than moderate intensity continuous exercise(MICE) but produced no difference in body fat mass. Auriemma⁷⁹ explained that the available studies on HIIT in overweight and obese patients remain limited by their short duration, small number of participants, and variation in the intensity and duration of their “on” intervals. Researchers should also consider the risk of potential injuries, even though more vigorous exercises might provide additional benefits.⁷³

In conclusion, our novel finding is that the effect of exercise on obesity is larger in outward appearance (BMI, waist circumference) than in practical factors (weight, % body fat). Moreover, the effect of exercise on TG was larger than that on LDL and HDL. With respect to exercise mode, the effects of exercise duration and intensity are more consistent and larger among obese adults than the effects of exercise time (minutes per week).

The limitations of this study include the generalization of the findings due to the small sample sizes in some of the studies, heterogeneity, and publication bias in the analysis of some variables. Some cases had very large SMDs with very small sample sizes. In future reviews, the number of studies included for each moderator effect analysis must be adequate to satisfy the normality of the statistical inference.

The authors declare no conflict of interest.

We thank the exercise committee members for participating in this article.

Study concept and design: KBK; acquisition of data: KK, CK, SJK, and HJK; analysis and interpretation of data: KBK and YAS; drafting of the manuscript: KBK and SY; critical revision of the manuscript: KBK and YAS; statistical analysis: KBK; administrative, technical, or material support: KBK and YAS; and study supervision: KK and YAS.