We conducted a systematic review and meta-analysis in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. The review protocol was registered on PROSPERO (CRD42024583729). The analysis was limited to previously published studies; no new clinical data were collected for this work.

A comprehensive literature search was performed in October 2024 using PubMed, Embase, and the Cochrane Library. The full search string for PubMed was: (myopia[MeSH Terms] OR myopia[Title/Abstract]) AND (atropine[MeSH Terms] OR atropine[Title/Abstract]) AND (child*[Title/Abstract] OR pediatric[Title/Abstract]) AND (randomized controlled trial[Publication Type] OR placebo[Title/Abstract]). Similar terms were adapted for Embase and Cochrane, including MeSH/Emtree. Complete, database-specific search strategies (including all field tags and limits) for PubMed, Embase, and Cochrane CENTRAL are provided in Supplementary Appendix S1. We also searched trial registries (ClinicalTrials.gov, WHO ICTRP) for unpublished trials, but none met inclusion. No language restrictions were applied; non-English studies (e.g., Chinese RCTs) were screened using translation tools if abstracts were available, but all included trials were in English. After removing duplicates, titles and abstracts were screened to exclude clearly irrelevant reports. The full text of remaining articles was then evaluated against the predefined inclusion criteria. We also hand-searched reference lists of relevant papers to identify any additional eligible studies.

We included studies that: (1) were randomized, double-blind, placebo-controlled trials of topical atropine eye drops; (2) enrolled children roughly 4–15 years old with bilateral myopia (spherical equivalent refraction −0.50 D to −8.00 D in each eye, confirmed by cycloplegic refraction-studies with minor variations in baseline SER were included if groups were balanced at randomization); (3) tested a low concentration of atropine (defined here as 0.01% in at least one study arm); (4) had a follow-up duration of at least 1 year; and (5) reported on at least one of the primary efficacy outcomes (change in refractive error and/or change in axial length). Baseline axial length thresholds were not predefined, as they vary by age/ethnicity, but included studies had typical ranges (∼23–25 mm) with balanced groups. We required that trials also provide data on safety or side effects (particularly photophobia, given the known mydriatic effect of atropine). If a trial included multiple atropine concentrations, only data from the 0.01% atropine and placebo arms were considered for our primary analysis, in order to focus on the ultra-low dose.

We excluded trials that combined atropine with other myopia control interventions (e.g., orthokeratology or multifocal lenses), studies in populations with coexisting ocular pathologies (e.g., strabismus, amblyopia) or prior myopia therapy, and studies with significant baseline differences between groups in age or refractive status (defined as p < 0.05 for between-group t-tests or equivalent on key variables like age, SER, AL). Non-randomized studies, open-label studies without a placebo control, and trials with follow-up shorter than 12 months were also excluded.

Two reviewers (L.M. and M.L.) independently screened all titles/abstracts and full-text articles for eligibility. Discrepancies were resolved by discussion with a third author (J.E.G.). In cases of multiple publications from the same trial, or an interim and final report, we included the most comprehensive or final dataset (to avoid double-counting patients).

Two reviewers independently extracted data using a standardized form; disagreements were resolved by consensus or, when needed, a third reviewer (J.E.G.). Extracted items included study design and setting, sample size, participant characteristics (age, ethnicity), atropine concentration and dosing frequency, follow-up duration, and outcomes.

The unit of analysis was the child (participant). When outcomes were reported per eye, we extracted a single participant-level observation to avoid unit-of-analysis error due to within-child (inter-eye) correlation. We used the trial’s prespecified eye when stated (commonly the right eye). If a trial reported only a participant-level summary (e.g., mean of both eyes), we extracted that summary as reported. If outcomes were reported separately for right and left eyes without a prespecified eye or participant-level summary, we extracted the right-eye outcome to maintain a single observation per participant. We did not enter outcomes from both eyes of the same child as independent observations and therefore did not apply a paired-eye correlation correction or effective sample-size adjustment. If a trial analyzed both eyes using a method that accounts for within-child correlation (e.g., mixed models or generalized estimating equations) and reported a participant-level estimate, we extracted the reported estimate.

Primary outcomes were the change from baseline in spherical equivalent refraction (SER; diopters) and axial length (mm), analyzed as 12-months (annualized) change (D/year and mm/year). When a 12-months endpoint was available, we extracted the 12-months change; otherwise, we annualized the reported change over follow-up duration T years, assuming approximately linear change over time, by μ_yr=μ_T/T. Dispersion was scaled accordingly (SD_yr = SD_T/T; SE_yr = SE_T/T; equivalently, variances were scaled by 1 / T²). When multiple follow-up timepoints were reported, we used the timepoint closest to 12 months for the primary analysis. Secondary outcomes included adverse events, particularly photophobia. For studies reporting confidence intervals but not standard deviations, we derived SE from the 95% CI as SE = (U – L)/3.92 and applied the same annualization scaling when applicable. Reported effect estimates and 95% CIs were recorded when available; otherwise, they were calculated from extracted raw data. Tables 1, 2 summarize study characteristics and extracted quantitative findings.

The risk of bias in each trial was assessed using the Cochrane Risk of Bias 2 (RoB 2) tool for randomized trials. This evaluates potential bias across five domains: (1) randomization process, (2) deviations from intended interventions, (3) missing outcome data, (4) outcome measurement, and (5) selective reporting. Two reviewers performed the bias assessments independently, with any disagreements resolved by consensus. Each trial was rated as “low risk,” “some concerns,” or “high risk” of bias on each domain and overall. Table 3 provides the full domain-level assessments for each study. In our cohort, all trials were blinded and placebo-controlled, minimizing performance and detection biases. No trial was judged to have a high overall risk of bias. Most were rated low risk across domains; however, five trials had some concerns in at least one domain [Hieda et al. (15): missing outcome data; Saxena et al. (16): randomization, missing outcome data, selection of the reported result; Sharma et al. (17): randomization, deviations, missing outcome data, selection of the reported result; Chan et al. (18) and Hansen et al. (19): selection of the reported result]. These methodological quality findings are noted when interpreting the results. A more complete explanation of each RoB2 domain for each of the studies can be found in Supplementary Appendix S1.

All meta-analyses were performed using Review Manager web version 9.17.0. Random-effects models were used throughout with inverse-variance weighting. For continuous outcomes (spherical equivalent refraction [D/year] and axial length [mm/year]), effects were pooled as mean differences (atropine − placebo) with 95% confidence intervals calculated using the Hartung–Knapp–Sidik–Jonkman (HKSJ) method, and between-study variance (τ²) estimated using restricted maximum likelihood (REML). For dichotomous outcomes (photophobia), effects were pooled as risk ratios with 95% confidence intervals calculated using HKSJ, and τ²; estimated using the DerSimonian–Laird method, consistent with the analysis settings used for that outcome. Statistical heterogeneity was assessed using Cochran’s Q and quantified with I². Prediction intervals were calculated for random-effects models to reflect the expected range of effects in a future similar study. A two-sided P < 0.05 was considered statistically significant.

The literature search yielded 312 unique records. After title/abstract screening, 27 full-text articles were reviewed. Of these, 18 were excluded (Figure 1 provides details of exclusions, such as ineligible study design or insufficient follow-up), leaving 9 randomized controlled trials (15–23) for inclusion in this review (Figure 1). These 9 trials were published between 2017 and 2024 and comprised a total of 1091 pediatric participants (576 in atropine 0.01% groups and 515 in placebo groups). Table 1 summarizes the key characteristics of the included studies. All trials were double-blind and placebo-controlled, and all but one were single-center studies (the exception being a multicenter trial in India). Study sample sizes ranged from 60 to 400 children, with follow-up durations of 1 year in most cases (one trial followed participants for 18 months, which we annualized for comparison). Geographically, the studies were conducted in East Asia (Hong Kong, mainland China, Japan), South Asia (India), Europe (Denmark), and the United States.

The mean age of participants at enrollment ranged from 8 to 10 years across trials, with most studies having slightly more female than male participants. Baseline refractive error and axial length were generally similar between the atropine and placebo groups within each study (mean spherical equivalent around −2 to −3 D, and mean axial length roughly 24 mm, with some variation by population). Ethnic composition varied: the U.S. trial included a multi-ethnic cohort (46% White, 18% East Asian, 15% Black, 15% Latino, 6% other), the Danish trial involved European Caucasian children, and the Asian trials predominantly involved East or South Asian children. These differences are noted because ethnicity can influence both myopia progression rates and treatment response.

Change in spherical-equivalent refraction (SER): Effects are expressed as mean difference (MD, atropine − placebo) in diopters per 12 months, where positive values favor atropine (less myopic shift). The pooled MD was +0.14 D/year (95% CI +0.04 to +0.24; p = 0.01; I² = 64%) using REML with Hartung–Knapp adjustment (Figure 2). Seven of nine studies showed point estimates favoring atropine; one was essentially null and one favored placebo. The 95% prediction interval was −0.11 to +0.40 D/year.

Change in axial length (AL): Effects are expressed as MD (atropine − placebo) in millimeters per 12 months, where negative values favor atropine (less axial elongation). The pooled MD was −0.05 mm/year (95% CI −0.08 to −0.01; p = 0.01; I² = 30%) (Figure 3). Most trials showed small reductions in axial elongation with atropine; one was essentially null and one slightly favored placebo. The 95% prediction interval was −0.11 to +0.02 mm/year.

Five trials contributed dichotomous safety data for photophobia, with definitions varying (e.g., self-reported light sensitivity vs. clinical grading); this and potential for selective reporting may bias estimates. The pooled risk ratio (RR) for photophobia with 0.01% atropine vs. placebo was 1.17 (95% CI 0.43–3.20; p = 0.69; I² = 15%) with a 95% prediction interval 0.28–4.85 (Figure 4). Absolute event counts were 42/427 (9.8%) in atropine and 22/372 (5.9%) in placebo arms. Overall, 0.01% atropine did not increase photophobia risk compared with placebo, and reported events were generally mild. Other adverse effects (e.g., accommodation difficulty, near-vision symptoms, pupil changes) were assessed inconsistently across trials, limiting synthesis.

To explore heterogeneity, we conducted sub-group analyses by region/ethnicity: Asian cohorts (7 trials: Chan, Yam, Wei, Hieda, Saxena, Chia, Sharma; n = 852) vs. Western/multi-ethnic (2 trials: Hansen [European], Repka [multi-ethnic US]; n = 239). For SER, the pooled MD was +0.17 D/year (95% CI +0.04 to +0.29; p < 0.03; I² = 56%) in Asian groups vs. +0.07 D/year (95% CI −1.18 to +1.32; p = 0.06; I² = 72%) in Western/multi-ethnic. For AL, −0.05 mm/year (95% CI −0.09 to −0.01; p < 0.28; I² = 25%) in Asian vs. −0.03 mm/year (95% CI −0.46 to +0.40; p = 0.14; I² = 54%) in Western/multi-ethnic. Test for sub-group differences: p = 0.39 for SER, p = 0.52 for AL. These suggest stronger effects in Asian populations, potentially due to faster baseline progression rates. However, limited studies per sub-group and inability to analyze other factors (e.g., age, baseline SER) restrict generalizability.

The individual risk-of-bias assessment for the nine included randomized controlled trials is summarized in Table 3. The evaluation was performed using the revised Cochrane risk-of-bias tool for randomized trials (RoB 2.0). Five trials [Yam et al. (20), Wei et al. (21), Repka et al. (22), Hansen et al. (19), and Chia et al. (23)] were judged to have an overall low risk of bias across all domains. In contrast, four studies–Hieda et al. (15), Saxena et al. (16), Sharma et al. (17), and Chan et al. (18)–were rated as having “some concerns.” These concerns were mainly related to missing outcome data (Hieda, Saxena, Sharma), the randomization process (Saxena, Sharma), deviations from intended interventions (Sharma), and selective reporting of results (Saxena, Sharma, Chan). Importantly, no trial was assessed as having a high risk of bias in any domain.