Two researchers independently and systematically searched multiple databases including PubMed, Web of Science, Cochrane, Embase, China National Knowledge Infrastructure (CNKI), VIP, and Wanfang. The initial literature search cutoff date was June 2024, with an updated search conducted in November 2025 during the manuscript revision phase. This covered all records from each database’s inception through November 2025. To minimize potential omissions, reference lists of relevant studies were also screened. Search terms included subject headings such as “diabetes mellitus, “ “sarcopenia, “ and “traditional Chinese medicine.” The specific search strategies for each database are detailed in Supplementary Table S1.

Inclusion criteria: ①In vivo studies in animals; ②Type 2 diabetes animal model was successfully built; ③The study set up an experimental group (receiving TCM treatment) and a control group (T2DM model group receiving placebo treatment or no administration); ④The study included at least one of the following outcomes: body weight, gastrocnemius muscle weight, grip strength, muscle cross-sectional area, muscle tissue morphology, blood glucose.

Exclusion criteria: ①Repeat published; ②Other kinds of studies (clinical trials, case reports, review articles, conference papers); ③Outcome data are insufficient.

Any disputes are resolved through discussion with a third reviewer.

Two reviewers independently conducted data abstraction, including the following information: characteristics of studies (the name of first author, publication year, study country); characteristics of the animal (species, sex, weight, sample size, modeling method); characteristics of intervention (intervention methods for the experimental and control groups; intervention dosage, route of administration, treatment duration, composition/acupoints, and method of preparing formulas for the experimental group); outcome index (body weight, gastrocnemius muscle weight, grip strength, muscle cross-sectional area, muscle tissue morphology, blood glucose).

For studies presenting data solely in graphical format without numerical text, numerical data extraction was performed using the Digitizer tool embedded in OriginPro software (Version 2021, OriginLab Corporation, Northampton, MA, USA). To ensure data reliability and minimize observer bias, data extraction was conducted independently by two reviewers. The results were cross-verified, and any discrepancies were resolved through discussion or by consulting a third reviewer to reach a consensus.

Regarding the handling of missing data, a strict protocol was followed. First, for studies reporting standard errors of the mean (SEM) or confidence intervals (CI), missing standard deviations (SD) were calculated using standard formulas recommended by the Cochrane Handbook. Second, for studies with undisclosed or ambiguous data (e.g., missing sample sizes or unclear grouping) that could not be derived calculationally, we attempted to contact the corresponding authors via email to request the raw data. However, no positive responses containing valid raw data were received. Consequently, to ensure the integrity of the results, studies with irretrievable critical data were excluded from the quantitative synthesis.

Two reviewers independently assessed the quality of the included studies using the risk of bias tool from the Systematic Review Centre for Laboratory Animal Experimentation (SYRCLE) for animal studies (14). Based on the specific descriptions in each study, the risk of bias for each of the 10 items was assessed as “low”, “uncertain”, or “high” to reflect the quality of the study design and methodology.

Data statistical analysis was performed using Review Manager 5.3 software. The standardized mean difference (SMD) served as the measure of relative effect size. Effect size and its 95% confidence interval (CI) reflected the reliability of the results. Studies with low heterogeneity (P ≥ 0.1 or I² ≤ 50%) were analyzed using a fixed-effect model, while those with high heterogeneity employed a random-effects model.

A systematic search of seven databases yielded 1, 216 relevant records. After plagiarism checking and the initial screening of the title and abstract, 1, 153 studies were excluded. The remaining 63 studies proceeded to full-text review. Using established selection criteria, 43 studies that did not meet requirements were further excluded, resulting in the inclusion of 20 studies (15–34). The detailed study selection process is outlined in Figure 1.

This study included 20 animal experiments involving a total of 351 animals. Among them, 177 animals were assigned to the experimental group and 174 to the control group. Table 1 lists the main characteristics of the 20 included studies.

Based on the SYRCLE bias risk assessment tool, the 20 included studies were evaluated, with specific results shown in Figure 2A and 2B. Regarding random sequence generation, 7 studies employed random number tables, indicating a low risk of selection bias (19, 21, 22, 25, 26, 28, 32). The remaining 13 studies mentioned only “randomization” without specifying the method, resulting in an unclear risk of selection bias (15–18, 20, 23, 24, 27, 29–31, 33, 34). Regarding baseline characteristics, all studies demonstrated consistency in baseline traits such as strain, sex, and age among experimental animals. No significant differences were observed in core indicators (e.g., body weight, blood glucose) between groups after randomization, indicating a low risk of selection bias. For allocation concealment, none of the studies mentioned whether the allocation sequence was concealed from the personnel, resulting in an unclear risk of selection bias. Regarding random housing and blinding of personnel, none of the studies reported randomization of animal placement or blinding of animal caretakers and medication administrators, leaving the risk of performance bias unclear. For random outcome assessment and blinding of outcome assessors, none explicitly stated whether animals underwent outcome measurements in randomized order or whether assessors were blinded, rendering the risk of detection bias unclear. Regarding incomplete outcome data, two studies reported animal withdrawals. One study did not specify the reasons for withdrawal, resulting in an unclear risk of attrition bias (26). The other study comprehensively reported withdrawals with no apparent abnormalities in distribution across groups, indicating a low risk of attrition bias (32). The remaining 18 studies reported no animal withdrawals and had complete data, resulting in a low risk of attrition bias (15–25, 27–31, 33, 34). Regarding selective reporting, no studies showed clear signs of selective reporting, indicating a low risk of reporting bias. For other biases, no other significant sources of bias were identified, resulting in a low risk of other biases.

Given the limited number of studies included for each outcome measure (N = 5–16), only univariate meta-regression was employed to explore sources of heterogeneity (multivariate models were not utilized due to insufficient statistical power). Although most variables did not reach statistical significance (P < 0.05), certain covariates demonstrated strong explanatory power regarding heterogeneity. For body weight, pathological state was the primary source of heterogeneity (P = 0.076), accounting for 25.41% of the between-study variation. Regarding gastrocnemius muscle weight, intervention type (P = 0.103) and disease model (P = 0.196) explained 25.44% and 13.35% of the heterogeneity, respectively. For grip strength, although animal species did not reach statistical significance (P = 0.142), it explained 52.67% of the heterogeneity. Regarding muscle cross-sectional area, treatment duration showed a marginal association with effect size (P = 0.158), explaining 40.14% of the heterogeneity. For blood glucose, treatment duration exhibited the highest explanatory power (Adj R² = 24.66%, P = 0.105). Overall, while some covariates explained a portion of the heterogeneity, residual heterogeneity remained high across all models (I²_res > 66%), suggesting the influence of other unaccounted factors. Detailed univariate regression values for each outcome are presented in Supplementary Table S2.

Sensitivity analyses were conducted using the leave-one-out method and by restricting inclusion to studies that explicitly reported randomization methods to verify the robustness of the results. The leave-one-out analysis revealed that the pooled effect sizes for muscle cross-sectional area, gastrocnemius muscle weight, body weight, and blood glucose did not exhibit any directional changes after excluding any single study, indicating the robustness of the results. Although the grip strength index showed some fluctuation, the overall conclusion remained unreversed.

When restricted to studies explicitly reporting randomization methods, the results for gastrocnemius muscle weight and blood glucose were largely consistent with the main analysis, demonstrating good robustness. Notably, for body weight, heterogeneity significantly declined after excluding low-quality studies (I² decreased from 83% to 59%), and the effect approached statistical significance (P = 0.05), suggesting that study quality is a major source of heterogeneity for this outcome. In contrast, grip strength failed to reach statistical significance after restricting by study quality (P = 0.09), indicating weaker robustness and requiring cautious interpretation. Analysis was not performed for muscle cross-sectional area as only one study explicitly reported the randomization method. For specific data, please refer to Supplementary Table S3 and Supplementary Figures S1, S2.

The analysis of publication bias revealed variations across different outcome measures. For body weight, Egger’s test detected no significant publication bias (P = 0.979), and the funnel plot was largely symmetrical. In contrast, significant publication bias was observed for gastrocnemius muscle weight (P < 0.001); after adjustment using the trim-and-fill method, two studies were imputed, reducing the effect size from 2.004 to 1.526, which suggests that the original results may have overestimated the intervention efficacy. Although grip strength (P = 0.033), muscle cross-sectional area (P = 0.005), and blood glucose (P < 0.001) all exhibited evidence of small-study effects, the trim-and-fill analysis showed that effect sizes remained stable before and after imputation (grip strength: 1.702; muscle cross-sectional area: 2.932; blood glucose: -3.384). This indicates that the current body of evidence is relatively complete, and the pooled results are reliable. Nevertheless, caution should be exercised when interpreting the results for gastrocnemius muscle weight and grip strength, considering the influence of potential bias. For specific data, please refer to Supplementary Table S4 and Supplementary Figures S3, S4.

This is the first systematic review and meta-analysis that has evaluated the therapeutic efficacy of TCM interventions in DS animal models. Results suggest that specific TCM interventions may significantly improve core pathological features of DS, including increasing gastrocnemius muscle weight and cross-sectional area, enhancing muscle tissue morphology, and improving grip strength. Additionally, the analysis indicates significant hypoglycemic effects in treated groups. Notably, a potential bidirectional regulatory effect on body weight was observed across different models: interventions appeared to promote weight gain in diabetic cachexia models whilst inducing weight loss in obesity with sarcopenia models. This bidirectional phenomenon may suggest that the metabolic impact of these TCM treatments is context-dependent, offering potential theoretical contributions to personalized diabetes management. Although the outcome measures exhibited considerable heterogeneity in the overall analysis, subgroup analyses and meta-regression identified ‘pathological state’ and ‘animal species’ as key factors contributing to therapeutic variability.

The development of DS is a complex pathological process involving multifactorial, multi-pathway interactions, whose essence appears to lie in the vicious cycle formed by systemic metabolic disorders and dysregulation of intramuscular molecular networks. The core pathophysiological alteration in this disease is thought to be the imbalance in skeletal muscle protein turnover homeostasis—specifically, the suppression of protein synthesis coupled with the overactivation of degradation pathways, which may ultimately lead to net muscle protein loss (35). This protein turnover imbalance is closely associated with key pathological mechanisms including insulin resistance (IR), chronic inflammation, oxidative stress, and mitochondrial damage. These mechanisms likely interact synergistically, collectively driving the pathological progression of skeletal muscle mass reduction, strength decline, and functional deterioration. This study systematically reviewed the potential mechanisms of various TCM intervention in DS, encompassing 20 studies (see Supplementary Material Table S5). These proposed mechanisms observed across different preclinical models can be broadly categorized into three aspects:

High heterogeneity was observed across multiple outcomes in this study, primarily stemming from differences in models and animal strains. Analysis of body weight revealed that different pathological states determined the direction of the effect, suggesting that TCM may possess a bidirectional regulatory characteristic dependent on the organism’s metabolic status. Heterogeneity in gastrocnemius muscle weight, muscle cross-sectional area, and blood glucose was largely concentrated within specific strains (e.g., C57BL/6J), indicating significant variability in response to intervention or substantial background genetic interference in this strain. Heterogeneity in grip strength was not fully explained by animal species, which may be attributed to insufficient standardization of measurement methods and a high proportion of small-sample studies.

Although limited by the number of included studies, meta-regression suggested that factors such as pathological state, intervention type, and treatment duration have potential impacts on effect sizes. However, high residual heterogeneity persisted across all models, implying that experimental details insufficiently reported in the existing literature (such as housing conditions) may be significant hidden sources of interference.

Regarding robustness, most primary outcomes showed consistent effect directions in sensitivity analyses. Furthermore, the restriction to high-quality studies further supported the reliability of the results for gastrocnemius muscle weight and blood glucose. Notably, heterogeneity for body weight decreased significantly after excluding low-quality studies, identifying study quality as a key confounding factor. It should be noted that the results for grip strength demonstrated insufficient stability, warranting cautious interpretation.

Regarding publication bias, although some outcomes suggested the presence of small-study effects, effect sizes remained substantially unchanged after adjustment using the trim-and-fill method, indicating good overall credibility of the evidence. While the effect size for gastrocnemius muscle weight decreased after adjustment, it remained statistically significant, suggesting that although the original results may have been somewhat overestimated, the conclusion remains valid.

Overall, despite heterogeneity caused by differences in models and methodologies, multiple robustness tests support the reliability of the study’s main conclusions. Future research should focus on standardizing model construction protocols and improving the reporting of experimental details to generate more reliable and translatable preclinical evidence.

1) Limited sample size and methodological quality: The sample sizes of the included studies were generally small, and reporting on randomization, allocation concealment, and blinding (for both caregivers and outcome assessors) was universally unclear across all studies. These methodological flaws not only compromise the reliability of the evidence but may also exert a directional influence on the results: the lack of allocation concealment predisposes the studies to selection bias, while the absence of blinding may lead to performance bias and detection bias. In animal research, such biases have been proven to systematically overestimate the effect sizes of interventions. Consequently, the pooled effect estimates presented in this meta-analysis may be higher than the true therapeutic effects, necessitating a cautious interpretation of the conclusions; 2) Potential language bias: The included studies were primarily sourced from Chinese databases, with a low proportion of English literature, which may compromise the comprehensiveness of the results; 3) Variations in TCM formulas and interventions: Significant differences exist across studies regarding herbal ingredients, dosages, and preparation methods. These variations increase inter-study incomparability and heterogeneity; 4) Inconsistent disease severity: Differences in the duration and severity of the disease in animal models may affect the stability of the effect sizes; 5) High heterogeneity of outcome measures: Although subgroup analyses identified pathological state and animal strain as primary confounding factors, residual heterogeneity remains high. This likely stems from insufficiently reported experimental conditions (e.g., housing environment, dosage details, and detection methods); 6) Limitations in clinical extrapolation: Due to differences in pathological complexity and metabolic characteristics between animal models and humans, caution is required when translating these results to clinical practice.

1) Rigorous methodological implementation: Future studies must strictly adhere to the SYRCLE tool and ARRIVE guidelines, strengthening the implementation and reporting of critical methodological steps such as randomization and blinding, while reasonably expanding sample sizes. Concurrently, animal model selection should be optimized to use models that better match human metabolic characteristics and disease stages, ensuring a systematic evaluation of intervention effects and guaranteeing result reliability and stability; 2) Promoting TCM standardization and mechanistic research: It is recommended to gradually unify the ingredients, dosages, and preparation processes of TCM formulas in both preclinical and clinical studies to enhance comparability and reduce heterogeneity. Mechanistic research should transcend simple efficacy verification to systematically elucidate multi-target interaction networks. In particular, the causal mechanisms underlying model-dependent effects, such as the bidirectional regulation of body weight, should be explored in depth; 3) Clinical translation and implementation strategies: Building upon reliable preclinical evidence, rigorously designed clinical trials should be advanced. It is recommended to stratify patients based on metabolic phenotypes (e.g., obesity status) and adopt multi-center, randomized, double-blind designs. Comprehensive assessments should include multidimensional indices such as muscle mass, grip strength, and physical function to provide high-level evidence-based support for the efficacy and safety of TCM in treating diabetic sarcopenia.