We followed a standard systematic review protocol according to the PRISMA guidelines, which were recently adapted for network meta-analyses ( Figure 1 ) (23). The following sources were searched for studies carried out in or before August 2021: MEDLINE, EMBASE, the Cochrane Central Register of Controlled Trials (CENTRAL), the China National Knowledge Infrastructure (CNKI), the Research Information Sharing Service, and the Korea Studies Information Service. Reviewers independently searched the articles using the following search terms: (“acupuncture” OR “acupressure OR “acupoint” OR “catgut” OR “embedding” OR “Cognitive Behavior* Therapy” OR “Cognitive Behavior* Treatment” OR “Cognitive Behavior* intervention” OR “meal replacement” OR “meal supplement” OR ((diet* OR “caloric restriction”) AND “exercise”)) AND (“weight loss” OR “simple obesity”). Any disagreements were resolved through discussion.

Randomized controlled trials (RCTs) that directly compared the effect of LCD alone with LCD-combined acupuncture, CBT, and exercise and an MR-based diet on weight loss in adults aged 18–65 years with simple obesity (BMI > 25) were included in the study. Further, RCTs that compared the effect of LCD-combined interventions such as LCD-combined CBT and LCD-combined exercise with other LCD-combined interventions were also included. However, RCTs in which a placebo, no treatment, and LCD combined with other interventions were assigned to the control group were excluded from the study. To synthesize the direct and indirect effects of the interventions, we conducted a lumping process in which treatment nodes were categorized to form single comparators. In our study, five interventions were combined: acupuncture plus LCD, CBT plus LCD, exercise plus LCD, MR-based LCD, and LCD alone. Among these interventions, acupuncture included laser acupuncture, manual acupuncture, and acupoint catgut embedding. CBT included psychoeducation, cognitive restructuring, and cognitive skills training on eating, weight, and emotions. Exercise training included any specific exercise prescription that included the mode of activity, frequency, intensity, and duration of the exercise. LCD was defined as a diet in which an energy deficit > 500 kcal/day was achieved with a dietary intake of 1,200 to 1,500 kcal/day in female participants and 1,500 to 1,800 kcal/day in male participants. An MR-based diet included calorie-restricted MR, in which a packaged food portion replaced one or more meals per day. However, MR-based diets combined with drugs were excluded from the study.

A predefined data template was prepared after determination of the articles for inclusion. Two investigators (Kim SY and Shin YS) extracted and coded the data using a template. Included in the data template were the characteristics of the study (author name, publication year, journal, study design), participants demographic information (sample size, age, sex, body weight, and BMI), interventions (duration, frequency, types of therapy, and control group), and weight-related outcomes (mean changes and standard deviation in weight or BMI). Any disagreements between investigators were resolved through discussion. The methodological quality of the included studies was assessed using the Cochrane risk of bias tool (24), in which random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other biases were examined. In random sequence generation, using a random number table, computer random number generator, coin tossing, throwing dice, or shuffling cards was considered to satisfy low risk, whereas using odd or even date of birth, some rules based on date of admission, or clinical record number were considered as high risk. In allocation concealment, low risk was accepted when participants and investigators enrolling participants could not foresee assignment, because equivalent methods, including sequentially numbered or sealed envelopes or drug containers of identical appearance, were used to conceal allocation. A high risk of allocation concealment was determined when participants or investigators could foresee assignments because of an open random allocation schedule, date of birth, or case record number. In blinding of participants and personnel, a low risk was accepted when the outcome was not likely to be influenced by lack of blinding or not likely that the blinding was broken for participants and personnel, despite no blinding or incomplete blinding. A high risk of participant blinding was considered when the outcome was likely to be influenced by a lack of blinding. In outcome assessment blinding, low risk was accepted when the outcome was not likely to be influenced by lack of blinding or not likely that the blinding was broken for outcome assessors, whereas high risk was determined when the outcome was likely to be influenced by lack of blinding for outcome assessors. In incomplete outcome data, low risk was accepted when no missing outcome data or reported reasons for missing outcome data were unlikely to be related to the true outcome, such as survival data. A high risk was determined when dichotomous outcome data or reasons for missing outcome data likely to be related to true outcomes were reported. In selective reporting, low risk was accepted when the study protocol was available and prespecified (primary and secondary) outcomes were reported, whereas high risk was determined when prespecified outcomes were not reported or the study report failed to include a key outcome. Regarding other biases, low risk was accepted when the study appeared to be free of other sources of bias, whereas high risk was determined when the study had a potential source of bias related to the specific study design used. Among the seven risks of bias, unclear risk was determined when there was no reporting related to low or high risk for each assessment item.

In our study, the primary outcome was a change in BMI (kg/m²) or weight (kg). Since all study outcomes were documented on a continuous scale, the concept of effect sizes was based on the standardized weighted mean differences with a 95% CI between the post-test and pre-test outcomes in the treatment and control groups. As the BMI scale differs from the weight scale, all study outcomes were transformed to Hedges’ g (24). Moreover, the effect sizes of small-sample studies tended to be overestimated, and Hedges’ g was used to adjust for sample size-related bias (25). Hedges’ g > 0.8 was considered “large,” that of 0.2–0.8 was considered “medium,” and Hedges’ g < 0.2 was considered “small” (25). Heterogeneity in the BMI or weight between studies was assessed using the Q-test and I² statistics. I² <40% was considered “non-important heterogeneity” and 30%-60% was considered “moderate heterogeneity” (24). Comprehensive meta-analysis software (version 3; Biostat Inc) was utilized to conduct a pairwise meta-analysis of the interventions in a direct, head-to-head manner using a random-effects model. The geMTC package (version 3.6.1) in R was used to conduct a network meta-analysis for the indirect effects of the interventions using an arm-based, Bayesian random-effects model.

Ranking of the intervention classes was presented using a league table, ranking probabilities, and rankograms (23). Network geometry was used to assess loop inconsistency assumptions according to closed intervention loops (26). The plausibility of the transitivity assumption was blindly evaluated by two investigators based on the design characteristics and methodologies of the included studies. Two investigators also evaluated clinical (patient, treatment) and methodological heterogeneity (study design and outcome measures). The consistency between the direct and indirect comparisons of all closed loops was evaluated using the ode-splitting method (23, 24).

The initial search strategy identified 4,014 potential studies. Ultimately, 32 RCTs met the inclusion criteria and they were included in the final systematic review and meta-analysis ( Figure 1 ). A total of 3,364 patients were included, with a mean age of 41.7 years. The proportion of females among the subjects was 78.1%, and the mean BMI was 33.0 kg/m². The characteristics of the studies included in the network meta-analysis are summarized in Table 1 . Thirty-five intervention arms were included, consisting of six types of direct pairwise comparisons: acupuncture plus LCD vs. LCD alone (27–30), CBT plus LCD vs. LCD alone (31, 35), exercise plus LCD vs. LCD alone (38–41), MR-based LCD vs. LCD alone (42–57), CBT plus LCD vs. exercise plus LCD (32–34, 36, 37), and MR-based LCD vs. exercise plus LCD (58). Figures 2A-C show the network geometry of all study arms, the short-term (≤ 12 weeks) study arms, and the long-term (>12 weeks) study arms. The network geometry formed two closed loops consisting of pairwise comparisons between interventions and a common comparator (LCD alone), as well as pairwise comparisons between interventions without a common comparator, and the possibility of loop inconsistency was excluded.

The results of the direct pairwise meta-analysis are summarized in Table 2 . Forest plots of the five lumped interventions are shown in Figure 3 . All LCD-combined interventions were more effective than LCDs alone. The effect sizes on weight loss of LCD-combined acupuncture (Hedges’ g = 0.48, 95% CI = 0.24-0.71), LCD-combined CBT (Hedges’ g = 0.49, 95% CI = 0.14-0.85), LCD-combined exercise (Hedges’ g = 0.27, 95% CI = 0.03-0.51), and MR-based LCD (Hedges’ g = 0.31, 95% CI = 0.18-0.45) were higher than those of LCD alone. The direct effects on weight loss of LCD-combined CBT (Hedges’ g = 0.14, 95% CI = -0.04-0.45) and MR-based LCD (Hedges’ g = 0.22, 95% CI = -0.42-0.85) were not significantly different from those of LCD-combined exercise.

Indirect comparisons are built on an assumption of transitivity, which refers to the similarity of one set of RCTs to another sets of RCTs in all important factors except the intervention comparison being made (26). Two investigators (Kim SY and Shin YS) independently assessed the transitivity of all RCTs, and the transitivity assumption was accepted. Figure 4 shows the forest plots of each intervention compared with the LCD alone via a network meta-analysis. LCD combined acupuncture was found to be the most effective (Hedges’ g = 0.48, 95% CI = 0.25-0.71), followed by LCD-combined CBT (Hedges’ g = 0.42, 95% CI = 0.20-0.63), MR-based LCD (Hedges’ g = 0.32, 95% CI = 0.19-0.45), and LCD-combined exercise (Hedges’ g = 0.27, 95% CI =0.06-0.46). The league table showing the relative rankings of the interventions is listed in Supplementary Table 1 in the supplement. The rankograms for each intervention are shown in Supplementary Figure 1 in the supplement. The league table and rankograms indicated that LCD-combined acupuncture was the most effective for weight loss, followed by LCD-combined CBT, MR-based diet, and LCD-combined exercise. The numerical ranking probability is summarized in Supplementary Table 2 in the supplement. LCD-combined acupuncture showed the highest probability of being ranked first in the first rank (63.70%), and LCD-combined CBT showed the highest probability of being ranked first in the second rank (48.98%). MR-based LCD (44.05%), LCD-combined exercise (61.22%), and LCD only (99.19%) showed the highest probability of being ranked first in the third, fourth, and fifth ranks, respectively.

The heterogeneity test results are presented in Table 3 . There was no significant within-intervention heterogeneity. However, the Q (df = 34) value of the 35 study arms was 51.03 (p = 0.03), indicating significant inter-intervention heterogeneity. The I² value of all interventions was 33.37%, indicating a boundary between the low and moderate levels.

We examined whether the intervention period affected the effect size of each intervention, using intervention periods of 12 weeks or less (short-term) and over 12 weeks (long-term). Figures 2B, C show the network geometry in the short and long terms. Table 4 summarizes the network meta-analysis results based on the intervention period. LCD-combined acupuncture was moderately effective in the short term (Hedges’ g = 0.39, 95% CI = 0.12-0.67), while LCD-combined acupuncture was highly effective in the long-term when compared with LCD alone (Hedges’ g = 0.89, 95% CI = 0.37-1.40). LCD-combined CBT in the long-term was moderately effective when compared with LCD alone (Hedges’ g = 0.51, 95% CI = 0.26-0.76), while LCD-combined CBT was not significantly effective in the short term (Hedges’ g = 0.23, 95% CI = -0.30-0.75). LCD-combined exercise was moderately effective in the long term (Hedges’ g = 0.37, 95% CI = 0.12- 0.59), while it was not significantly effective in the short-term (Hedges’ g = 0.11, 95% CI = -0.27-0.48). The MR-based diet was moderately effective in the short term (Hedges’ g = 0.35, 95% CI = 0.18-0.53), while it was not significantly effective in the long term (Hedges’ g = 0.22, 95% CI = -0.01-0.47). Non-standardized mean differences in BMI between LCD-combined interventions and LCD alone are summarized in Supplementary Table 3 in the supplement.

The risk of bias in the 32 studies is summarized in Table 5 . Twelve studies used adequate methods of sequence generation, such as a computerized random number generator (34, 37, 39, 45, 52, 55, 57), the blocked random method (30, 44, 56), website randomization (36), or coin flip (49). The remaining 20 studies failed to report a precise randomization method. Allocation concealment was conducted in seven studies using opaque sealed envelopes (53), computer-based allocations (34, 37, 39, 45, 55), or website randomization (36). The remaining 25 studies failed to specifically mention the concealment methods used therein. Regarding the patient blinding process, the risk of bias was high in six studies. The risk of bias was also high in five studies with regard to outcome assessor blinding. As the included studies were not placebo-controlled, it was difficult for them to be designed as double-blind trials. There was no evidence of a high risk of bias in the other studies. Publication bias was examined using a funnel plot with trim and fill. The funnel plot had a symmetrical distribution, indicating no publication bias ( Figure 5 ) (59).

Of the 32 studies, seven reported that there were no adverse events (28, 43, 44, 47, 49, 51, 54), and five reported adverse events (45, 46, 50, 54, 55). The remaining 20 studies did not report any results regarding adverse events. The five studies that reported adverse events were related to an MR-based diet, and the adverse events reported were not serious. Allison et al. reported abdominal inflation, indigestion, abnormal taste, and lethargy (45). Fisberg et al. reported one case of a transitory elimination of flatus, which discontinued after 2 weeks of drug usage (46). Tsai et al. reported adverse effects of a green tea MR, including diarrhea, indigestion, and sleep loss (50). Metzner et al. reported that one patient was withdrawn from the study due to diarrhea (54). Fuller et al. reported two cases of gastrointestinal distress, constipation, and bloating (55).