Globally, 16% of pregnant women experience hyperglycemia1,2. The prevalence of maternal diabetes mellitus (MDM) has increased worldwide, which is associated with the growing epidemic of obesity, advancing maternal age and improved diagnostic approaches for MDM3,4. There are calls for greater attention to the risks associated with diabetes in pregnancy given the increasing trend of gestational diabetes and preexisting type 2 diabetes5. Animal studies have demonstrated the adverse effects of hyperglycemia during pregnancy on inflammatory responses, intrauterine oxidative stress and imbalance in epigenetic mechanisms, which may contribute to poor neurodevelopment in the offspring6,7.

Attention-deficit/hyperactivity disorder (ADHD) is a neurodevelopmental disorder characterized by hyperactivity, impulsivity and inattentiveness8. Currently, ADHD is estimated to affect 2% to 7% of children worldwide9,10, making it one of the most common disorders among school-aged children. ADHD not only adversely impacts the affected individuals but also poses a substantial burden on their families and the wider society11,12,13,14,15,16,17. A complex interaction between genetic, environmental and psychosocial risk factors is thought to be responsible for the etiology of ADHD8.

Emerging evidence has suggested that both pregestational diabetes mellitus (PGDM) and gestational diabetes mellitus (GDM) are associated with ADHD. A previous meta-analysis found that the offspring of diabetic mothers were at 40% higher risk of ADHD18. However, some of the included studies used self-reported data19,20,21, had limited statistical power20,22 or had limited adjustment for confounders19,20,23,24, especially familial factors, and were predominantly conducted in White populations7,19,20,21,22,23,25,26. To account for these limitations, we conducted the current cohort study based on population-based data covering over 3.6 million mother–child pairs in Hong Kong, New Zealand, Taiwan, Finland, Iceland, Norway and Sweden, with extensive coverage of relevant covariates, to assess the association between MDM and the risk of ADHD in offspring.


This study consisted of children from all live births within the site-specific study period (Table 1). All mother–child pairs were linked with exact deterministic linkage27. Children without valid mother–child linkage or with incomplete birth information (for example, sex or gestational age) or without at least 6 years of follow-up were excluded to allow sufficient follow-up time to capture ADHD outcomes, as a diagnosis is often deferred until a child is of school age28. The follow-up period for each child started on the date of birth and ended on the date of outcome occurrence, date of death, or end of data source-specific study period, whichever came first. Details of the sample size calculation are given in Extended Data Fig. 1.

Table 1 Data characteristics

We identified the start and end of pregnancy by the date of the last menstrual period (LMP) and the child’s date of birth (Fig. 1). In Hong Kong, New Zealand and the Nordic countries, LMP was determined by subtracting gestational age at birth (determined by ultrasound) from the date of birth; in Taiwan, this was defined as the date of delivery minus 280 days29. As hyperglycemia may affect neurodevelopment differently at different trimesters, we divided the pregnancy period into first trimester (0–90 days after the LMP), second trimester (91–180 days after the LMP) and third trimester (181 days after the LMP to delivery). The primary exposure is MDM (including GDM and PGDM). MDM was further classified as GDM (including those receiving and not receiving medications) and PGDM (including type 1 and type 2 PGDM; Extended Data Fig. 2). PGDM refers to existing diabetes before pregnancy, and GDM refers to diabetes diagnosed only during pregnancy. Data source-specific criteria were applied to ensure the best identification of the exposure status (Supplementary Table 1).

Fig. 1: Illustration of cohort inclusion and pregnancy periods.
figure 1

aThe earliest date of mothers’ health data in each data source. bBaseline conditions included: demographics, maternal conditions and medication use. For HK analysis, maternal age at delivery and birth year were assessed at the date of delivery, body mass index (BMI) was assessed from LMP − 365 days to LMP − 1 day and all other covariates were assessed before LMP; for analysis in Nordic countries and New Zealand, medication use was assessed from LMP − 365 days to LMP − 1 day, and diagnoses were assessed from LMP − 365 days to delivery date; for Taiwan analysis, all covariates were assessed within 2 years before the date of delivery. cExposure window: (1) Period before pregnancy, (2) first trimester: LMP to LMP + 90 days, (3) second trimester: LMP + 91 days to LMP + 180 days, and (4) third trimester: LMP + 181 days to delivery date. dEarliest of: date of ADHD diagnosis, date of first ADHD medication prescription, date of death, end of database catchment period.

ADHD was defined by each site using specific diagnosis and medication codes as reported in previous studies and detailed in Supplementary Table 1 (refs. 30,31). In Hong Kong, New Zealand and Taiwan, children with either one ADHD diagnosis or one ADHD medication prescription were regarded as having the outcome. In Nordic countries, the outcome was defined as (1) ≥2 records of ADHD diagnoses or (2) ≥1 record of ADHD diagnosis and ≥2 records of ADHD medication prescription fills. In Hong Kong, New Zealand and Nordic countries, these criteria needed to be met at or after the age of 3 years to exclude invalid ADHD diagnoses (Supplementary Table 1).

We identified 3,619,717 mother–child pairs to be included in the analysis (Fig. 2). Overall, 8.0% (n = 30,396), 4.1% (n = 21,326), 13.7% (n = 107,898) and 6.6% (n = 126,425) of children were born to mothers with diabetes in Hong Kong, New Zealand, Taiwan and the Nordic countries, respectively. Child and maternal characteristics are summarized in Supplementary Table 2. Covariate balances were achieved after propensity score (PS) weighting with standardized differences <10%, except for obesity in New Zealand (Supplementary Table 3).

Fig. 2: Flowchart of cohort identification.
figure 2

aNot applicable to Finland data, where residence/migration data are not available. bIndividuals could fulfill more than one exclusion criteria.

Primary analyses and sibling-matched comparisons

When comparing children born to mothers with any diabetes during pregnancy (NMDM = 286,045) and unexposed children (Nnon-MDM = 3,333,672), we identified a higher risk of ADHD (pooled hazard ratio (HR): 1.16, 95% confidence interval (CI): 1.08-1.24). The cumulative incidence of ADHD for different data sources is shown in Extended Data Fig. 3. Similarly, we identified a higher risk of ADHD across different types of MDM when comparing children whose mothers had GDM (NGDM = 230,798), PGDM (NPGDM = 54,825), type 1 PGDM (Ntype 1-PGDM = 11,444) and type 2 PGDM (Ntype 2-PGDM = 42,977) to those whose mothers did not have diabetes (GDM pooled HR: 1.10, 95% CI: 1.04 to 1.17; PGDM pooled HR: 1.39, 95% CI: 1.25-1.55, type 1 PGDM pooled HR: 1.46, 95% CI: 1.24-1.71; type 2 PGDM pooled HR: 1.38, 95% CI: 1.24-1.53; Fig. 3 and Table 2). We applied sibling-matched analysis for GDM to control for shared familial confounding including unmeasured lifestyle factors. Siblings who were born to the same mother but with discordant exposure to GDM during the respective pregnancy episodes did not differ in the risks of ADHD (pooled HR: 1.05, 95% CI: 0.94-1.17, NGDM = 72,791, Nnon-MDM = 75,082; Fig. 4 and Table 2). Risks of ADHD for children were similar across those whose mothers had GDM diagnosed at different trimesters in Hong Kong; in Taiwan and New Zealand, the risk of ADHD was highest for children born to mothers with GDM diagnosed in the first trimester (Supplementary Table 4).

Fig. 3: Meta-analyses of maternal diabetes and the risk of ADHD in offspring.
figure 3

Data are presented as HRs and 95% CIs, which were adjusted for demographics, socioeconomic status, birth year, multifetal pregnancies, maternal conditions and use of relevant medications using Cox proportional hazard regression, with a significance level of 5% for a two-sided test. No adjustments were made for multiple comparisons. df, degrees of freedom; IV, inverse variance; s.e., standard error; T1DM, type 1 pregestational diabetes; T2DM, type 2 pregestational diabetes.

Table 2 Risks of ADHD in offspring with different maternal diabetes by data sources
Fig. 4: Meta-analyses of discordant GDM exposure in siblings and the risk of ADHD.
figure 4

Data are presented as HRs and 95% CIs, which were adjusted for demographics, socioeconomic status, birth year, multifetal pregnancies, maternal conditions and use of relevant medications using Cox proportional hazard regression, with a significance level of 5% for a two-sided test.

Comparisons in different MDM exposures

Children born to mothers with GDM may have a lower risk of ADHD when compared with children born to mothers with PGDM (pooled HR: 0.76, 95% CI: 0.61-0.96, NGDM = 230,798, NPGDM = 54,825). Risks of ADHD did not differ between children born to mothers with type 2 PGDM and type 1 PGDM (pooled HR: 1.04, 95% CI: 0.89-1.21, Ntype 1-PGDM = 11,444, Ntype 2-PGDM = 42,977). Children whose mothers had GDM requiring antidiabetic medication had a similar risk of ADHD when compared with children born to unmedicated mothers with GDM (pooled HR: 1.14, 95% CI: 0.92-1.42, Nmedicated = 25,206, Nunmedicated = 205,592; Table 2 and Extended Data Fig. 4).

Sensitivity analyses

Results from the sensitivity analyses were similar to the primary analyses when the analytic cohorts included children with less than 6 years of follow-up and were stratified by sex (Supplementary Table 5). We computed E-values to facilitate the interpretation of results in the presence of unmeasured confounding. The E-values of the pooled results range from 1.43 to 2.28, suggesting that unmeasured confounding with an association magnitude equal or greater to both the exposure and outcome could explain away the observed associations respectively (Supplementary Table 6). In other words, any residual confounding is likely to lead to an even smaller estimate. A post hoc analysis also yielded similar results when the analytic cohorts included only children with 9 or more years of follow-up (Supplementary Table 7). Finally, results from the Poisson and negative binomial regression were similar to the Cox regression in the main analysis (Supplementary Table 8).


In this large multinational cohort study, including over 3.6 million mother–child pairs and leveraging a common data model and analytic approach, we found that MDM overall, GDM and PGDM were associated with a small-to-moderate risk of ADHD in offspring. After controlling for shared familial genetic and social factors in the sibling-matched analyses, risks of ADHD did not differ between siblings with discordant exposure to GDM in pregnancy. Due to the discrepancy around the within- and between-family analyses, we speculate that the relationship between GDM exposure and ADHD may be confounded by familial factors.

Compared with a previous meta-analysis which showed a 40% and twofold increased risk of ADHD in children born to mothers with any diabetes and GDM respectively, our study demonstrates a relatively smaller risk estimate of ADHD in children born to mothers with any diabetes after comprehensively controlling for potential confounders. The E-values for the primary analyses ranged from 1.43 to 2.28 and were considered relatively small32. Thus, it is likely that there are unmeasured confounders of this magnitude, such as disease severity or paternal factors, that may explain the identified association in our primary analyses. More importantly, although we observed a higher risk of ADHD in offspring whose mothers had GDM at the population level, the association was null in the sibling-matched analysis. Thus, shared familial or genetic factors are likely to at least partly explain the observed association at least between GDM and ADHD in our primary analyses.

In our study, children born to mothers with medicated GDM had a similar risk of ADHD when compared with those with unmedicated GDM. Existing literature regarding the effects of GDM treatment on the risk of ADHD in offspring is scarce. A study using US private healthcare data found a 38% increased risk in children born to mothers requiring treatment for GDM than those with unmedicated GDM33. Combining our results with currently available evidence, it remains unclear whether maternal antidiabetic medication during pregnancy could increase the risk of ADHD in offspring or if the severity of GDM requiring antidiabetic medication underlies the association. This further poses a question on the role of glycemic control during pregnancy in mitigating the risk of ADHD in children whose mothers have diabetes.

Our study has limitations. First, most data originated from reimbursement or other administrative purposes rather than research purposes, which represent a variety of data sources, healthcare settings, coding practices, diagnostics criteria, and treatment approaches. Nevertheless, data sources included in this study are all based on high-quality territory-wide electronic health records with robust mother–child linkage and comprehensive information on maternal and child medical records that have been used extensively for pregnancy-related studies30,34,35,36. Moreover, we applied data source-specific identification criteria for not only the exposure and outcome but also the covariates to maximize the comparability of variables included in the study. Second, there may be inaccuracy in coding and incompleteness of diagnoses, prescriptions and laboratory test records. However, we used electronic records and the included pregnant women are likely to have received more frequent clinical monitoring throughout pregnancy. Furthermore, we used a comprehensive set of data ranging from diagnosis records, prescription records and laboratory test records for pregnant women. We defined the study outcome as a diagnosis for ADHD or with ADHD medication prescriptions to capture all possible cases to increase the power of our study, but we acknowledge that misclassification could exist, and some children with minor symptoms of ADHD might have been included in the non-ADHD group. Similarly, nondifferential misclassification of ADHD patients may direct our results towards the null, leading to a smaller effect estimate than previous studies. First, ADHD may not be diagnosed until later in life, which may lead to inclusion of undiagnosed ADHD patients in our non-cases. We therefore conducted a post hoc sensitivity analysis (Supplementary Table 7) to assess the effect of insufficient follow-up duration in our main analyses. After only including mother-baby pairs with at least 9 years of follow-up duration, we found a similar result to the main analyses, demonstrating the robustness of our study conclusion. Similarly, the rates of detection and diagnosis of ADHD were expected to vary with age. This finding implies that the hazards for exposed and unexposed children would naturally exhibit deviations from proportionality throughout the observation period (Extended Data Fig. 3)8. We therefore recommend interpreting the HRs we present as a weighted average of the time-varying hazard ratios within our observation period. Nonetheless, sensitivity analyses using Poisson regression and negative binomial regression, which do not rely on the proportional hazards assumption, yielded consistent estimates that support our interpretation and conclusion (Supplementary Table 8). Second, some ADHD medication used in the Nordic countries may be used for other non-ADHD conditions such as narcolepsy. Therefore, ADHD case identification in Nordic countries required at least one ADHD diagnosis and at least two prescription fills for an ADHD medication and thus we were not likely to capture children using ADHD medication for other non-ADHD conditions. Third, maternal lifestyle factors such as physical activity and diet may not be fully captured in electronic health records such that inadequate adjustment for these factors could lead to residual confounding. We thus applied the current study design and analytic approaches to address this concern: 1) the use of sibling-matched analysis for GDM to control for shared familial confounding including lifestyle factors; 2) the computation of E-value to aid the interpretation of the results in presence of unmeasured confounding. For MDM, E-value computation showed that any residual confounding is likely to lead to an even smaller estimate. Therefore, even if these behavioral factors could explain the association, it is unlikely to affect our conclusion that there is a small-to-moderate association between MDM and ADHD, whereas the association between GDM and ADHD is unlikely to be causal. Finally, although the sibling-matched analyses allowed us to control for unmeasured, shared confounding, the design could amplify confounding from factors unique to each sibling37. Therefore, we draw our conclusion based on complementary study designs, including the unrelated cohort analyses, sibling-matched comparisons and computation of E-values. More importantly, our results remain robust in all sensitivity analyses and consistent across all data sources from various populations.

It was hypothesized that hyperglycemia may alter the intrauterine environment with increased inflammation, metabolic stress and lipotoxicity, which may affect the neurodevelopment of offspring6. Our study, however, only found a small-to-moderate effect between MDM and ADHD where the effect is likely to be confounded by shared genetic and familial factors, at least in the case of GDM. Future studies should explore the specific roles of genetic factors and glycemic control during different developmental stages of the human embryonic brain.


Study design

This is a population-based cohort study with linked mother–child pairs based on healthcare data in Hong Kong, New Zealand, Taiwan and Nordic countries (Finland, Iceland, Norway and Sweden; data nested within the NorPreSS collaboration38). The study results were reported following the STrengthening the Reporting of OBservational studies in Epidemiology statement.

All study sites used pseudonymized patient-level electronic health data derived from the respective territory-wide administrative, clinical, or register databases. The list of data sources is provided in Table 1. We applied a distributed network approach with a common data model to harmonize the data structure and standardize the contents from different data sources. Briefly, the coordinating center at the University of Hong Kong distributed a common analytic package for generating aggregated results based on the common data model38,39. Site investigators conducted the analyses locally in Hong Kong, New Zealand, Taiwan and Nordic countries where data from individual Nordic countries were pooled into one cohort and analyzed centrally38, before sharing aggregated results with the study coordinator. This approach preserved data confidentiality as the individual-level data remained at each site40. Moreover, we were able to maintain the consistency of analyses among sites with common analytics41. The codes used to identify relevant diagnoses and medication prescriptions from each site are presented in Supplementary Table 8.

Covariate assessment

Covariates were selected based on known confounders for the study association and risk factors for the study outcomes30,42,43,44,45,46,47,48,49,50, including demographic factors such as maternal age47,51,52,53, infant sex (as directly recorded by clinicians)54, socioeconomic status55,56, birth year, multifetal pregnancies57 and other maternal factors including smoking44,47,58,59, alcohol consumption47,60,61, psychiatric and neurological conditions47,48,62,63,64, other chronic medical conditions (hypertension51,65, renal disease, inflammatory bowel disease66,67, autoimmune disease46,68,69, thyroid disorders43,47,70,71 and polycystic ovary syndrome45,72,73), BMI44,47,51,74 and use of psychotropic medication42,47,49, antihypertensives, ADHD medication and known or suspected teratogenic medication. Various measures were applied as the proxy of socioeconomic status for each data source according to their respective practice (namely, median household income in Hong Kong, education level in Nordic countries, insurance fees in Taiwan and deprivation quintile in New Zealand). Definitions of the covariates by study site are available in Supplementary Table 8. A schematic directed acyclic graph illustrating the causal relationships between the different covariates and the exposure and/or outcome is shown in Extended Data Fig. 5.

Comparison groups

In our primary analyses, we compared the ADHD status in children born to mothers with any MDM, GDM, any PGDM, type 1 PGDM and type 2 PGDM, with children born to mothers without any diabetes. We also conducted sibling-matched analyses which compared the ADHD status in children born to the same mothers but with discordant GDM exposure. In our secondary analyses, we compared the ADHD status between children born to mothers with different diabetes subtypes, namely GDM and PGDM, type 2 PGDM and type 1 PGDM, and medicated GDM and unmedicated GDM.

Statistical analysis

We estimated HRs of average treatment effect with 95% CIs to study the associations between MDM status and ADHD using Cox proportional hazard regression models. Propensity score (PS) fine-stratification weighting was used to address the differences in baseline covariates. PS, the probability of receiving treatment conditional on the observed characteristics at baseline, can be applied to account for confounding effects efficiently in observational studies50,75. We used PS fine-stratification weighting because of the greater precision, less residual and equivalent bias control compared to traditional PS methods75,76. The PS is first used to create 50 fine strata; weights for both exposure and reference patients in all strata are subsequently calculated based on the total number of patients within each stratum, whereas strata with no exposed or reference patients are dropped out before weight calculation76. We applied robust standard errors to adjust for data clustering. All the covariates listed in Supplementary Table 8 were included in the PS model. Factors with standardized differences greater than 10% were further adjusted in the Cox models77. For missing data, indicator variables for missing maternal characteristics were included in the Nordic (folic acid use, education level, cohabitation, parity, and non-Nordic place of birth) and New Zealand (mother’s socioeconomic status and BMI) models; median imputation was applied for socioeconomic status in Taiwan for missing insurance fees (missing rate = 0.13%).

We conducted sibling-matched analyses to control for shared genetic, familial and environmental confounding factors and used stratified Cox regression with a separate stratum for each family identified by the mother’s unique identification number. Only sibling pairs with discordant exposure and outcome statuses were informative and contributed to the effect estimates.

We pooled the effect estimates from each data source in a meta-analysis using a random-effect model. Meta-analyses were represented in forest plots and the I2 statistic was used to quantify heterogeneity between sites. CIs not overlapping 1.0 were considered statistically significant. Statistical Analysis System (SAS) v9.4 (SAS Institute) and R Foundation for Statistical Computing version 3.6.0 were used for data analysis.

Sensitivity analyses

We conducted sensitivity analyses to test the validity and robustness of the study results. First, we repeated the main analysis in the entire mother-baby cohort, including offspring with less than 6 years of follow-up. Second, we stratified the analyses by offspring’s sex due to a higher prevalence of ADHD in males. Third, to assess the impact of any unmeasured confounders, we computed the E-value, which is defined as the minimum strength of association that an unmeasured confounder would need to have with both exposure and outcome, conditional on the measured covariates, to explain away an observed association78. Fourth, a post hoc sensitivity analysis including only mother-baby pairs with at least 9 years of follow-up was conducted to assess the effect of insufficient follow-up duration in our main analyses. Finally, Poisson and negative binomial regression models were also applied in the sensitivity analyses to test the robustness of our results in the presence of model uncertainty.

Ethics and inclusion statement

The study used healthcare data obtained from Hong Kong, New Zealand, Taiwan and Nordic countries. These data encompassed various sources such as electronic health records, registers, and insurance records. Each participating site followed the relevant local ethics and regulatory frameworks for study approval, namely the Finnish Institute for Health and Welfare (THL/1551/6.02.00/2018, THL/1673/5.05.00/2019) and the Social Insurance Institution of Finland (Kela 148/522/2018 and Kela 117/522/2019) in Finland; the University of Hong Kong/Hospital Authority Hong Kong West Cluster (UW20-051) in Hong Kong; the National Bioethics Committee (VSNb2018060017/03.01) in Iceland; the Norwegian Data Inspectorate (17/02068/Norwegian Data Inspectorate) and the Regional Committee for Medical and Health Research Ethics (2017/2546/REC South-East Norway) in Norway; the New Zealand Health and Disability Ethics Committee (13789) in New Zealand; the Swedish Ethical Review Authority (Dnr 2015/1826-31/2, 2017/2238-32, 2018/1790-32, 2018/2211-32, 2022-04004-02) in Sweden; and the National Cheng Kung University Human Research Ethics Committee (110-453) in Taiwan.

We fully endorse the Nature Portfolio journals’ guidance on authorship and inclusion. All collaborators of this study have fulfilled the criteria for authorship required by Nature Portfolio journals have been included as authors, as their participation was essential for the design and implementation of the study. Roles and responsibilities were agreed among collaborators ahead of the research. This work includes findings that are locally relevant, which have been determined in collaboration with local partners. This research was not severely restricted or prohibited in the setting of the researchers and does not result in stigmatization, incrimination, discrimination or personal risk to participants. Local and regional research relevant to our study was taken into account in citations.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.