Early‐Onset Versus Late‐Onset Gestational Diabetes Mellitus: A Systematic Review and Meta‐Analysis of Maternal, Perinatal and Neonatal Outcomes

ABSTRACT

Aim

To determine the precise relationship between the timing of gestational diabetes mellitus (GDM) diagnosis and adverse maternal, perinatal and neonatal outcomes.

Materials and Methods

PubMed, EMBASE, Cochrane Central, Web of Science and CINAHL from inception to March 5, 2026 were systematically searched for observational studies and post hoc analyses or subanalyses of randomised controlled trials comparing outcomes between early‐onset GDM (eGDM) and late‐onset GDM (LGDM). The tested outcomes included insulin initiation, caesarean delivery, preterm birth, large‐for‐gestational‐age (LGA) infants and neonatal intensive care unit (NICU) admission. Pooled odds ratios (ORs) were calculated using random‐effects models. Heterogeneity was assessed using Cochran's Q test and I ² statistic. Meta‐regression and subgroup analyses were used to explore the sources of heterogeneity and sensitivity analyses confirmed robustness. The protocol was registered in PROSPERO (CRD420261328602).

Results

Thirty‐five studies encompassing 105 123 pregnancies met inclusion criteria. eGDM is associated with significantly elevated odds of insulin initiation (OR, 95% CI: 2.26, 1.75–2.91), caesarean delivery (1.16, 1.01–1.33), preterm birth (1.32, 1.05–1.64), LGA (1.29, 1.06–1.58) and NICU admission (1.24, 1.03–1.50). Heterogeneity is substantial (I ² = 57.3%–86.5%). Certainty is rated by GRADE as moderate for preterm birth and low for the other four outcomes.

Conclusions

eGDM versus LGDM is associated with higher risks of adverse outcomes, particularly insulin initiation. Gestational age at diagnosis may serve as a key risk stratifier, highlighting the need for risk‐stratified approaches rather than universal early surveillance. Future research should prioritise individual‐participant data meta‐analyses and randomised trials of tailored protocols.

1. Introduction

Gestational diabetes mellitus (GDM) affects ~14% of pregnancies worldwide, with substantial geographical variation [1]. A growing proportion, 15%–70% of all GDM cases, is diagnosed before 24 weeks of gestation (early‐onset GDM, eGDM [2]). Unlike late‐onset GDM (LGDM, 24–28 weeks), eGDM is a distinct phenotype characterised by preexisting insulin resistance, β‐cell dysfunction and elevated risks of adverse perinatal outcomes [3]. Its pathophysiology involves chronic hyperglycemia, early placental maladaptation and foetal glucose steal; especially, excessive foetal glucose uptake worsens maternal hyperglycemia [3].

Relevant guidelines recommend universal screening of GDM at 24–28 weeks, but the existing opinions on managing earlier hyperglycemia lack consensus [4]. Moreover, the current observational data are inconsistent, as some studies find no benefit, but other studies report an increased risk of small‐for‐gestational age [5, 6]. To reflect this uncertainty, Huhn et al. noted that most expert panels withhold routine early screening owing to insufficient evidence from randomised controlled trials (RCTs) [7].

Although most observational studies associate eGDM with worse outcomes, several cohort studies paradoxically report lower risks of preterm birth, caesarean delivery or large‐for‐gestational age (LGA), potentially because of intensified monitoring and timely treatment. Some studies demonstrate lower odds of preterm birth (odds ratio [OR] 0.72 [8], 0.59 [9]) and neonatal intensive care unit (NICU) admission (OR 0.77) [8]. Other studies document lower rates of preterm birth (OR 0.41) [10] and caesarean (adjusted risk ratio [aRR] 0.80, 95% CI 0.68–0.94) [11]. Collectively, these findings suggest that early diagnosis may attenuate some adverse outcomes, but the evidence remains heterogeneous.

In contrast, a recent RCT meta‐analysis concludes that detection and treatment of eGDM offer no indisputable benefits [12] and the only benefit is a reduction in neonatal respiratory distress. Population‐based strategies increase GDM detection and reduce primary caesarean delivery, but are associated with higher rates of pregnancy‐induced hypertension and preeclampsia [12]. This discrepancy between observational evidence and randomised evidence highlights persistent uncertainty.

The TOBOGM trial reported the reduced neonatal respiratory distress after early treatment, but demonstrated no benefit in other outcomes [6]. A subsequent meta‐analysis confirms this benefit, but shows no consistent effects on other outcomes, with evident certainty rated low to very low [12]. The 2024 Flemish consensus proposes specific diagnostic thresholds for eGDM [4], but their generalizability remains uncertain and the comparative risk profile of eGDM versus LGDM requires systematic evaluation [13].

Therefore, we conducted this systematic review and metaanalysis to compare maternal (insulin use, caesarean delivery), perinatal (preterm birth) and neonatal (LGA, NICU admission) outcomes between eGDM and LGDM, to clarify whether eGDM confers a higher risk independent of treatment and to inform clinical decision‐making.

2. Materials and Methods

2.1. Search Strategy and Selection Criteria

This systematic review and meta‐analysis was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) 2020 statement [14]. The review protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO) (CRD420261328602).

Two independent reviewers systematically searched PubMed, EMBASE, Cochrane Central Register of Controlled Trials, Web of Science and CINAHL from database inception to March 5, 2026. Our strategy incorporated the use of Medical Subject Headings (MeSH) terminology in PubMed, EMBASE ETREE terms and headings in CINAHL and keyword phrases related to ‘gestational diabetes mellitus’, ‘early‐onset’, ‘late‐onset’ and ‘pregnancy outcomes’. The search was restricted to human studies published in English.

2.2. Inclusion and Exclusion Criteria

Studies were eligible for inclusion if they met the following criteria:

• P (Population): Pregnant women diagnosed with GDM;

• I (Intervention/Exposure): eGDM, defined as diagnosis before 24 weeks of gestation (or as defined by the original study);

• C (Comparator): LGDM, defined as diagnosis at or after 24 weeks of gestation (or as defined by the original study);

• O (Outcomes): at least one of the following outcomes: insulin use, caesarean delivery, preterm birth (< 37 weeks), LGA or NICU admission;

• S (Study Design): Observational studies (prospective or retrospective cohort, cross‐sectional) or post hoc analyses or subanalyses of RCTs (p‐h/sa of RCT).

Exclusion criteria included:

• Studies with inadequate or unclear data to calculate effect estimates;

• Case reports, case series, reviews, editorials or conference abstracts;

• Animal or laboratory‐based studies;

• Studies including women with pre‐existing type 1 or type 2 diabetes;

• Studies with multiple pregnancies (twins, triplets) unless data for singletons could be extracted separately.

When multiple publications from the same cohort reported overlapping outcomes, the study with the largest sample size was retained.

2.3. Data Retrieval and Risk of Bias Assessment

Two reviewers separately evaluated the titles and abstracts of all potentially relevant electronic search studies. The full texts of potentially eligible studies were assessed against the predefined eligibility criteria. Disagreements between them were resolved through discussion or, if necessary, consultation with a third reviewer. Inter‐rater agreement was assessed using Cohen's kappa coefficient (κ): 0.00–0.20, slight; 0.21–0.40, fair; 0.41–0.60, moderate; 0.61–0.80, substantial; 0.81–1.00, almost perfect [15].

A standardised data extraction form was used to collect the following information from each included study: first author, year of publication, country, study design, sample size (early‐onset and late‐onset groups), diagnostic criteria for GDM, screening type (universal vs. selective), maternal age, body mass index (BMI), parity, outcomes assessed, effect measures (OR, HR, RR) with 95% CIs and adjustment variables. For studies reporting only crude event counts, we calculated unadjusted ORs using the raw data. The definitions of eGDM and LGDM both varied across the included studies. We systematically extracted the specific diagnostic criteria used in each study, which were primarily based on glycemic thresholds (e.g., International Association of Diabetes and Pregnancy Study Groups (IADPSG)/World Health Organization (WHO) 2013, Carpenter‐Coustan, Australasian Diabetes in Pregnancy Society (ADIPS)). These criteria are comprehensively summarised in Table S1.

The risk of bias for each included study was assessed using the Risk Of Bias In Non‐randomized Studies of Interventions (ROBINS‐I), as recommended by the Cochrane Handbook for Systematic Reviews of Interventions. For the p‐h/sa of RCTs, ROBINS‐I was also applied because the comparison of interest (eGDM vs. LGDM) involved non‐randomised groups based on the timing of diagnosis, which introduces potential confounding and selection bias analogous to observational studies. ROBINS‐I evaluates seven key domains: confounding, selection of participants, classification of exposures, deviations from intended interventions, missing data, measurement of outcomes and selective reporting [16]. Overall risk of bias was categorised as low, moderate, serious or critical. Two reviewers independently assessed quality and disagreements between them were resolved by consensus or third‐party adjudication. Inter‐rater agreement was quantified using Cohen's kappa (κ).

2.4. Statistical Analysis

The primary effect measure was the OR for each outcome comparing eGDM with LGDM. When a study reported HRs or RRs, these data were treated as equivalent to ORs for pooling, given the low event rates of most outcomes (except caesarean delivery and insulin use) [17].

For the primary analysis, we defined eGDM as diagnosis before 24 weeks of gestation. Studies with overlapping participant cohorts from the same institution were handled by excluding the smaller or less comprehensive report. Two studies [18, 19] (same institution as Refs [11, 20]. respectively) were excluded from the primary analysis to avoid double‐counting. Moreover, adjusted effect estimates (adjusted OR [aOR], aRR, adjusted prevalence ratio [aPR]) were prioritised when available. When adjusted estimates were not reported for the outcome of interest (e.g., insulin use in [11]), crude ORs calculated from raw event counts were used.

Pooled effect estimates with 95% CIs were calculated using a random‐effects model (DerSimonian‐Laird method) to account for anticipated between‐study heterogeneity [21]. Statistical heterogeneity was assessed using the I ² statistic and I ² of 25%, 50% and 75% indicates low, moderate and high heterogeneity, respectively [22]. When I ² exceeded 50%, the sources of heterogeneity were explored through meta‐regression and subgroup analyses.

Meta‐regression was conducted to examine effect modifiers when ≥ 10 studies were available. Subgroup analyses were prespecified to investigate potential sources of heterogeneity based on adjustment status (aOR vs. crude OR), diagnostic criteria (IADPSG/WHO 2013 criteria vs. other criteria), screening type (universal vs. selective vs. mixed) or geographic region (high‐income Western countries [Europe, North America, Oceania] vs. Asia vs. Middle East).

Robustness was assessed via a comprehensive series of sensitivity analyses. First, we evaluated the impact of varying eGDM definitions by gestational age at diagnosis: (i) < 22 weeks [23]; (ii) < 20 weeks [13, 19, 24, 25, 26]; (iii) very narrow definitions (< 16 weeks [27], < 13.1 weeks [28] and 6–14 weeks [29]). Second, to mitigate bias from cohort overlap, we substituted Ref. [20] with Ref. [18] and Ref. [11] with Ref. [19] (same underlying registries). Third, we restricted analyses to (a) adjusted effect estimates (≥ 3 key confounders); (b) IADPSG/WHO 2013 criteria only; (c) studies with ≥ 100 participants (excluding Ref. [30]). Fourth, we sequentially excluded the large cross‐sectional study [23] and the post hoc RCT analysis [28, 31], both individually and collectively. Finally, leave‐one‐out analyses were performed to assess the influence of individual studies.

For analyses including five or more studies, publication bias was assessed visually via funnel plots. Egger's linear regression test was used to quantify funnel plot asymmetry. If publication bias was detected (Egger's test p < 0.10 or visual asymmetry), the Duval‐Tweedie trim‐and‐fill method was applied to impute missing studies. The adjusted effect estimate was not materially different from the original analysis, confirming the overall conclusion is robust to publication bias [32].

The Grading of Recommendations Assessment, Development, and Evaluation (GRADE) framework was used to rate the overall certainty of evidence for each outcome and risk of bias, inconsistency, indirectness, imprecision and publication bias were considered [33]. The certainty of evidence was categorised as high, moderate, low or very low.

All statistical analyses were performed using Stata 14.0 (StataCorp LLC, College Station, TX, USA). Statistical significance was defined as two‐sided p < 0.05 for all analyses, except for tests of heterogeneity where p < 0.10 was considered significant.

3. Results

3.1. Study Identification

Of the 6113 identified publications, 92 articles were selected for full‐text review (Figure 1). Several studies were excluded from the primary analysis for predefined reasons, such as reporting only composite outcomes [34], defining LGDM as > 29 weeks (≥ 24 weeks required) [35] or lacking a late‐onset comparator [36]. Overlapping cohorts from the same institution were resolved by retaining the larger or more recent study (e.g., Most et al. [20] retained over Gupta [18]; Sagili et al. [11] over Nekkanti [19]). Additional overlapping studies came from Emek Medical Center, Israel (Yefet et al. [37]; Abu Shqara et al. [35]) and JIPMER, India (Nekkanti et al. [19]; Sagili et al. [11]). For Emek, the study from Abu Shqara et al. [35] was excluded (late‐onset definition mismatch); only the study from Yefet et al. [37] was retained for all outcomes. Both JIPMER studies were retained as they used distinct early‐onset definitions (< 20 weeks vs. < 24 weeks), enabling subgroup analyses. The study from Sagili et al. [11] (larger sample) was prioritised for primary analysis and the study from Nekkanti et al. [19] was prioritised for sensitivity analyses.

FIGURE 1.

Flow‐diagram of study selection.

Ultimately, 35 studies were included in the systematic review: one cross‐sectional study [23], two p‐h/sa analyses of RCTs [28, 31] and 31 cohort studies [8, 9, 10, 11, 13, 19, 20, 24, 25, 26, 27, 29, 30, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54]. Of them, 26 studies met the prespecified criteria for the primary metaanalysis (eGDM defined as < 24 weeks and LGDM as ≥ 24 weeks). Nine studies employing alternative definitions (< 22 weeks [23], < 20 weeks [13, 19, 24, 26], 16–18 weeks [25], < 16 weeks [27], < 13.1 weeks [28] or 6–14 weeks [29]) were included only in sensitivity analyses to test the robustness of the findings across different diagnostic thresholds. An additional study [18] from the same institution as Most et al. [20] was also used in sensitivity analyses as a replacement to assess robustness to temporal cohort choice (in Section 2). Interrater agreement for study inclusion was substantial (Cohen's κ = 0.76). Diagnostic criteria also differ: IADPSG/W‐HO 2013 criteria in 15 studies, Carpenter‐Coustan criteria in six studies, ADIPS criteria in two studies, NDDG criteria in two studies and other criteria (O'Sullivan, NZ OGTT, WHO 1999 or JSOG) (Table S1).

As for study settings, 34 studies were conducted at single centres and one study was a multicenter investigation [31]. The geographical origins of the studies were categorised by established socioeconomic development and metabolic disease burden profiles into three regions: high‐income Western countries (n = 18), Asia (n = 11) and the Middle East (n = 6). Screening strategies differed across studies: universal screening (n = 17), selective screening targeting individuals with predefined clinical risk factors (e.g., maternal obesity, prior gestational diabetes or advanced maternal age) (n = 17) and a hybrid strategy integrating both approaches (n = 1). Sample sizes ranged from 64 to 75 546 participants, with a cumulative total of 105 123 pregnant individuals across all 35 included studies. All studies reported at least one of the five prespecified core outcomes: insulin use, caesarean delivery, preterm birth, LGA or NICU admission. Fourteen studies reported adjusted effect estimates (aORs, aRRs or aPR) for at least one outcome; the remaining 21 studies reported crude ORs derived from unadjusted event counts.

3.2. Risk of Bias Assessment

According to the ROBINS‐I tool, the risk of bias varied across the 35 included studies, with outcome‐specific assessments summarised in Table S2. Confounding was the most frequent source of serious risk of bias, which was primarily due to inadequate adjustment for key maternal confounders, including maternal age, pre‐pregnancy BMI, parity and prior GDM history. Inter‐rater agreement was substantial (Cohen's κ = 0.78).

3.3. Maternal Outcomes

3.3.1. Insulin Use in eGDM vs. LGDM

A random‐effects meta‐analysis of 16 studies demonstrated that eGDM was associated with significantly higher odds of insulin therapy (pooled OR 2.26, 95% CI 1.75–2.91; p < 0.001), with substantial heterogeneity (I ² = 86.5%, p < 0.001, Figure 2).

FIGURE 2.

Forest plot comparing the odds of insulin use between women with eGDM and those with LGDM.

Univariate meta‐regression showed that diagnostic criteria (p = 0.934), screening type (p = 0.780) and effect type (p = 0.405) were not significantly associated with the effect size (adjusted R ² ≤ 12.4%, Table S3). Subgroup analyses by adjustment status, diagnostic criteria, region or screening type showed no significant differences. Subgroup findings were summarised as follows: is presented in Table 1 and detailed forest plots are available in Figures [Link], [Link], [Link], [Link].

TABLE 1.

Subgroup analyses for primary outcomes.

Outcome	Overall OR (95% CI)	I 2	Adjustment status (aOR vs. crude OR)	Diagnostic criteria (IADPSG vs. Other)	Region (high‐income vs. Asia vs. Middle East)	Screening type (universal vs. selective vs. mixed)
Insulin use	2.26 (1.75–2.91)	86.5%	aOR: 1.97 (1.04–3.72)	IADPSG: 2.02 (1.33–3.05)	High‐income: 2.34 (1.80–3.04)	Universal: 2.61 (1.56–4.38)
			cOR: 2.35 (1.82–3.03)	Other: 2.47 (1.86–3.27)	Asia: 2.21 (1.01–4.82)	Selective: 2.17 (1.73–2.71)
					Middle East: 2.39 (1.77–3.25)	Mixed: —
Caesarean delivery	1.16 (1.01–1.33)	74.9%	aOR: 1.05 (0.74–1.49)	IADPSG: 1.11 (0.89–1.39)	High‐income: 1.19 (0.97–1.45)	Universal: 1.14 (0.93–1.39)
			cOR: 1.20 (1.04–1.37)	Other: 1.20 (1.01–1.43)	Asia: 1.02 (0.83–1.25)	Selective: 1.21 (0.99–1.48)
					Middle East: 1.27 (1.06–1.53)	Mixed: 1.11 (0.88–1.41)
Preterm birth	1.32 (1.05–1.64)	57.3%	aOR: 1.70 (1.00–2.90)	IADPSG: 1.30 (0.95–1.49)	High‐income: 1.18 (0.83–1.68)	Universal: 1.33 (1.03–1.71)
			cOR: 1.22 (0.94–1.58)	Other: 1.34 (1.01–1.77)	Asia: 1.41 (0.84–2.37)	Selective: 1.28 (0.92–1.77)
					Middle East: 1.50 (1.01–2.25)	Mixed: —
LGA	1.29 (1.06–1.58)	75.3%	aOR: 1.35 (0.80–2.26)	IADPSG: 1.51 (1.08–2.10)	High‐income: 1.54 (1.24–1.91)	Universal: 1.14 (0.87–1.50)
			cOR: 1.27 (1.03–1.57)	Other: 1.21 (0.94–1.55)	Asia: 0.92 (0.73–1.16)	Selective: 1.50 (1.14–1.96)
					Middle East: 1.19 (0.92–1.54)	Mixed: 1.38 (1.05–1.82)
NICU admission	1.24 (1.03–1.50)	65.2%	aOR: 1.48 (0.89–2.45)	IADPSG: 1.17 (0.84–1.64)	High‐income: 1.16 (0.87–1.55)	Universal: 1.05 (0.79–1.41)
			cOR: 1.19 (0.96–1.46)	Other: 1.31 (1.06–1.63)	Asia: 1.16 (0.86–1.56)	Selective: 1.38 (1.08–1.77)
					Middle East: 1.97 (1.35–2.88)	Mixed: —

Note: High‐income, high‐income Western countries (Europe, North America, Oceania); “—” indicates 95% CI not reported in the primary manuscript text; subgroup differences were not statistically significant for any outcome (all p for interaction > 0.05, as reported in the Section 3).

Abbreviations: aOR, adjusted odds ratio; cOR, crude odds ratio; IADPSG, International Association of Diabetes and Pregnancy Study Groups; LGA, large‐for‐gestational‐age; NICU, neonatal intensive care unit.

Sensitivity analyses across alternative eGDM definitions (< 22 weeks, < 20 weeks, very narrow, all combined) and restrictions to adjusted estimates or IADPSG criteria yielded consistent results (OR 1.05–1.23). Leave‐one‐out analysis identified no unduly influential study. The exclusion of the study from Sagili et al. [11] produced the largest change (OR 1.24) (Figure S5 and Table S4).

Substantial heterogeneity was observed (I ² = 86.5%; Q = 111.40, df = 21, p < 0.001). The funnel plot showed mild asymmetry (Figure S6), supported by Egger's test (p = 0.007). Trim‐and‐fill analysis imputed two missing studies and attenuated the OR from 2.26 (95% CI 1.75–2.91) to 2.05 (1.60–2.63), indicating modest upward bias without any change in significance.

The overall certainty of evidence was rated as low (GRADE) and was downgraded for serious inconsistency (I ² = 86.5%; p < 0.001) and publication bias (Egger's p = 0.007; trim‐and‐fill attenuated OR from 2.26 to 2.05). No downgrading was applied for risk of bias or imprecision.

3.3.2. Caesarean Delivery in eGDM Versus LGDM

A meta‐analysis of 19 studies showed that eGDM was associated with a 16% higher odds of caesarean delivery (pooled OR 1.16, 95% CI 1.01–1.33; p = 0.032), with moderate heterogeneity (I ² = 74.9%, p < 0.001, Figure 3).

FIGURE 3.

Forest plot comparing the odds of caesarean delivery between women with eGDM and those with LGDM.

Univariate meta‐regression showed that diagnostic criteria (p = 0.682), screening strategy (p = 0.938) and effect type (p = 0.582) were not significantly associated with effect estimates for caesarean delivery, explaining negligible between‐study variance (ranged from −15.5% to −2.8%, Table S3). Subgroup analyses were summarised in Table 1. No significant differences were observed by adjustment status, diagnostic criteria, region or screening type (all p for interaction > 0.05). Figures [Link], [Link], [Link], [Link] provided corresponding forest plots.

Sensitivity analyses across alternative eGDM definitions (< 22, < 20, very narrow, all definitions combined) yielded consistent ORs (range 1.11–1.15). Restriction to adjusted estimates (OR 1.05) or IADPSG criteria (OR 1.11) did not alter the conclusion. Leave‐one‐out analysis identified no unduly influential study; excluding Sagili et al. [11] produced the largest change (OR 1.20, 95% CI 1.06–1.37) (Figure S11 and Table S4).

No publication bias was detected (Begg's p = 0.834; Egger's p = 0.847, Figure S12). The overall certainty of evidence was rated as low (GRADE) and downgraded for serious inconsistency (I ² = 74.9%). Sensitivity analyses confirmed the robustness of the direction of effect.

3.4. Perinatal Outcome

3.4.1. Preterm Birth in eGDM Versus LGDM

A random‐effects meta‐analysis of 15 studies indicated that eGDM was associated with significantly higher odds of preterm birth (pooled OR 1.32, 95% CI 1.05–1.64; p = 0.015), with moderate heterogeneity (I ² = 57.3%, p = 0.003, Figure 4).

FIGURE 4.

Forest plot comparing the odds of preterm birth between women with eGDM and those with LGDM.

Univariate meta‐regression showed that adjustment status (p = 0.262), diagnostic criteria (p = 0.966) and screening type (p = 0.945) were not associated with the effect size (adjusted R ² ≤ −11.8%, Table S3). Subgroup analyses were presented in Table 1 and Figures [Link], [Link], [Link], [Link]; none of the prespecified factors (adjustment status, diagnostic criteria, region, screening type) significantly modified the effect estimate (all p > 0.05).

Sensitivity analyses confirmed robustness across alternative eGDM definitions (< 22 weeks: OR 1.27; < 20 weeks: OR 1.29; very narrow: OR 1.30; all definitions: OR 1.27) and when restricted to adjusted estimates (OR 1.70) or IADPSG criteria (OR 1.41). Leave‐one‐out analysis showed no single study dominated the pooled estimate (range 1.25–1.40). Removal of the study from Sagili et al. [11] gave the largest decrease (OR 1.28, 95% CI 1.03–1.60) (Figure S17 and Table S4). Visual inspection of the funnel plot suggested mild asymmetry (Figure S18). Begg's test (p = 0.166) revealed no asymmetry, whereas Egger's test (p = 0.014) suggested potential small‐study effects, but trim‐and‐fill imputed no missing study (OR unchanged at 1.32), indicating that the findings were robust to publication bias.

The overall certainty of evidence for preterm birth was rated as moderate (GRADE), downgraded for inconsistency (I ² = 57.2%). No serious concern regarding risk of bias, imprecision or publication bias was identified. Sensitivity analyses confirmed the robustness of the pooled estimate.

3.5. Neonatal Outcomes

3.5.1. LGA in eGDM Versus LGDM

eGDM was associated with a 29% higher odds of LGA infants (pooled OR 1.29, 95% CI 1.06–1.58; p = 0.013) based on 16 studies, with substantial heterogeneity (I ² = 75.3%, p < 0.001, Figure 5).

FIGURE 5.

Forest plot comparing the odds of LGA between women with eGDM and those with LGDM.

Univariate meta‐regression showed that adjustment status (p = 0.786), diagnostic criteria (p = 0.348) and screening strategy (p = 0.321) were not associated with the effect size for LGA, which explains negligible between‐study variance (adjusted R ² ranged from −11.1% to 6.7%; Table S3). Subgroup analyses (Table 1 and Figures [Link], [Link], [Link], [Link]) revealed no significant effect modification by adjustment status, diagnostic criteria, region or screening type (all p > 0.05).

Sensitivity analyses confirmed robustness across alternative definitions of eGDM (< 22 weeks: OR 1.27; < 20 weeks: OR 1.25; very narrow: OR 1.29; all definitions: OR 1.23) and when restricted to adjusted estimates (OR 1.35) or IADPSG criteria (OR 1.31). Leave‐one‐out analysis showed no disproportionate influence from any single study (range 1.22–1.35). Omission of the study from Mustafa et al. [38] produced the greatest reduction (OR 1.22, 95% CI 1.01–1.48) (Figure S23 and Table S4). No evidence of publication bias was detected (Begg's p = 0.753; Egger's p = 0.432; Figure S24).

The overall certainty of evidence for LGA was low (GRADE) and downgraded for serious inconsistency (I ² = 75.3%). No serious bias, imprecision or publication bias was identified. Sensitivity analyses confirmed robustness (OR 1.22–1.35).

3.5.2. NICU Admission in eGDM vs. LGDM

A meta‐analysis of 15 studies showed that eGDM was associated with a 24% higher odds of NICU admission (pooled OR 1.24, 95% CI 1.03–1.50; p = 0.026), with substantial heterogeneity (I ² = 65.2%, p < 0.001, Figure 6).

FIGURE 6.

Forest plot comparing the odds of NICU between women with eGDM and those with LGDM.

Univariate meta‐regression showed that adjustment status (p = 0.543), diagnostic criteria (p = 0.524) and screening type (p = 0.274) were not significantly associated with the effect size for NICU admission, which explains negligible between‐study variance (adjusted R ² ranged from −28.8% to 19.3%; Table S3). Subgroup analyses were detailed in Table 1 and Figures [Link], [Link], [Link], [Link]. Importantly, none of the examined covariates (adjustment status, diagnostic criteria, region, screening type) significantly explained the between‐study heterogeneity (all p > 0.05).

Sensitivity analyses confirmed robustness across alternative definitions of eGDM (< 22 weeks: OR 1.18; < 20 weeks: OR 1.27; very narrow: OR 1.25; all definitions: OR 1.23) and when restricted to adjusted estimates (OR 1.48) or IADPSG criteria (OR 1.25). Leave‐one‐out analysis showed no disproportionate influence from any single study (range 1.19–1.31). Removing the study from Bashir et al. [39] led to the greatest attenuation (OR 1.19, 95% CI 0.98–1.44) (Figure S29 and Table S4). No evidence of publication bias was detected (Begg's p = 0.843; Egger's p = 0.314; Figure S30).

The overall certainty of evidence for NICU admission was low (GRADE), downgraded for serious inconsistency (I ² = 65.2%). No serious bias, imprecision or publication bias was detected. Sensitivity analyses confirmed robustness (OR range 1.19–1.48).

4. Discussion

In this systematic review and meta‐analysis of 35 studies (105 123 pregnancies), eGDM was consistently associated with higher odds of insulin requirement (OR 2.26), preterm birth (OR 1.32), LGA (OR 1.29), NICU admission (OR 1.24) and caesarean delivery (OR 1.16). The largest effect was for insulin use, followed by preterm birth and LGA with comparable magnitude, while NICU and caesarean showed lower odds. This pattern reinforces gestational age at diagnosis as a key phenotypic discriminator. Our findings extend a prior meta‐analysis [55] by providing updated estimates from a larger cohort (105 123 vs. 15 270 pregnancies) and systematically exploring heterogeneity.

4.1. Metabolic Severity: Insulin Requirement and Foetal Overgrowth

eGDM was associated with a 2.26‐fold increased odds of maternal insulin requirement and a 29% higher odds of LGA neonates, two outcomes that most directly reflect the metabolic burden of prolonged in utero hyperglycemic exposure. Insulin requirement (OR 2.26; 95% CI 1.75–2.91) exceeds those observed in other GDM subgroups stratified by pre‐pregnancy BMI or gestational weight gain [56], confirming eGDM as a metabolically distinct subtype marked by severe baseline insulin resistance and limited β‐cell reserve. Preexisting insulin resistance, frequently linked to obesity, is further amplified by placental hormone‐induced insulin resistance during mid‐to‐late gestation, ultimately exhausts β‐cell functional capacity and necessitates insulin therapy [57, 58].

Prolonged maternal hyperglycemia drives foetal overgrowth via the Pedersen hypothesis: maternal hyperglycemia → foetal hyperinsulinemia → IGF signalling → increased adipogenesis and somatic growth [59]. The HAPO study confirms this cascade and shows dose‐dependent associations between maternal glucose and cord C‐peptide, birth weight or adiposity [60]. Earlier GDM onset extends exposure to dysregulated glycemia and amplifies foetal insulin secretion and growth velocity, thus increasing LGA risk. Moreover, the ‘fetal glucose steal’ creates a self‐sustaining cycle: established foetal hyperinsulinemia enhances foetal glucose disposal and steepens the maternal‐foetal glucose gradient, even when maternal glucose is normal [61]. This result explains why foetal overgrowth can persist in eGDM despite adequate glycemic control, reinforcing that intervention must precede entrenched foetal hyperinsulinemia.

The heterogeneity in LGA effect estimates across studies stemmed from differences in population characteristics and screening practices. Mustafa et al. reported a high OR (> 2.0), which is likely due to a cohort enriched for obesity and selective high‐risk‐based screening [38]. In contrast, studies using universal IADPSG criteria (e.g., [31, 40]) found estimates closer to the null, which reflects broader case detection, including milder, later‐onset or transient hyperglycemia. Our pooled LGA estimate (OR 1.29; 95% CI 1.06–1.58) is lower than the OR of 1.60 from a prior GDM meta‐analysis that did not distinguish onset timing [62], suggesting that the overall LGA risk attributed to GDM is largely driven by early‐onset cases. This result highlights the need for temporal stratification in both research and clinical risk assessment.

4.2. Delivery Complications: Caesarean Delivery and Preterm Birth

Women with eGDM had 16% higher odds of caesarean delivery (OR 1.16; 95% CI 1.01–1.33) and 32% higher odds of preterm birth (OR 1.32; 95% CI 1.05–1.64), which align with prior meta‐analyses of heterogeneous GDM populations [62, 63]. This is the first head‐to‐head comparison providing rigorous estimates for eGDM versus LGDM and is directly applicable to antenatal risk stratification and delivery planning. The elevated caesarean risk appears to be driven by two interrelated pathways: increased LGA incidence (predisposing to cephalopelvic disproportion) and placental maladaptation (e.g., impaired trophoblast invasion, oxidative stress) [64, 65]. Geographic variation (OR > 1.5 in Western studies [46, 47] vs. ~1.1–1.2 in Asian cohorts [36, 45]) likely reflects differences in the baseline caesarean rates, obesity prevalence and adjustment for confounders. Notably, 19 of 20 included studies reported ORs > 1.0, reinforcing the robustness of this association. Paradoxically, some studies reported lower risks (e.g., preterm birth OR 0.72 [8], 0.59 [9]; caesarean aRR 0.80 [11]), which is probably due to intensified surveillance and earlier intervention rather than biological advantage.

Preterm birth arises through spontaneous or iatrogenic pathways. Spontaneous preterm birth is accelerated by prolonged hyperglycemia‐driven placental senescence, sterile inflammation (e.g., elevated IL‐6, TNF‐α) and premature parturition signalling [66]. Iatrogenic preterm birth often follows early detection of complications (e.g., polyhydramnios, gestational hypertension, preeclampsia or abnormal foetal surveillance), prompting provider‐initiated delivery. Two studies reported ORs < 1.0 (0.98 [23], 0.94 [42]), both used universal IADPSG screening, had low maternal obesity prevalence and defined eGDM with strict gestational age cutoffs (< 22–24 weeks), suggesting that earlier diagnosis and optimised management may attenuate the risk. Our pooled estimates for caesarean delivery (OR 1.16; 95% CI 1.01–1.33) and preterm birth (OR 1.32; 95% CI 1.05–1.64) align with those in broader GDM populations [63, 67], confirming that these complications are elevated even with later‐onset diagnosis. Unlike metabolic severity outcomes (e.g., insulin requirement OR 2.26), which shows a large effect, the associations for delivery complications are more modest and less timing‐dependent, indicating they reflect shared GDM pathophysiology rather than a distinct early‐gestation effect.

4.3. Neonatal Morbidity: NICU Admission as a Composite Endpoint

The 24% higher odds of NICU admission associated with eGDM (OR 1.24; 95% CI 1.03–1.50) represents a clinically meaningful neonatal signal. To our knowledge, this is the first meta‐analysis to provide pooled estimates specifically comparing NICU admission risk between eGDM and LGDM and extend prior syntheses that treat GDM as a binary exposure without temporal stratification [62]. NICU admission is a clinically meaningful composite endpoint reflecting multiple neonatal morbidities, such as respiratory distress, hypoglycemia, hyperbilirubinemia and complications of prematurity or birth trauma—all of which are more common after eGDM. These outcomes share overlapping pathophysiological mechanisms: (1) prolonged maternal hyperglycemia drives foetal hyperinsulinemia, leading to postnatal hypoglycemia [68]; (2) delayed pulmonary surfactant maturation increases respiratory morbidity, worsened by longer hyperglycemic exposure [69]. Elevated risks of preterm birth and LGA‐related birth trauma further compound neonatal vulnerability [70].

Two mechanisms link eGDM to neonatal respiratory morbidity. First, foetal hyperinsulinemia, driven by prolonged maternal hyperglycemia, delays pulmonary surfactant synthesis and alveolar maturation [69]. Second, trophoblast‐derived exosomes may impair foetal lung development during the canalicular phase (16–26 weeks) [3]. Together, these mechanisms, compounded by elevated preterm birth rates in eGDM, substantially increase risks of respiratory morbidity and NICU admission.

4.4. Sources of Heterogeneity and the Observational‐ RCT Gap

Substantial heterogeneity (I ² = 57.3%–86.5%) stemmed from differences in populations, diagnostic criteria and clinical management. This degree of heterogeneity precludes the formulation of a universally applicable strategy for early screening or management of GDM. Several study‐level factors may be accountable, including differences in baseline maternal BMI, glycemic thresholds for insulin initiation, local protocols and healthcare systems. Higher obesity prevalence may increase effect sizes for insulin use and LGA [71], whereas stricter glycemic targets may attenuate outcome differences [72].

An RCT meta‐analysis [12] demonstrates no overall benefits of early GDM treatment, except for reduced respiratory distress. This observational‐RCT discrepancy suggests unmeasured confounding, but the consistent direction (OR > 1.0) across studies supports eGDM as a genuine high‐risk phenotype. Nevertheless, future individual‐participant data meta‐analyses are needed to formally assess effect modification by these clinical and system‐level variables.

4.5. Strengths and Limitations

This meta‐analysis has several key strengths. It is the most comprehensive contemporary comparison of eGDM and LGDM across five prespecified and clinically critical outcomes: maternal insulin requirement, caesarean delivery, preterm birth, LGA and NICU admission. To address cohort overlap, we applied rigorous methods, explicit exclusion criteria and conservative effect‐size recalculation where overlap could not be ruled out. Prespecified meta‐regression, clinically‐informed subgroup analyses and leave‐one‐out sensitivity tests consistently confirm robust pooled estimates. GRADE enables a transparent systematic assessment of evidence certainty for all outcomes.

Several limitations warrant consideration. First, substantial heterogeneity remained unexplained by metaregression (adjusted R ² ≤ 12.4%) for all outcomes. Second, most included studies were observational. Despite prioritising adjusted estimates, residual confounding from unmeasured or imprecisely measured covariates cannot be excluded, including adherence to dietary and pharmacologic therapy, health literacy, socioeconomic status and pregestational metabolic health. Additionally, substantial heterogeneity in glucose monitoring protocols and insulin titration across studies may have biased effect estimates toward or away from the null. Third, definitions of eGDM vary across studies (ranging from < 20 to < 24 weeks, with some using < 22 or < 16 weeks), which compromises cross‐study comparability. Standardised diagnostic thresholds for eGDM are urgently needed. Fourth, one large cross‐sectional study (Regnault et al.) using a non‐standard < 22‐week definition contributes to 71.9% of participants, but leave‐one‐out analysis confirms its exclusion does not meaningfully alter pooled estimates. Fifth, small‐study effects are evident for insulin use (Egger's p = 0.007; OR attenuated from 2.26 to 2.05) and suggested for preterm birth (p = 0.014; no missing studies imputed). Sixth, the observed association between eGDM and adverse outcomes may be confounded by treatment intensity. Several studies paradoxically reported lower risks of preterm birth, caesarean delivery or NICU admission in eGDM (e.g., Clarke, Rowan, De Muylder), which likely reflects intensified surveillance and earlier glycemic intervention rather than a genuine biological advantage. Thus, pooled ORs may be attenuated by differential management between early‐ and late‐diagnosed groups, an intervention bias inherent to observational designs that cannot be fully adjusted for.

5. Conclusions

In this meta‐analysis, eGDM is associated with higher odds of adverse maternal, perinatal and neonatal outcomes compared with LGDM and the strongest association is found with insulin use, followed by preterm birth, LGA and NICU admission with comparable magnitudes. These findings suggest that early diagnosis identifies a higher‐risk phenotype, supporting future studies of risk‐stratified management strategies. Future priorities include individual‐participant data meta‐analyses, standardised diagnostic thresholds, randomised trials of tailored protocols and longitudinal studies to clarify long‐term outcomes for mothers and offspring.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Forest plot comparing the odds of insulin use between women with eGDM and those with LGDM, stratified by adjustment status.

Figure S2: Forest plot comparing the odds of insulin use between women with eGDM and those with LGDM, stratified by diagnostic criteria.

Figure S3: Forest plot comparing the odds of insulin use between women with eGDM and those with LGDM, stratified by geographic region.

Figure S4: Forest plot comparing the odds of insulin use between women with eGDM and those with LGDM, stratified by screening type.

Figure S5: Leave‐one‐out sensitivity analysis comparing the odds of insulin use between women with eGDM and those with LGDM.

Figure S6: Funnel plot comparing the odds of insulin use between women with eGDM and those with LGDM.

Figure S7: Forest plot comparing the odds of caesarean delivery between women with eGDM and those with LGDM, stratified by adjustment status.

Figure S8: Forest plot comparing the odds of caesarean delivery between women with eGDM and those with LGDM, stratified by diagnostic criteria.

Figure S9: Forest plot comparing the odds of caesarean delivery between women with eGDM and those with LGDM, stratified by geographic region.

Figure S10: Forest plot comparing the odds of caesarean delivery between women with eGDM and those with LGDM, stratified by screening type.

Figure S11: Leave‐one‐out sensitivity analysis comparing the odds of caesarean delivery between women with eGDM and those with LGDM.

Figure S12: Funnel plot comparing the odds of caesarean delivery between women with eGDM and those with LGDM.

Figure S13: Forest plot comparing the odds of preterm birth between women with eGDM and those with LGDM, stratified by adjustment status.

Figure S14: Forest plot comparing the odds of preterm birth between women with eGDM and those with LGDM, stratified by diagnostic criteria.

Figure S15: Forest plot comparing the odds of preterm birth between women with eGDM and those with LGDM, stratified by geographic region.

Figure S16: Forest plot comparing the odds of preterm birth between women with eGDM and those with LGDM, stratified by screening type.

Figure S17: Leave‐one‐out sensitivity analysis comparing the odds of preterm birth between women with eGDM and those with LGDM.

Figure S18: Funnel plot comparing the odds of preterm birth between women with eGDM and those with LGDM.

Figure S19: Forest plot comparing the odds of LGA between women with eGDM and those with LGDM, stratified by adjustment status.

Figure S20: Forest plot comparing the odds of LGA between women with eGDM and those with LGDM, stratified by diagnostic criteria.

Figure S21: Forest plot comparing the odds of LGA between women with eGDM and those with LGDM, stratified by geographic region.

Figure S22: Forest plot comparing the odds of LGA between women with eGDM and those with LGDM, stratified by screening type.

Figure S23: Leave‐one‐out sensitivity analysis comparing the odds of LGA between women with eGDM and those with LGDM.

Figure S24: Funnel plot comparing the odds of LGA between women with eGDM and those with LGDM.

Figure S25: Forest plot comparing the odds of NICU between women with eGDM and those with LGDM, stratified by adjustment status.

Figure S26: Forest plot comparing the odds of NICU between women with eGDM and those with LGDM, stratified by diagnostic criteria.

Figure S27: Forest plot comparing the odds of NICU between women with eGDM and those with LGDM, stratified by geographic region.

Figure S28: Forest plot comparing the odds of NICU between women with eGDM and those with LGDM, stratified by screening type.

Figure S29: Leave‐one‐out sensitivity analysis comparing the odds of NICU between women with eGDM and those with LGDM.

Figure S30: Funnel plot comparing the odds of NICU between women with eGDM and those with LGDM. eGDM, early‐onset GDM, LGDM, late‐onset GDM, LGA, large‐for‐gestational‐age NICU, neonatal intensive care unit.

Data S1: dom70982‐sup‐0031‐SupinfoS1.docx.

Table S1: Detailed characteristics of studies included in the meta‐analysis and systematic review.

Data S2: Additional file 2. PRISMA Checklist.

Table S2: ROBINS‐I Risk of Bias Assessment.

Table S3: Meta‐regression analysis exploring sources of heterogeneity.

Table S4: Sensitivity analyses for the association between age of onset or diabetes duration and all‐cause mortality.

Mei G., Xia Y., Gan J., and Chen B., “Early‐Onset Versus Late‐Onset Gestational Diabetes Mellitus: A Systematic Review and Meta‐Analysis of Maternal, Perinatal and Neonatal Outcomes,” Diabetes, Obesity and Metabolism 28, no. 9 (2026): 7849–7863, 10.1111/dom.70982.

Data Availability Statement

The datasets used and analyzed in this investigation are accessible from the corresponding author in response to a legitimate request.