This systematic review and meta-analysis (SRMA) adhered strictly to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines to ensure transparency and reproducibility. ¹³ The research questions were formulated using the PECO framework (Participants: pregnant women; Exposure: PM2.5; Comparison: varying levels of PM2.5 exposure; Outcome: maternal depression), following established methodologies for environmental health reviews. ¹⁴ Does exposure to PM2.5 influence the risk of depression in pregnant women?

A comprehensive literature search was conducted to identify studies published between January 2015 and January 2025, focusing on recent evidence to reflect current PM2.5 exposure levels and diagnostic standards for depression. Earlier studies were excluded due to potential inconsistencies in exposure assessment methods and lower relevance to contemporary air quality standards. Three databases were searched: PubMed, Open Science, Google Scholar, and a manual search of reference lists. The search initially used broad terms such as “air pollution” and “particulate matter” to capture relevant studies, followed by specific terms including “PM2.5,” “fine particulate matter,” “maternal depression,” “postpartum depression,” “pregnant women,” and “prenatal exposure.” These were combined using Boolean operators (AND, OR, NOT) to enhance precision. In PubMed, Medical Subject Headings (MeSH) terms (e.g., “Air Pollution”[MeSH Terms] AND “Depression”[MeSH Terms]) were incorporated alongside keywords, with equivalent terms adapted for Embase and Google Scholar to ensure consistency across databases ( Table 1). Manual searches of reference lists from included studies supplemented the database search to ensure a thorough collection of relevant literature.

Studies were included if they: (1) were original articles published between 2015 and 2025; (2) assessed outdoor PM2.5 exposure using ambient air quality data or modeled estimates (personal exposure monitors were not used in included studies); (3) reported maternal depression outcomes (diagnosed via clinical criteria, e.g., DSM-5, or validated scales, e.g., Edinburgh Postnatal Depression Scale [EPDS] ≥ 13); (4) focused on pregnant women; (5) were written in English; and (6) had five or more citations to ensure a minimum level of peer recognition. Exclusion criteria included: (1) reviews or studies on indoor air pollution; (2) studies not specifically addressing PM2.5 (e.g., focusing on PM10 or NO2); (3) studies reporting non-depressive outcomes (e.g., anxiety); (4) studies published before 2015; (5) studies with fewer than five citations; and (6) non-English studies. Social media-related studies were not considered, as they were irrelevant to the research question. The selection criteria are tabulated on Table 2 below.

The search identified 208 studies across databases (Google Scholar: 28; Science Open: 106; PubMed: 58) and 39 additional studies via manual search, totalling 231 studies. After removing 49 duplicates, 182 studies remained. Screening titles and abstracts excluded 110 studies, leaving 72 for full-text review. Of these, 48 were excluded based on predefined inclusion/exclusion criteria (e.g., irrelevant outcomes, non-pregnant populations, or non-depression focus). During data extraction, 14 additional studies were excluded due to incompatible exposure metrics (e.g., PM2.5 not specifically measured) or insufficient data for effect size conversion (e.g., missing sample sizes or variance estimates). This left 10 eligible studies, of which 7 were ultimately included in the meta-analysis. Three studies were excluded at the final stage: 2 lacked PM2.5-specific data, and 1 reported odds ratios (ORs) without sufficient data (e.g., sample sizes, confidence intervals) to convert to risk ratios (RRs) using standard methods. ¹⁵ A PRISMA flow diagram detailing the selection process is provided ( Figure 1).

Figure 1: Summarizes each of the selected articles that are included in this review.

For each included study, the following were extracted using a predefined form: author’s first name, year of publication, location/country, participant details, research design, sample size, perinatal exposure window, outcome definition (e.g., clinical diagnosis, EPDS score), mean/median exposure duration, adjustment variables, and effect estimates (RRs) with 95% confidence intervals (CIs). Quality was assessed using the Effective Public Health Practice Project (EPHPP) tool, which evaluates study design, selection bias, confounders, and data collection reliability.

Exposure data were standardized to a 10 μg/m ³ PM2.5 increment by converting reported effect sizes (e.g., using linear scaling for studies reporting per 5 μg/m ³). Outcomes were harmonized by including only studies reporting clinical depression diagnoses or validated scale cutoffs, ensuring consistency. All effect sizes were standardized to RRs; studies reporting ORs were converted to RRs where possible using the method of. ¹⁵ When studies reported both unadjusted and adjusted RRs, adjusted estimates were used. For studies with multiple adjusted models, the model with the largest number of covariates was selected to minimize confounding bias.

A random-effects model (DerSimonian-Laird) was used to pool RRs, accounting for between-study heterogeneity, which was assessed using Cochran’s Q statistic, I ², and Tau ². I ² values were reported to indicate the proportion of variance due to heterogeneity, with I ² ≥ 50% considered substantial. Subgroup analyses by exposure window (first, second, third trimesters, whole pregnancy) and region (e.g., China, USA) explored heterogeneity sources. Leave-one-out sensitivity analysis evaluated the robustness of the pooled estimate. Publication bias was assessed using funnel plots, Egger’s test, and Begg’s test, with significant bias (Egger’s p < 0.05) noted but not adjusted due to the small sample size (n = 7). Quality-effects meta-regression was performed using weighted least squares (WLS) with Newcastle-Ottawa Scale (NOS) scores and sample size as covariates, despite the small sample size, to assess quality effects, acknowledging limitations (e.g., invalid Omnibus test). Analyses were conducted in Python using statsmodels and visualized with plotly.

The Cochrane Risk of Bias tool was initially considered but deemed less applicable for observational studies. Instead, the NOS was used to assess study quality, focusing on selection, comparability, and outcome assessment, ensuring a robust evaluation of methodological rigor.

The primary outcome is maternal depression in pregnant women associated with PM2.5 exposure. Depression is defined as a clinical diagnosis (e.g., DSM-5, ICD-10) or a score above a validated threshold on a standardized scale (e.g., EPDS ≥ 13, PHQ-9 ≥ 10). Outcomes are measured at any point during pregnancy or up to 12 months postpartum. The effect measure for synthesis is the risk ratio (RR) with 95% CIs, pooled using a random-effects model (DerSimonian-Laird). Odds ratios (ORs) are converted to RRs using the method of. ¹⁵ No additional outcomes were included, as the study focuses exclusively on depression, with non-depressive outcomes (e.g., anxiety) excluded per the selection criteria.

Subgroup analysis stratified by exposure window, first trimester exposures showed a 23% increased risk (RR = 1.23, 95% CI: 1.10-1.36;, I ² = 0.0%, 2 studies), second trimester RR = 9.24 (95% CI: 1.04–82.10, I ² = 0.0%, n = 1), third trimester RR = 1.04 (95% CI: 1.00–1.08, I ² = 0.0%, n = 1), and whole pregnancy RR = 1.41 (95% CI: 1.23–1.63, I ² = 29.57%, n = 3). Significant between-group heterogeneity was observed (Q_between = 37.46, df = 3, p < 0.001), with the second trimester showing an outlier effect, though with a wide confidence interval. Geographically, Chinese studies (RR = 1.23, 95% CI: 1.10-1.36) showed more consistent effects than US studies (RR = 1.36, 95% CI: 1.07-1.72; p = 0.03 for difference) ( Figure 3).

The leave-one-out sensitivity analysis ( Figure 4) assessed the impact of omitting each study. Pooled RRs ranged from 1.25 (95% CI: 1.08–1.45, I ² = 85.8%, when ¹⁷ excluded) to 1.33 (95% CI: 1.19–1.49, I ² = 44.3%, when ¹⁰ excluded), with percent changes from -2.81% to 3.56% relative to the original RR of 1.29 ( Figure 4). The stability of the pooled estimate across omissions (all CIs overlapping the original 1.10–1.50) supports the robustness of the primary finding. All sensitivity estimates remained statistically significant.

The funnel plot ( Figure 5) assessed publication bias, with Egger’s test showing an intercept of 1.914 (p = 0.013) and Begg’s test a correlation of 0.893 (p = 0.007). The plot exhibited asymmetry, with studies scattered unevenly around the pooled effect (log RR ≈ 0), particularly at lower precision (1/SE), indicating significant publication bias that may inflate the pooled RR.

The scatter plot of study quality vs. effect size ( Figure 6) explored the influence of NOS scores. Quality-effects meta-regression found no significant association between quality score and log RR (coefficient = -0.0505, p = 0.530) or sample size and log RR (coefficient = -2.483e-07, p = 0.732), with an R-squared of 0.694 (F-statistic p = 0.0937). The small sample size (n = 7) precluded valid Omnibus normality testing, and a high condition number (4.47e+06) suggested potential multicollinearity. Studies included from USA are ^{10, 12, 17, 18} and studies from China includes. ^{16, 19, 20}