A systematic search of five electronic databases (PubMed, Web of Science, The Cochrane Library, Scopus, and Embase) was conducted up to 3 September 2025 using a combination of MeSH terms and free-text keywords related to “sepsis” and “artesunate” (Supplementary Table S1). The reference lists of relevant articles were manually screened.

Studies were included based on the PICOS framework, where, P (population): in vivo septic animal models; I (intervention): AS intervention; C (control): vehicle, saline, or no treatment; O (outcomes) primary outcome: survival rate, secondary outcomes: colony-forming unit (CFU), body weight, serum levels of inflammatory cytokines [tumor necrosis factor-alpha (TNF-α), interleukin-6 (IL-6), and interleukin-1 beta (IL-1β)], markers of organ injury [aspartate aminotransferase (AST), alanine aminotransferase (ALT), lung injury score, myeloperoxidase (MPO) activity, and wet-to-dry (W/D) ratio], and apoptosis-related biomarkers [terminal deoxynucleotidyl transferase dUTP nick end labeling (TUNEL)-positive cells, the ratio of phosphorylated mechanistic target of rapamycin to total mTOR (p-mTOR/mTOR), and caspase-3 expression]; S (study design): controlled animal studies.

The exclusion criteria were as follows: 1. the use of combined therapies that could confound the results; 2. incomplete or unavailable data for meta-analysis; 3. secondary literature types (reviews, abstracts, etc.); and 4. duplicate publications, with only the most complete dataset retained.

Two researchers independently extracted data (publication details, animal characteristics, modeling method, AS administration, and outcomes) and resolved discrepancies with a third reviewer. Original data from figures were extracted by contacting authors or via WebPlotDigitizer. In studies involving multiple intervention doses or time-points—resulting in several datasets for the same outcome—the dataset corresponding to the most effective dose or time-point was selected for meta-analysis. Methodological quality was assessed using the Systematic Review Centre for Laboratory Animal Experimentation (SYRCLE) RoB tool (for animal studies) (Hooijmans et al., 2014) and Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach (https://gdt.gradepro.org/app/) (Guyatt et al., 2008; Balshem et al., 2011). Two reviewers conducted assessments independently, with discrepancies resolved via discussion.

The meta-analysis was conducted using Review Manager 5.4 (Cochrane Collaboration, Oxford, UK) and Stata 17.0 (StataCorp LLC) software. Based on data types, the corresponding effect measures were applied. For dichotomous outcomes, odds ratios (ORs) with 95% confidence intervals (CIs) were calculated and pooled using the Mantel–Haenszel method. For continuous variables, mean differences (MD) with 95% CIs were computed. A p-value < 0.05 was considered statistically significant. Heterogeneity among studies was assessed using the chi-square test (χ²) and I² statistic. A fixed-effect model was applied when p > 0.1 and I² < 50% (Pereira et al., 2010; Thorlund et al., 2012; Hoaglin, 2016), indicating low heterogeneity; otherwise, a random-effects model was used (Borenstein et al., 2010). Publication bias of the included studies was evaluated using a funnel plot, Egger’s test (Egger et al., 1997), and the trim-and-fill method (Peters et al., 2007) based on standard error (Lau et al., 2006) when >9 studies were included for a single outcome measure. We performed sensitivity analyses in outcomes with more than two studies to investigate the influence of individual studies on the overall pooled effect size.

The methodological quality of the 15 included studies was assessed using the SYRCLE risk-of-bias tool, with the detailed results visualized in Figure 2 and summarized in Supplementary Table S2. Reporting of key methodological details was generally incomplete. Only one study adequately described random sequence generation, and none reported sufficient information regarding allocation concealment. Blinding of personnel and outcome assessors was explicitly stated in only one study, and no study described random housing or blinding during outcome assessment. While all studies reported basic animal characteristics, none provided baseline data for specific outcome measures. All studies were considered to be at low risk for incomplete outcome data and selective reporting. Nevertheless, the widespread unclear or high risk of bias in critical domains such as allocation concealment and blinding underscores substantial methodological limitations in the available evidence. These shortcomings led to a reduction in the overall certainty of evidence in the GRADE assessment, which was predominantly rated as low or moderate for critical outcomes (Supplementary Table S3).

The basic characteristics of the included studies are summarized in Table 1. The relevant publications were issued between 2008 and 2025, with sample sizes ranging from 20 to 168. The sepsis animal models in the included studies were primarily induced by cecal ligation and puncture (CLP) or intraperitoneal injection of lipopolysaccharide (LPS), whereas some studies utilized bacterial injection or a combination of methods. The animals used included various strains of mice and rats. Regarding the intervention, the effective dose of AS ranged from 5 to 300 mg/kg. Administration was predominantly via the intraperitoneal route (7/15 studies), with the remaining studies using intravenous, intramuscular, intratracheal, or oral routes. The timing of intervention varied relative to sepsis induction, including pretreatment, concurrent administration, and post hoc treatment at time-points ranging from immediately to after 72 h. These variabilities are critical considerations when interpreting the pooled results and their translational implications. All studies reported sepsis-related outcomes, with survival rate and systemic inflammatory cytokine levels being the most frequently assessed parameters.

The analysis of ten studies (Li et al., 2008; Li et al., 2010; Jiang et al., 2011; Li et al., 2014; Cao et al., 2016; Shang et al., 2020; Bang et al., 2021; Chen et al., 2023; He et al., 2023; Yuan et al., 2023) provides evidence that AS improved survival in septic animals compared with the controls [OR: 6.87 (95% CI: 3.81, 12.41), p < 0.00001; heterogeneity: I² = 0%, p = 0.84, GRADE of evidence: moderate; Figure 3A]. Data from two studies (Chen et al., 2023; Chen et al., 2025) indicated an increase in body weight associated with AS treatment [MD: 1.75 (95% CI: 0.05, 3.46), p = 0.04; heterogeneity: I² = 65%, p = 0.09, GRADE of evidence: low; Figure 3B]. Furthermore, a decrease in bacterial load was observed across four studies (Jiang et al., 2011; Liu et al., 2020; Shang et al., 2020; Yuan et al., 2023) [MD: −1.68 (95% CI: −3.10, −0.27), p = 0.02; heterogeneity: I² = 81%, p = 0.001, GRADE of evidence: low; Figure 3C].

Meta-analysis of three studies (Cao et al., 2016; Zhang et al., 2020; Liu et al., 2023) indicates that AS attenuated lung injury, as evidenced by reduced histological scores [MD: −6.97 (95% CI: −8.98, −4.97), p < 0.00001; heterogeneity: I² = 75%, p = 0.02, GRADE of evidence: low; Figure 4A]. Data from two studies (Zhang et al., 2020; Liu et al., 2023) showed a decrease in pulmonary myeloperoxidase activity [MD: −0.16 (95% CI: −0.23, −0.10), p < 0.00001; heterogeneity: I² = 87%, p = 0.006, GRADE of evidence: low; Figure 4B]. Furthermore, analysis of three studies (Cao et al., 2016; Zhang et al., 2020; Liu et al., 2023) suggests a reduction in the wet-to-dry ratio [MD: −2.68 (95% CI: −3.76, −1.60), p < 0.00001; heterogeneity: I² = 66%, p = 0.05, GRADE of evidence: moderate; Figure 4C].

AS treatment appears to confer hepatoprotection, as shown by a marked reduction in serum aspartate aminotransferase levels across three studies (Li et al., 2014; He et al., 2023; Yang et al., 2024) [MD: −113.31 (95% CI: −144.64, −81.99), p < 0.00001; heterogeneity: I² = 39%, p = 0.20, GRADE of evidence: moderate; Figure 5A]. Similarly, a decrease in alanine aminotransferase levels was observed in the same three studies [MD: −56.49 (95% CI: −85.00, −27.90), p = 0.0001; heterogeneity: I² = 67%, p = 0.05, GRADE of evidence: moderate; Figure 5B].

When synthesizing data grouped by sampling time-points, a pattern of phase-dependent regulation was retrospectively inferred. AS demonstrated bidirectional modulation of systemic inflammation. During the late immunosuppressive phase, AS restored circulating levels of TNF-α [three studies (Liu et al., 2020; Shang et al., 2020; Yuan et al., 2023); MD: 102.16 (95% CI: 33.06, 171.27), p = 0.004; heterogeneity: I² = 90%, p < 0.00001, GRADE of evidence: low; Figure 6A] and IL-6 [two studies (Shang et al., 2020; Yuan et al., 2023); MD: 422.42 (95% CI: 99.28, 745.55), p = 0.01; heterogeneity: I² = 3%, p = 0.31, GRADE of evidence: moderate; Figure 6C]. Conversely, during the early hyperinflammatory phase, it suppressed the levels of both TNF-α [six studies (Li et al., 2008; Li et al., 2010; Jiang et al., 2011; Li et al., 2014; Cao et al., 2016; He et al., 2023); MD: −70.29 (95% CI: −101.92, −38.68), p < 0.00001; heterogeneity: I² = 96%, p < 0.00001, GRADE of evidence: low; Figure 6B] and IL-6 [four studies (Jiang et al., 2011; Li et al., 2014; Cao et al., 2016; He et al., 2023); MD: −227.42 (95% CI: −386.12, −168.71), p < 0.00001; heterogeneity: I² = 84%, p = 0.0003, GRADE of evidence: low; Figure 6D]. Additionally, AS treatment elevated IL-1β levels across two studies (Shang et al., 2020; Yuan et al., 2023) [MD: 58.67 (95% CI: 34.21, 83.14), p < 0.00001; heterogeneity: I² = 0%, p = 0.49, GRADE of evidence: moderate; Figure 6E]. Additionally, although one study (He et al., 2023) reported a reduction in IL-1β levels with AS treatment, meta-analysis was not performed due to the inclusion of only a single study. Analyses of cytokine levels and bacterial load in the lung and spleen tissues (Supplementary Figures) further support the time-dependent immunomodulatory effects and effective clearance of local infection by AS.

Meta-analysis of apoptosis-related biomarkers provides preliminary evidence that AS attenuated apoptosis in septic models. AS treatment markedly reduced the number of apoptotic cells [two studies (Zhang et al., 2020; Liu et al., 2023), MD: −43.80 (95% CI: −50.17, −37.43), p < 0.00001; heterogeneity: I² = 0%, p = 0.90, GRADE of evidence: moderate; Figure 7B] and enhanced the phosphorylation of the pro-survival regulator mTOR, as indicated by an increased p-mTOR/mTOR ratio [two studies (Zhang et al., 2020; Chen et al., 2025), MD: 0.61 (95% CI: 0.46, 0.77), p < 0.00001; heterogeneity: I² = 89%, GRADE of evidence: very low; Figure 7C]. In contrast, the overall effect on caspase-3 expression was not statistically significant [two studies (Zhang et al., 2020; Yuan et al., 2023), MD: −3.17 (95% CI: −8.74, 2.39), p = 0.26, GRADE of evidence: very low; Figure 7A], with substantial heterogeneity (I² = 99%) that is likely attributable to methodological variations in protein normalization.

All included studies involved the use of AS for treating septic animal models. Using survival outcomes as the main variable, the included studies were evaluated for the effect of the study size. The funnel plot was largely symmetrical (Figure 8). Although Egger’s test was not statistically significant (p = 0.053, Table 2), the statistical power of both the funnel plot and Egger’s test is insufficient due to the small number of included studies, and thus, publication bias cannot be fully excluded. Given its marginal significance and to preclude any potential bias, a trim-and-fill analysis was conducted. This analysis imputed two missing studies, and the adjusted effect size (HR = 1.646) was nearly identical to the original (HR = 1.807) and remained highly significant (Table 3), indicating that the primary survival outcome is relatively robust despite the inability to rule out publication bias.

The results of the sensitivity analysis were evaluated by systematically excluding individual studies and comparing the recalculated pooled effect sizes with the overall estimates. As shown in Figure 9, no significant differences were observed between the recalculated results and the original overall results across all outcome measures. This consistency indicates that the findings of our study are robust and are not unduly influenced by any single dataset.

This meta-analysis provides preliminary evidence that AS improves survival and exerts multidimensional protection in septic animal models. AS was associated with treatment-enhanced survival rates while reducing bacterial load and promoting body weight recovery. The agent attenuated multiple organ injuries, evidenced by improved lung histopathology and liver function biomarkers. Furthermore, analysis stratified by time-point suggested that AS exerts phase-dependent immunomodulation, being associated with the suppression of early hyperinflammation and potential restoration of immune function in later stages, accompanied by reduced apoptosis and enhanced autophagy. For the primary survival outcome, publication bias assessment did not show statistically significant evidence of bias, but the small number of included studies led to insufficient statistical power for the funnel plot and Egger’s test; thus, publication bias cannot be fully excluded. Sensitivity analysis demonstrated that no individual study unduly influenced the overall effect estimates.

AS exerts its therapeutic efficacy against sepsis through a multimodal mechanism. The following synthesis distinguishes between effects supported by the present meta-analysis and deeper mechanistic insights that are primarily derived from individual preclinical reports. Although the latter provide valuable hypotheses for specific molecular pathways, they require further independent replication and validation. This integrated perspective encompasses phase-dependent immunomodulation, organ-protective effects, regulation of apoptosis and autophagy, and direct antibacterial and endotoxin-clearing actions (Figure 10).

This study is the first systematic review to evaluate the efficacy of AS in septic animal models, providing a solid evidence base for clinical trial design and novel therapeutic development. It strictly adhered to PRISMA guidelines, used the SYRCLE tool and GRADE approach to assess methodological quality and evidence certainty respectively, registered the protocol in PROSPERO, and conducted publication bias and sensitivity analyses for key outcomes, enhancing the robustness of findings.

The limitations of this study should be considered from the following aspects. First, the reporting of critical methodological elements in the included studies, such as randomization, allocation concealment, and blinding, was generally incomplete, which may lead to an overestimation of the true treatment effect and necessitates cautious interpretation of the pooled effect estimates. The overall certainty of evidence, as reflected in the low–moderate GRADE ratings, directly stems from these methodological shortcomings. Consequently, the internal validity and, thus, the translational validity of this evidence base are constrained. Although the pooled results are promising, they should be interpreted as preliminary and hypothesis-generating, underscoring the need for future preclinical studies adhering to higher methodological standards to strengthen the evidence for potential clinical translation.

Second, substantial statistical heterogeneity (I² > 50%) was observed for several critical outcomes, including bacterial load, lung injury scores, and cytokine levels. Although random-effects models and sensitivity analyses were utilized, heterogeneity constrained interpretation, and subgroup analyses were not feasible due to the limited number of studies (n < 3) per comparison. This heterogeneity predominantly stems from variations in animal species, sepsis modeling methods, and AS intervention protocols. The presence of such marked heterogeneity inherently reduces the certainty with which a single, precise treatment effect can be defined. Consequently, the point estimates for these outcomes, while suggesting a directional benefit of AS, should be interpreted with caution. They are best viewed as indicative of a potential treatment signal within a range of plausible effects rather than as precise measures of efficacy. This limitation underscores that our findings primarily highlight the need for and inform the design of more standardized future investigations. Furthermore, the considerable heterogeneity in AS dosing, route, and timing of administration across studies (Table 1) limits the comparability of the findings and complicates the direct translation of a specific therapeutic protocol. The lack of standardized protocols in preclinical studies introduces significant methodological variability, which limits the clinical relevance and translational interpretation of pooled findings. To bridge this translational gap, future research should prioritize standardized models that enhance clinical applicability. For instance, using intraperitoneal administration—the most common route in existing studies—and initiating treatment within a narrow, clinically feasible window (e.g., 1 h–4 h post-sepsis induction) would better simulate emergency intervention scenarios. Adopting such standardized, clinically aligned dosing regimens is essential to more definitively evaluate the therapeutic potential of AS and to inform the design of subsequent clinical trials.

Third, the exclusion of gray literature and reliance on published studies may have introduced publication bias. Although funnel plot symmetry and Egger’s test did not indicate significant bias, the relatively small number of studies limits the statistical power of these assessments, and thus, publication bias cannot be entirely excluded. This limitation is compounded by the substantial heterogeneity in experimental designs (e.g., dose and timing) across studies, which may itself be a source of selective reporting and could influence the magnitude and direction of the pooled effect estimates. To mitigate this issue in future research and improve the reliability of evidence synthesis, preclinical studies should aim for standardized, dose-ranging experimental designs that are more directly comparable and less susceptible to such biases. Fourth, when studies reported multiple doses or time-points, we selected the most effective dataset for synthesis. Although this strategy was intended to capture the maximal therapeutic potential of AS and facilitate a clearer synthesis, we acknowledge that it may introduce a selection bias and lead to an overestimation of the overall treatment effect. This constitutes an inherent limitation of our analysis. Additionally, although data extraction from figures was performed meticulously using standardized software, potential measurement errors cannot be entirely ruled out. Finally, as the conclusions are derived from animal studies, clinical translatability requires validation through further preclinical and clinical investigations. Nonetheless, this meta-analysis offers a systematic summary of current research in the field and provides a valuable reference to support the further development of AS for the treatment of sepsis.