A systematic search was conducted across PubMed, Embase, Cochrane, WOS, China National Knowledge Internet (CNKI), VIP Information Chinese Periodical Service Platform (VIP), China Biology Medicine disc (CBM disc), and Wanfang Data Knowledge Service Platform (Wanfang) databases. The search period encompassed the whole duration of each database, starting from its establishment and concluding in May 2025. The primary objective of the investigation centered on studying the effect of LE and its active components on animals suffering from atopic dermatitis using a Boolean search strategy with the operators “AND,” “OR,” and “NOT”: atopic dermatitis, atopic neurodermatitides, Lithospermum, Radix Arnebiae, Zicao, shikonin, and acetylshikonin (the search syntax can be found in Supplementary Material). This search was conducted without any limitations on the language, country, or publication.

To mitigate potential publication bias and the effects of access restrictions due to institutional journal subscriptions, the following supplementary strategies were implemented: searching Google Scholar and ProQuest Dissertations & Theses Global for unpublished studies, conference abstracts, and academic theses that could have useful information. We used several methods to obtain the full articles for records that were chosen for full-text review but were not available through our institutional subscriptions. These methods included direct requests to the authors, searches on academic social networks such as ResearchGate, and interlibrary loan services.

This meta-analysis aimed to evaluate the efficacy of interventions derived from Zicao or its primary bioactive constituents (e.g., shikonin) in animal models of atopic dermatitis. The review included randomized controlled animal trials that reported on symptomatic improvement and associated mechanisms: (1) AD animal models, without any limitations on gender, species, or method of modeling; (2) treatment groups using Zicao (defined as the root of Lithospermum erythrorhizon Sieb. and Zucc., Arnebia euchroma (Royle) Johnst., or Arnebia guttata Bunge) in the form of crude extracts, its purified chemical derivatives (e.g., shikonin and acetylshikonin), or compound preparations in which Zicao serves as the principal component. There were no restrictions on the dosage, formulation, route of administration, or treatment duration. (3) The control group received either the vehicle alone (vehicle control) or were left untreated (untreated control). (4) The primary outcome included dermatitis severity (e.g., SCORing Atopic Dermatitis (SCORAD) and Eczema Area and Severity Index), scratching behavior (including scratch frequency, scratch duration, and scratching intensity). (5) The secondary outcome included histological alterations in the skin, immune-associated cytokines, and other relevant outcome measures. The exclusion criteria were as follows: (1) studies involving non-mammalian species, in vitro experiments, animal models of comorbidities, human studies, and in silico studies; (2) reviews, case reports, patents, and commentaries; (3) without a control group or only including before-and-after controlled studies in the treatment group; (4) studies where LE and its derivatives were not the main intervention; and (5) no reporting of the outcome.

These works were independently examined by two writers using the pre-established inclusion and exclusion criteria. The process started with a brief review of the titles and abstracts, followed by an in-depth examination of the entire text in order to evaluate its content. Any disagreements during screening were resolved by consensus. The extracted data were then recorded in a standardized Excel spreadsheet. The extracted information included (1) the title; (2) the first author and the publication date; (3) animals (species, sex, number, age, and weight); (4) the approach of disease modeling; (5) medication administered to the experimental group, including the name of the medication, the dose, and the dosing frequency; (6) the non-functional substances in the control group (the name, dose, and dosing frequency); (7) the intervention and the duration; and (8) the outcome: the data for the outcome were expressed as the mean ± SD. For data presented only in graphical form, numerical values were digitized using GetData Graph Digitizer (version 2.26.0.20), and to ensure accuracy, all digitized values were independently extracted by two reviewers. If conflicts arose, a third reviewer was involved in resolving them. For studies featuring multiple intervention groups with varying doses, the outcomes from these groups were pooled to create a single, combined estimate. This procedure, endorsed by the Cochrane Handbook, prevents unit-of-analysis errors by ensuring that no participants are double-counted in the meta-analysis (Cochrane, 2004). If the data presented in the text were initially expressed as standard errors (SEs), they were afterward transformed into standard deviations (SDs) (Lee et al., 2015).

Additionally, all medicinal plant species mentioned in this systematic review were taxonomically validated to ensure nomenclatural accuracy. The accepted scientific name, author citation, and family for each species were verified using the Plants of the World Online (POWO) database and the Medicinal Plant Names Services (MPNS) portal. The standard pharmacopeial drug name (where applicable) is also provided.

The Systematic Review Center for Laboratory Animal Experimentation (SYRCLE) risk-of-bias tool was used for bias assessment (Hooijmans et al., 2014). It assesses bias in the experimental design, execution, measurement, reporting, and other sources, each of which is categorized into three risk-of-bias levels: “Yes,” meaning low risk, “No,” meaning high risk, and “Unclear,” meaning unclear. The assessment of the reporting quality of the included studies was performed using a scoring system adapted from Liu et al. (2024), where each domain received a score of 1 = Yes, 0 = No, 0.5 = partly, and NA = not applicable, contributing to a total score of 10 points. Two authors assessed the studies independently. Any disagreements were resolved by consultation with a third reviewer.

Statistical analysis was performed using R version 4.3.2 software (the {meta} package). The outcome indicators were represented as continuous variables; the combined total effect sizes were presented using the standardized mean difference (SMD) and 95% confidence intervals (95% CI). A p-value < 0.05 indicates statistical significance. Consolidation of subgroup effect sizes via Hedges’ g—in accordance with Cochrane Collaboration guidelines—reduced bias attributable to the duplicate use of shared control group data in SMD calculations. A random-effects model was used since the actual effect size varied across studies because of diverse animal and treatment factors (Muka et al., 2020). Confidence intervals for the pooled effect estimates were computed using Knapp–Hartung adjustments (Harrer et al., n.d.).

Heterogeneity was quantitatively assessed through three complementary metrics: the percentage of variability in the effect sizes not attributable to sampling error (I²), the variance of the distribution of true effect sizes (Tau²/τ²), and the prediction intervals (PIs) (Harrer et al., n.d.; IntHout et al., 2016). The magnitude of heterogeneity was quantified by (1) Cochrane Handbook thresholds: I² = 25% (low), 50% (moderate), and 75% (substantial); and (2)τ² (the DerSimonian–Laird method estimator), which represents the variance of the true effect sizes across studies, quantifying the absolute magnitude of between-study heterogeneity. τ² = 0 indicates the absence of between-study heterogeneity (all observed variability is attributable to sampling error alone). τ² > 0 signifies the presence of between-study heterogeneity, with larger values corresponding to greater heterogeneity magnitude. (3) Computation of the 95% prediction interval bounds for the true effect size. Using the prediction interval delineates the expected range of effects in future studies based on current evidence. Crucially, when the interval lies entirely on the intervention-beneficial side of the null line, it predicts consistent therapeutic benefit despite heterogeneity. Conversely, intervals spanning the null value (e.g., risk ratio = 1 or mean difference = 0) indicate uncertain clinical relevance.

Subgroup analyses were restricted exclusively to the primary outcome measures. Systematic investigation of outcome variation across experimental settings and sources of heterogeneity was performed using predetermined subgroup criteria, including, but not limited to, animal species, modeling method, drug formulation, drug source, and intervention protocol. We used the Influence Analysis function from the {dmeta} package (R) to perform the sensitivity analysis and assess the presence of outliers, the former using the leave-one-out method to examine the effect of each finding on the effect size and I² one-by-one effect (Viechtbauer and Cheung, 2010). Publication bias assessment for the primary outcomes utilized dual analytical approaches: visual inspection of funnel plot symmetry and the quantitative Egger’s linear regression test.

A thorough search across eight databases found a total of 445 relevant studies. Among them, 18 papers were obtained from PubMed, 38 papers were from Embase, 4 papers were from Cochrane, 43 papers were from Web of Science (WOS), 29 papers were from Wanfang, 17 papers were from CNKI, 69 papers were from CQVIP, 27 papers were from China Biology Medicine (CBM), 188 papers were from ProQuest, and 12 papers were from Google Scholar. Following relevant exclusions, a total of 12 studies ultimately met all the inclusion criteria. The literature selection process is shown in Figure 1.

Following systematic screening, 12 studies met the final inclusion criteria (Oh et al., 2021; Kadoyama et al., 2019; Ku et al., 2018; Choi et al., 2017; Lee et al., 2009; Kim et al., 2009; Wu et al., 2021; Xiaochao et al., 2019; Liyuan et al., 2019b; Xu et al., 2019), comprising four Chinese and six English publications. The experimental models utilized NC/Nga mice (k = 4), BALB/c mice (k = 4), KM mice (k = 1), SD rats (k = 2), and unspecified murine species (k = 1); the detailed animal characteristics (sex/age/weight) are tabulated in Table 1. The majority of studies (k = 8) employed dinitrochlorobenzene (DNCB) sensitization to induce dermatitis. Other induction methods comprised ovalbumin (OVA) challenge (k = 1), oxazolone (OX) exposure (k = 1), a spontaneous model (k = 1), and the combined use of Biostir AD ointment with 4% SDS (k = 1). The therapeutic interventions involved commercial Chinese polyherbal preparation (k = 6), shikonin (k = 2), an investigator-formulated botanical preparation (k = 1), and standardized extracts (k = 3), which were administered via oral gavage (k = 4) or external application (k = 8). Furthermore, none of the included studies reported any toxicological or safety assessments for the administered preparations.

The risk-of-bias assessment for all 12 included studies (summarized in Figure 2) revealed critical methodological limitations: (1) allocation methodology was unreported in one study (Lee et al., 2009), while only three of the remaining nine studies detailed random sequence generation procedures (Wu et al., 2021; Liyuan et al., 2019b; Xu et al., 2019); (2) allocation concealment was universally unreported; (3) all studies failed to report the blinding of investigators, randomization implementation, or outcome assessor masking; (4) one study reported the death of animals during disease modeling after randomization but did not provide details regarding the missing data from dead animals and the potential impact on the outcomes (Liyuan et al., 2019b); and (5) one included study (Lee et al., 2009) reported sample sizes with uncertainty, indicating ranges of n=(7–8) for both the experimental and control groups. In accordance with the principle of conservative estimation when handling imprecise numerical data in meta-analysis, we adopted the lower bound of the reported range (n = 7) for effect size calculations. This approach minimizes the risk of overestimating the precision in pooled effect estimates as using smaller sample sizes yields wider confidence intervals and appropriately reflects the underlying uncertainty in the primary data. (6) As detailed in the table of the included study characteristics (Table 1), all studies were assessed as having a potential risk of bias.

The reporting quality was assessed with a 10-item checklist, with scores ranging from 3 to 10 (Figure 3). Among 12 studies, one study received a score of 10 (Liyuan et al., 2019b). All studies demonstrated high quality, with 78.3% of the parameters receiving a “yes” response. Nine studies (75%) clearly defined animal characteristics (species, age, sex, and weight). A total of 11 studies (91.6%) clearly reported the number of animals per group at the start of the experiment. All studies clearly defined the disease model, with nine studies (75%) providing validation of the model. Six studies (50%) clearly defined the study drug preparation method, 10 studies (83.3%) clearly defined the dosages, and all studies defined the administration route and treatment duration. While the reporting of methodological procedures such as randomization was limited (reported in 25% of studies), all included studies provided well-defined outcome measures and complete datasets, allowing for their inclusion in the quantitative synthesis.

The outcomes are organized and presented according to the core pathophysiological axes of AD: beginning with metrics of skin barrier integrity, followed by an analysis of cytokines related to Th1/Th2, and measures of important effector cells and systemic sensitization (mast cells and IgE levels in the serum).

For SCORAD and EASI, the influence diagnostics via R-based outlier assessment and leave-one-out sensitivity analysis identified Kadoyama et al. (2019) as the primary contributor to substantial heterogeneity. Exclusion of these outliers yielded a marked reduction in heterogeneity indices: I² decreased from 84.0% to 55.5%, while τ² = 0.5931 (95% CI: 0.0000–5.6277) encompassed the null value of 0, and Cochran’s Q heterogeneity test remained statistically significant (p = 0.0278).

In addition, elimination of studies with disproportionate effect weights (SMD = −3.5431, CI: 4.3086, -2.7776; p < 0.0001) did not significantly alter the primary outcome trajectory, thus further validating the result stability (Supplementary Material).

Enhanced funnel plot analysis revealed asymmetry, with Egger’s test indicating potential subtle publication bias (intercept = −4.057, 95% CI: −6.95 to −1.16; p = 0.0286). Trim-and-fill imputation added four studies: one significant study (Oh 2021-like) at p < 0.1 and three non-significant studies in the funnel’s midzone, yielding an adjusted estimate of SMD = −2.2983 (95% CI: −3.31 to −1.2847; p < 0.0001; I² = 84.8%) (Figure 19).

This systematic review and meta-analysis includes data from 12 studies involving 316 animals, evaluating the therapeutic profile of Zicao and its primary bioactive naphthoquinones (e.g., shikonin) in experimental models of AD.

The primary outcomes demonstrated significant treatment effects on the dermatitis severity scores (SMD = −3.30, I² = 84.0%) and scratching behavior (SMD = −2.60, I² = 79.3%). However, the prediction intervals for both crossed the null value (dermatitis: −6.80 to 0.20; scratching: −7.76 to 2.57), indicating that while the pooled estimates indicate substantial average efficacy, the effects in the subsequent studies may vary from significantly beneficial to null or oppositely directed under varying experimental conditions. Subgroup analyses of dermatitis severity scores indicated the AD induction model as a statistically significant effect modifier (p = 0.009), while DNCB-induced models demonstrated reduced heterogeneity (I² = 35.9%). However, the uneven distribution of studies among subgroups and limited statistical power precluded definitive conclusions regarding the influence of the formulation, route of administration, botanical source, or host species. Sensitivity analysis excluding one influential outlier reduced heterogeneity to I² = 55.5% while maintaining significant effects (SMD = −3.5431, CI: 4.3086,-2.7776; p < 0.0001). Trim-and-fill adjustment for detected publication bias (Egger’s test, p = 0.0286) yielded a more conservative, though still statistically significant, estimate (adjusted SMD = −2.30, p < 0.0001), supporting the general robustness of the primary findings.

For IL-4, a highly consistent inhibitory effect was observed at both the protein level in the serum (SMD = −1.46, I² = 0%) and the mRNA level in skin tissue (SMD = −3.28, I² = 75.4%). The former analysis demonstrated lower heterogeneity, with a prediction interval (−2.02 to −0.90) that entirely excluded 0 and remained below the null value, suggesting that future individual studies are likely to yield effects in the same direction, whereas the latter analysis revealed a larger effect size alongside substantial heterogeneity. Similarly, a potent and homogeneous reduction in TSLP was confirmed in the serum (SMD = −2.13, I² = 0%). In stark contrast, the effect on IL-13 mRNA, while significant on average (SMD = −2.35), was markedly heterogeneous (I² = 70.1%) with a prediction interval (−18.72 to 14.02); this wide-spanning interval, which crosses the null value, indicates the potential for future studies to demonstrate null or even opposing effects. TNF-α was reduced at the mRNA level in the skin (SMD = −2.66, I² = 10.6%) and the protein level in the serum (SMD = −2.68, I² = 86.7%), whereas IL-6 exhibited non-significant effects (SMD = −2.62, p = 0.16) with extreme heterogeneity (I² = 92.9%; prediction interval: −43.05 to 37.81), rendering the pooled estimate uninterpretable. Conversely, an increase in IFN-γ protein was observed (SMD = 2.37, p = 0.001; I² = 84.1%) with a prediction interval (−3.41–8.14); the pronounced heterogeneity combined with a null-crossing prediction interval suggests that this finding may lack robustness, and future individual studies may plausibly demonstrate null or contradictory results.

Beyond the preclinical evidence presented in our meta-analysis, some clinical studies indicated the therapeutic potential of Zicao (the root of LE Siebold and Zucc., Arnebia euchroma (Royle) Johnst., or Arnebia guttata Bunge) in AD. A clinical trial found that an ointment containing shikonin significantly improved dermatitis and reduced the EASI score in AD patients (Yen and Hsieh, 2016). In addition, the same research team used human dendritic cells derived from AD patients and showed that shikonin potently suppressed the expression of multiple allergen-induced pro-inflammatory cytokines and chemokines, with its inhibitory effect on IL-9, MIP-1β, and RANTES exceeding that of dexamethasone (Chung-Yang et al., 2020). In addition, a randomized controlled trial by Cho et al. (2008) involving 28 AD subjects showed that oral administration of LE extract (1.5 g daily for 10 weeks) significantly increased stratum corneum hydration and ceramide levels compared to that with placebo, with improvements positively correlating with baseline disease severity. Another clinical trial also indicated that a compound Zicao oil preparation had comparable efficacy to pimecrolimus for symptomatic relief, alongside a more favorable safety profile (Liyuan et al., 2019a).

It is worth noting that the barrier repair effects observed in the human trial by Cho et al. (2008) appear to contrast with the insignificant findings for epidermal thickness and filaggrin expression in our meta-analysis. This discrepancy may be attributed to several factors, including potentially limited sample sizes and animal model differences in skin physiology and disease manifestation.

The variety of AD modeling methods, each focusing on different pathophysiological mechanisms, represents a potential confounding factor affecting therapeutic efficacy outcomes. The substantial heterogeneity observed across most indicators in this meta-analysis, along with subgroup analyses of the primary outcome measure (dermatitis score), suggests that AD animal modeling is a statistically significant effect modifier. It requires an extensive examination of modeling methodologies.

There are three main types of AD animal models: The first category consists of hapten-induced models, which use repeated applications of haptens [oxazolone, DNCB (2,4-dinitrochlorobenzene)] to induce a transition from delayed-type hypersensitivity to Th2-dominated chronic inflammation, characterized by elevated levels of Th2-associated cytokines, including IL-4, IL-5, and IL-13. The widely used DNCB-induced AD model causes disease by directly stimulating keratinocytes and sensory neurons and causing keratinocytes to release cytokines and chemokines. The molecular mechanisms behind this include the mitogen-activated protein kinase (MAPK)/nuclear factor-κB (NF-κB) and erythroid 2-related factor 2 (Nrf2)/heme oxygenase-1 (HO-1) signaling pathways, along with the involvement of several immune cell types, such as Th1, Th2, mast cells, and eosinophils.

The oxazolone (OX)-induced model primarily operates through the TRPV1, MAPK-AP1, and ROS pathways to promote inflammation. While DNCB modeling emphasizes MAPK/NF-κB signaling, OX modeling predominantly activates the MAPK-AP1 pathway (Jin et al., 2009).

The second category includes allergen-induced models, such as ovalbumin (OVA) and house dust mite (HDM) models, wherein the OVA-induced AD model demonstrates relative independence from the mouse strain, age, and sex, exhibiting a notable immunological shift from Th2 to a mixed Th1 + Th2 response and involving IgE-independent mast cell activation pathways. The HDM model induces epidermal and dermal thickening, increased infiltration of mast cells and eosinophils, and downregulation of skin barrier proteins (filaggrin and loricrin), with inflammatory responses being significantly amplified when combined with barrier disruption protocols, and NC/Nga mice are frequently used for this modeling approach (Sanjel and Shim, 2022).

The third category consists of spontaneous models, represented by the NC/Nga mouse strain, which develops pruritic skin lesions spontaneously at approximately 8 weeks of age under conventional housing conditions, exhibiting impaired water-retention properties and barrier dysfunction, reduced ceramide content, intrinsic susceptibility to AD-like dermatitis, and high expression of Th2-specific chemokines, including thymus and activation-regulated chemokine (TARC) and macrophage-derived chemokine (MDC), along with their receptor CCR4 in lesional skin (Zheng et al., 2024).

The use of DNCB-induced models utilizing non-BALB/c mouse strains in the current analysis may have contributed to the observed heterogeneity. A critical interspecies discrepancy requires attention: while human AD patients demonstrate FLG gene downregulation in lesional skin, FLG expression is paradoxically upregulated in NC/Nga and OVA models (Ewald et al., 2017), which may explain the discordance between the clinical study (Cho et al., 2008) showing significant FLG upregulation following Zicao treatment and the non-significant results observed in this meta-analysis. Since Zicao may primarily exert its therapeutic effects via Th2 modulation and its associated mechanisms, increased efficacy is expected in Th2-dominant models.

This systematic review has inherent limitations that necessitate cautious interpretation of the findings. First, critical methodological deficiencies were identified via SYRCLE assessment—inadequate randomization reporting (only 3/12 studies described sequence generation), universal lack of allocation concealment and blinding, and unaccounted animal exclusions post-randomization in one study—collectively inflating efficacy estimates. Second, substantial unexplained heterogeneity persisted in primary outcomes, which may be attributable to variable disease induction methods, inconsistent outcome measurement methods, and divergent interventions (oral vs. topical administration; whole-herb vs. isolated compounds). Third, pharmacological standardization was severely compromised by the universal absence of phytochemical quantification (e.g., shikonin/acetylshikonin concentrations) and inadequate botanical documentation (voucher specimens and extraction protocols), precluding dose–response interpretation. Finally, critical translational gaps exist: none of the studies reported pharmacokinetics, toxicology profiles, adverse events, or long-term efficacy data. Furthermore, to maintain methodological rigor, our quantitative synthesis was restricted to outcome measures that were consistently reported in at least two independent studies. Consequently, several potentially important parameters, such as specific cell-signaling pathway components and microbiome composition, could not be evaluated due to insufficient reporting.