Maternal diabetes and risk of attention-deficit/hyperactivity disorder in offspring in a multinational cohort of 3.6 million mother–child pairs

Hyperglycemia affects an estimated 16% of pregnancies globally, and the prevalence of maternal diabetes mellitus (MDM) has increased alongside obesity, older maternal age, and improved diagnostics. ADHD is a common neurodevelopmental disorder (2–7% prevalence) with substantial personal and societal burden. Animal and mechanistic studies suggest hyperglycemia during pregnancy may influence offspring neurodevelopment via inflammation, oxidative stress, and epigenetic changes. Observational studies have reported associations between MDM (including gestational diabetes mellitus, GDM, and pregestational diabetes mellitus, PGDM) and ADHD in offspring, but many prior studies relied on self-report, had limited power, insufficient adjustment for confounders, and were conducted predominantly in White populations. The research question was whether maternal diabetes (overall and by subtype) is associated with ADHD risk in offspring, and to what extent associations may be causal versus confounded by shared familial factors. To address this, the authors conducted a large multinational cohort study with harmonized analytics to quantify the association between MDM (any, GDM, PGDM, type 1 and type 2) and ADHD, including within-family (sibling-matched) analyses to assess confounding by shared genetics and environment.

A prior meta-analysis reported approximately 40% higher ADHD risk in offspring of mothers with any diabetes, with some studies suggesting up to twofold risk with GDM. However, several contributing studies had methodological limitations: use of self-reported exposure, small sample sizes limiting statistical power, inadequate control for important confounders (particularly shared familial factors and maternal comorbidities), and restricted racial/ethnic diversity. These limitations motivate the need for large, population-based datasets with rigorous exposure and outcome definitions, comprehensive covariate control, and quasi-experimental designs (such as sibling comparisons) to discern causality from confounding.

Design and data sources: Population-based cohort study using linked mother–child health data from Hong Kong, New Zealand, Taiwan, and the Nordic countries (Finland, Iceland, Norway, Sweden; Nordic data pooled via NorPreSS). Data comprised territory-wide electronic health records, national registers, and insurance claims. A distributed network with a common data model and shared analytic package was used to harmonize data and analyses across sites; each site ran standardized code locally and shared aggregated results for meta-analysis. Cohort assembly and follow-up: Included all live births within site-specific periods with sufficient follow-up to capture ADHD diagnoses. For main analyses, only births with at least 6 years of follow-up were included (yielding 3,619,717 mother–child pairs: Hong Kong 362,338; Nordic countries 1,902,081; Taiwan 789,730; New Zealand 520,143). Exclusions included missing mother–child linkage, unknown infant sex, perinatal death or abortion, missing gestational age, missing mother’s birth date, and site-specific data issues. Follow-up began at birth and ended at ADHD outcome, death, or the site-specific end of data (2017–2020). Exposure assessment: Primary exposure was maternal diabetes mellitus (MDM), classified as gestational diabetes mellitus (GDM) and pregestational diabetes mellitus (PGDM), with PGDM further subtyped into type 1 and type 2. In Hong Kong, New Zealand, and Nordic countries, the last menstrual period (LMP) was derived by subtracting ultrasound-based gestational age from birth date; in Taiwan, LMP was defined as delivery date minus 280 days. Pregnancy was segmented into trimesters (0–90, 91–180, and 181 days to delivery). Site-specific algorithms based on diagnoses, prescriptions, and laboratory data identified exposure status. GDM medication status (medicated vs unmedicated) was also captured. Outcome definition: ADHD was identified using diagnosis and medication codes. In Hong Kong, New Zealand, and Taiwan, ≥1 ADHD diagnosis or ≥1 ADHD medication prescription defined cases. In Nordic countries, cases required either ≥2 ADHD diagnoses or ≥1 ADHD diagnosis plus ≥2 ADHD medication fills. In Hong Kong, New Zealand, and the Nordics, criteria had to be met at or after age 3 years. Covariates: A comprehensive set of potential confounders and risk factors was assessed, including demographics (maternal age, infant sex, birth year), multifetal pregnancy, proxies of socioeconomic status (income, education, insurance fees, deprivation), maternal lifestyle (smoking, alcohol), psychiatric and neurological conditions, chronic medical conditions (hypertension, renal disease, inflammatory bowel disease, autoimmune disease, thyroid disorders, polycystic ovary syndrome), BMI, and use of psychotropic and other relevant medications (including known or suspected teratogens). Assessment windows varied by site according to data availability and practice. Missing data were handled with indicator categories or imputation as appropriate by site. Statistical analysis: Cox proportional hazards models estimated hazard ratios (HRs) with 95% confidence intervals (CIs) for average treatment effect, using propensity score (PS) fine-stratification weighting (50 strata) to balance baseline covariates; robust standard errors addressed clustering. Residual imbalances (standardized difference >10%) were further adjusted in models. Sibling-matched analyses used stratified Cox models with mothers’ IDs defining strata, restricting to siblings discordant for exposure and outcome to control for shared genetic/familial factors. Site-specific estimates were combined via random-effects meta-analysis; heterogeneity was quantified by I2. Sensitivity analyses included: repeating main analyses including births with <6 years of follow-up; sex-stratified analyses; E-value computation for unmeasured confounding; restricting to ≥9 years follow-up; and alternative modeling with Poisson and negative binomial regression. Power: Based on observed sample sizes and risks, power to detect small differences was high (e.g., normal approximation power ~98.6% for MDM vs non-MDM). Ethics: Each site obtained approvals from relevant ethics and data custodians; data remained pseudonymized and onsite.

- Population and exposure: 3,619,717 mother–child pairs included. Proportion exposed to maternal diabetes: Hong Kong 8.0% (n=30,396), New Zealand 4.1% (n=21,326), Taiwan 13.7% (n=107,898), Nordic countries 6.6% (n=126,425). - Any maternal diabetes (MDM) vs no diabetes: pooled HR 1.16 (95% CI 1.08–1.24). Site-specific PS-weighted HRs: Hong Kong 1.16 (1.10–1.24), Nordics 1.20 (1.16–1.24), Taiwan 1.08 (1.05–1.10), New Zealand 1.22 (1.13–1.32). Heterogeneity I2=91%. - Gestational diabetes (GDM) vs no diabetes: pooled HR 1.10 (95% CI 1.04–1.17). Site-specific HRs: Hong Kong 1.15 (1.08–1.23), Nordics 1.15 (1.10–1.21), Taiwan 1.06 (1.04–1.09), New Zealand 1.00 (0.88–1.13). I2=79%. - Pregestational diabetes (PGDM) vs no diabetes: pooled HR 1.39 (95% CI 1.25–1.55). Site-specific HRs: Hong Kong 1.30 (1.06–1.59), Nordics 1.28 (1.21–1.36), Taiwan 1.57 (1.42–1.73), New Zealand 1.43 (1.29–1.59). I2=77%. - Type 1 PGDM vs no diabetes: pooled HR 1.46 (95% CI 1.24–1.71); site HRs: Hong Kong 1.44 (1.09–1.89), Nordics 1.32 (1.18–1.47), Taiwan 1.90 (1.48–2.43), New Zealand 1.35 (1.07–1.71). I2=58%. - Type 2 PGDM vs no diabetes: pooled HR 1.38 (95% CI 1.24–1.53); site HRs: Hong Kong 1.16 (0.87–1.55), Nordics 1.28 (1.21–1.37), Taiwan 1.52 (1.36–1.69), New Zealand 1.45 (1.30–1.63). I2=69%. - Sibling-matched comparison (discordant GDM exposure within families): pooled HR 1.05 (95% CI 0.94–1.17), indicating no significant within-family association. Site HRs: Hong Kong 1.11 (0.92–1.34), Nordics 1.07 (0.93–1.23), Taiwan 0.96 (0.91–1.01), New Zealand 1.26 (0.96–1.65). - Comparisons among diabetes types: GDM vs PGDM pooled HR 0.76 (95% CI 0.61–0.96), suggesting lower ADHD risk with GDM than PGDM; Type 2 vs Type 1 PGDM pooled HR 1.04 (95% CI 0.89–1.21), no difference. - Medicated vs unmedicated GDM: pooled HR 1.14 (95% CI 0.92–1.42), no statistically significant difference; site results varied. - Incidence rates (examples from Table 2): MDM-exposed vs unexposed IR per 1,000 person-years: Hong Kong 4.40 vs 3.67; Nordics 3.13 vs 2.47; Taiwan 11.71 vs 10.95; New Zealand 3.29 vs 2.57. - Sensitivity analyses: Results robust when including children with <6 years of follow-up, stratifying by sex, restricting to ≥9 years follow-up, and using Poisson/negative binomial models. E-values for pooled results ranged 1.43–2.28, indicating that modest unmeasured confounding could explain away the observed population-level associations, particularly for GDM.

The study addressed whether maternal diabetes is associated with offspring ADHD and evaluated causality using both population-level and within-family designs. Population-level analyses indicated small-to-moderate increased ADHD risk with any maternal diabetes, with stronger associations for pregestational diabetes (both type 1 and type 2) than for gestational diabetes. However, in sibling-matched analyses, discordant exposure to GDM within families did not translate into differential ADHD risk, suggesting that shared genetic and familial environmental factors likely confound the GDM–ADHD association. Compared with prior literature reporting larger effects (e.g., ~40% increased risk for any diabetes and up to twofold for GDM), this study’s risk estimates were smaller after extensive control for measured confounders and familial factors. The findings imply that while maternal metabolic disease may be associated with ADHD risk, particularly for PGDM, the association for GDM may not be causal. The E-value analyses support the interpretation that residual unmeasured confounding of modest magnitude could attenuate or explain the observed associations, especially at the population level. These results refine understanding of the etiologic role of intrauterine hyperglycemia versus familial/genetic liability in ADHD and emphasize the importance of quasi-experimental designs in perinatal epidemiology.

In a harmonized multinational cohort of over 3.6 million mother–child pairs, maternal diabetes was associated with a small-to-moderate increase in ADHD risk in offspring. Associations were strongest for pregestational diabetes (type 1 and type 2), while the population-level association observed for gestational diabetes attenuated to null in sibling comparisons, indicating likely confounding by shared familial factors. These findings contrast with prior reports of larger risks and call for reevaluation of the relative contributions of intrauterine hyperglycemia, glycemic control, disease severity, and genetic/familial factors to offspring ADHD. Future research should leverage family-based and genetically informed designs, incorporate detailed measures of glycemic control and disease severity, and investigate mechanistic pathways to clarify causality and identify modifiable targets during pregnancy.

- Use of administrative and clinical registry data entails variability across healthcare systems in coding practices, diagnostics, and treatment, which could introduce measurement error despite harmonized definitions. - Potential inaccuracy or incompleteness in diagnoses, prescriptions, and laboratory records; ADHD case definition may misclassify some cases and non-cases, potentially biasing estimates toward the null. - ADHD diagnoses often occur later in childhood; although a ≥6-year follow-up criterion (and a ≥9-year sensitivity analysis) was applied, some undiagnosed cases may remain. - Outcome definition differences across sites (e.g., stricter criteria in Nordic countries to avoid non-ADHD indications for ADHD medications) may contribute to heterogeneity. - Residual confounding may persist due to limited capture of maternal lifestyle factors (diet, physical activity) and paternal or family-level factors; E-values suggest modest unmeasured confounding could account for population-level associations. - Sibling-matched design, while controlling for shared familial factors, may be susceptible to bias from sibling-specific confounders and reduced generalizability to families without multiple births or discordant exposures/outcomes. - Heterogeneity across sites (high I2 in some comparisons) indicates variation by population or data context that may limit generalizability of pooled estimates.