This study obtained data from the UK Biobank (UKB), a large-scale prospective cohort that recruited more than 500,000 participants aged 40-69 years from 2006 to 2010. At enrollment, participants completed detailed questionnaires on demographic characteristics, behavioral factors, environmental exposures, and medical history, and underwent physical exams and standardized biological measurements^[18]. During follow-up, the UKB was linked to nationwide databases, such as primary care data, hospitalization records, death registries, and disease-specific registers, to identify disease diagnoses. The UK Biobank study received ethical approval from the North West Multi-Centre Research Ethics Committee (REC reference: 21/NW/0157, IRAS project ID: 299116), and all participants provided written informed consent. The present analysis was conducted under UK Biobank Application Number 103011 and required no additional institutional ethical approval because only de-identified data were used.

We conducted both cross-sectional and prospective analyses. For the cross-sectional analyses, we excluded 51,079 women without a history of childbirth and 218,033 women without a prior history of GDM or HDP. GDM was identified using ICD-10 code O24.4 (and corresponding ICD-9 codes 648.0 and 648.8). HDP was considered present if any of the following ICD-10 codes were recorded: O11, O13, O14.0, O14.1, O14.9, O15.0, O15.1, O15.2, and O15.9 (and corresponding ICD-9 codes 642.3, 642.4, 642.5, 642.6, and 642.7). We also excluded 672 women with missing data on lifestyle factors, 86 women with missing genetic data, and 177 women lacking information on major covariates. Consequently, data from 3,251 women were used for the cross-sectional analyses [Figure 1]. For the prospective analyses, we further excluded 142 women with prevalent diabetes at recruitment, resulting in a final sample of 3,109 women.

A combined lifestyle score was constructed using six behavioral indicators [Supplementary Tables 1 and 2]. Following a published approach^[19], each indicator was initially classified into three categories and then dichotomized as healthy (1) or unhealthy (0). The combined lifestyle score (range: 0-6) was derived by summing the six dichotomized items, with higher scores indicating a healthier lifestyle. For diet, we assessed adherence to the dietary priorities for cardiovascular health^[20] across seven core food groups: fruits, vegetables, fish, refined grains, whole grains, processed red meat, and unprocessed red meat. Participants were classified as having a healthy diet if they met ≥ 6 categories, a moderate diet if they met 2-5 categories, and a poor diet if they met ≤ 1 category; only the healthy category contributed 1 point. Physical activity was evaluated based on World Health Organization recommendations^[21] and categorized as active (≥ 150 min of moderate or ≥ 75 min of vigorous activity/week), moderate (partial compliance), or inactive (no compliance); only the active category contributed 1 point. Smoking status was categorized as never, former, or current smoker; only never smoking contributed 1 point. For alcohol consumption, never drinkers were assigned 1 point, whereas low-risk drinkers (≤ 7 standard drinks/week; 1 drink = 14 g ethanol) and high-risk drinkers (> 7 drinks/week or daily)^[22] were grouped into unhealthy. Sleep duration was categorized as healthy (7-8 h/night), moderate (6 or 9 h), and poor (≤ 5 or ≥ 10 h/night), with healthy sleepers assigned 1 point^[23]. BMI, calculated as weight (kg)/height²(m²), was categorized as healthy (< 25.0 kg/m²; 1 point) or unhealthy (≥ 25.0 kg/m²)^[24]_. To ensure adequate statistical power, we divided the combined lifestyle score into five categories (0, 1, 2, 3, and 4-6 points) in the main analyses and further classified it into three groups: unhealthy (0-1 points), intermediate (2-3 points), or healthy (4-6 points). In addition, we constructed an alternative combined lifestyle score excluding BMI. Comparing effect estimates between the two scores allowed us to indirectly examine whether BMI mediated the association between the combined lifestyle score and T2DM risk.

The PRS for T2DM quantifies genetic susceptibility using a standardized metric available in the UKB (Field ID: 26285). The PRS was calculated as the weighted sum of allele dosages across genome-wide risk variants, with weights defined by posterior effect sizes estimated from independent genome-wide association study summary statistics^[25]. Previous studies have validated the clinical utility of this standardized PRS within the UKB^[26,27]. In this study, the PRS was categorized into quartiles: Q1 (lowest genetic risk), Q2, Q3, and Q4 (highest genetic risk). To increase statistical power for the joint analysis of combined lifestyle and genetic susceptibility, the cohort was stratified into a low genetic risk group (PRS quartiles Q1-Q2) and a high genetic risk group (PRS in quartiles Q3-Q4).

Prevalent T2DM at baseline was defined as meeting one or more of the following criteria: (1) self-reported diabetes at the baseline assessment (field 20002); (2) linked historical hospital records indicating diabetes (ICD-9/ICD-10 codes); or (3) baseline HbA1c ≥ 48 mmol/mol in accordance with the American Diabetes Association guidelines^[28]. Incident T2DM cases during follow-up were ascertained from linked inpatient hospital records and death registries using ICD-10 codes (E11) [Supplementary Table 3]. For each participant, follow-up was calculated from the baseline assessment to the first occurrence of T2DM, death, loss to follow-up, or the end of follow-up (December 2021), whichever occurred first.

Major covariates were collected through the UKB baseline assessment system [Supplementary Table 2]. Demographic characteristics included age (continuous), race/ethnicity (White/non-White), and education level (high: university degree or professional qualification; low: other qualifications). Reproductive history variables included parity (categorized as 0, 1, 2, or ≥ 3 live births) and oral contraceptive use (never/ever). Family history of diabetes was defined as self-reported diabetes in parents or siblings (binary: yes/no). Other covariates used in sensitivity analyses included the Townsend deprivation index, social support, postmenopausal status, and medication use. The Townsend deprivation index, derived from participants’ residential postcode at baseline, was included as a continuous measure of socioeconomic deprivation, with higher values indicating greater deprivation^[29]. Social support was derived from baseline questionnaire data^[30]. Postmenopausal status was coded as yes or no based on self-reported menopausal status at baseline. Medication use for metabolic diseases was categorized as self-reported regular intake of antihypertensive, lipid-lowering, or antidiabetic medications at baseline.

Baseline characteristics of women with prior GDM or HDP were compared across combined lifestyle score categories (0, 1, 2, 3, and ≥ 4). Continuous variables were compared across groups using analysis of variance, and categorical variables were compared using the chi-square test or Fisher’s exact test, as appropriate.

In the cross-sectional analyses, we used logistic regression models to assess associations between the combined lifestyle score and prevalent T2DM. Odds ratios (ORs) with 95% confidence intervals (CIs) were computed. We adjusted for covariates sequentially. In Model 1, we accounted for age (continuous) and race/ethnicity (White or non-White). Model 2 additionally included educational level (categorical), parity (1, 2, or ≥ 3), oral contraceptive use (binary: yes/no), and family history of diabetes (yes/no). When analyzing individual lifestyle factors, Model 2 was additionally adjusted for other lifestyle factors (excluding BMI).

In the prospective analyses, we used Cox proportional hazards regression models to evaluate associations of individual lifestyle factors, the combined lifestyle score, and the PRS with new-onset T2DM, following the same analytical strategy as in the cross-sectional analyses. The proportional hazards assumption of the Cox models was assessed using Schoenfeld residuals. No significant violations were detected (all global P > 0.05), indicating that the assumption was satisfied. Hazard ratios (HRs) and corresponding 95%CIs were reported. We examined whether the relationship between lifestyle factors and T2DM varied across population subgroups using stratified analyses. Participants were stratified by PRS (low vs. high), age (< 55 years vs. ≥ 55 years), BMI (< 30 kg/m² vs. ≥ 30 kg/m²), and family history of diabetes (yes vs. no). To assess effect modification (interaction), we added interaction terms to our Cox models and tested their statistical significance using the likelihood ratio test. To evaluate the joint effect of PRS and lifestyle, we conducted a joint association analysis. The previously defined dichotomous genetic risk variable (low vs. high) was cross-classified with the three-category lifestyle variable (unhealthy, intermediate, and healthy), generating six combined exposure groups.

Five sensitivity analyses were conducted to assess the robustness of the association between the combined lifestyle score and incident T2DM. First, to evaluate the contribution of lifestyle factors independent of body weight, we reconstructed an alternative lifestyle score excluding BMI (score range: 0-5). Second, to minimize reverse causality, we excluded participants diagnosed with T2DM during the first two years of follow-up. Third, we handled missing covariate data using multiple imputation via the mice package in R to avoid potential bias^[31], generating 10 imputed datasets. Continuous covariates were imputed using predictive mean matching (pmm), and categorical covariates using logistic regression (logreg). Analyses were repeated in each imputed dataset, and estimates were pooled according to Rubin’s rules^[32]. Fourth, based on Model 2 in the prospective analyses, further adjustments were made for the Townsend deprivation index (continuous), menopausal status (yes/no), social support (high/low), and medication use for metabolic diseases (yes/no). Fifth, to assess the robustness of our findings, we performed sensitivity analyses separately among women with a history of GDM and those with a history of HDP.

All analyses were conducted with R software (version 4.3; R Foundation for Statistical Computing). A P value < 0.05 (two-sided) was regarded as indicating statistical significance.

The cross-sectional analyses included 3,251 women [Figure 1], of whom 3,109 were eligible for prospective analyses. Within the prospective cohort, 1,017 (32.7%) had a history of GDM only, 2,081 (66.9%) had HDP only, and 11 (0.4%) had a history of both conditions. Five groups were created according to the combined lifestyle scores among the 3,109 participants: 0 (n = 111), 1 (n = 585), 2 (n = 1,087), 3 (n = 956), and ≥ 4 (n = 370). No participants achieved the maximum score of 6. Compared with those in the lower score groups, participants with higher combined lifestyle scores were mostly younger and had higher education levels [Table 1]. They also had lower BMI and were more likely to have healthier lifestyle behaviors, including diet, smoking status, alcohol consumption, physical activity, and sleep duration (all P < 0.05). Similar findings were observed when the combined lifestyle score (range 0-6) was recategorized into healthy, intermediate, and unhealthy categories [Supplementary Table 4].

In the logistic regression, after adjusting for age, race/ethnicity, education, parity, oral contraceptive use, and family history of diabetes, we observed a significant inverse dose-response relationship (P for trend < 0.01; Table 2). Compared with participants with a combined lifestyle score of 0, the adjusted ORs (95%CIs) for those with scores of 1, 2, 3, and ≥ 4 were 0.82 (0.49, 1.38), 0.49 (0.29, 0.83), 0.31 (0.17, 0.55), and 0.36 (0.19, 0.69), respectively.

In prospective analyses of 3,109 women with a history of GDM or HDP (median follow-up 14.5 years), 247 incident T2DM cases were documented. Cox proportional hazards models, adjusted for age, race/ethnicity, education, parity, oral contraceptive use, and family history of diabetes, showed that most modifiable lifestyle factors were independently associated with T2DM [Table 3]. Compared with participants with a combined lifestyle score of 0, the multivariable-adjusted HRs (95%CIs) for incident T2DM were 0.83 (0.49, 1.39) for 1 point, 0.54 (0.32, 0.93) for 2 points, 0.37 (0.22, 0.64) for 3 points, and 0.29 (0.16, 0.53) for ≥ 4 points (P for trend < 0.001), indicating a pronounced inverse association between the combined lifestyle score and T2DM risk [Table 2 and Figure 2].

No significant interaction was observed between PRS and lifestyle (P for interaction = 0.247, Supplementary Table 5). In the high-PRS group, compared with the unhealthy reference category, the adjusted HR for participants in the healthy category (≥ 4 points) was 0.32 (0.16, 0.65). The joint analysis of PRS and the combined lifestyle score showed that all PRS-lifestyle combined groups exhibited a markedly reduced T2DM risk compared with the reference group (unhealthy lifestyle/high PRS) [Figure 3]. In addition, a combined healthy lifestyle was similarly associated with lower T2DM risk across subgroups defined by BMI, age, and family history of diabetes, with no statistical heterogeneity detected [Supplementary Table 5]. For instance, the adjusted HRs were 0.42 (95%CI: 0.22, 0.78) among individuals with BMI < 30 kg/m² and 0.61 (95%CI: 0.24, 1.54) among those with BMI ≥ 30 kg/m² (P for interaction = 0.704). Similarly, the association was observed in both age groups: HR was 0.32 (95%CI: 0.16, 0.65) for age < 55 years and 0.44 (95%CI: 0.21, 0.90) for age ≥ 55 years (P for interaction = 0.780).

In a sensitivity analysis of a lifestyle score based on five factors (excluding BMI), the negative association between higher lifestyle scores and reduced T2DM risk remained robust. Compared with the reference group (0 points), the adjusted HRs and 95%CIs were 0.72 (0.43, 1.18) for 1 point, 0.55 (0.33, 0.89) for 2 points, 0.35 (0.20, 0.61) for 3 points, and 0.40 (0.20, 0.80) for ≥ 4 points. Across sensitivity analyses that excluded 16 participants diagnosed within the first two years of follow-up, used multiple imputation for missing covariates, and further adjusted for the Townsend Deprivation Index, menopausal status, social support, and medication use, the association between higher lifestyle scores and lower T2DM risk remained robust [Table 4]. In addition, the inverse association between a combined healthy lifestyle and T2DM risk was consistently observed in both women with a history of GDM (HR: 0.43; 95%CI: 0.24, 0.77 for healthy vs. unhealthy) and those with a history of HDP (HR: 0.21; 95%CI: 0.07, 0.59) [Supplementary Table 6].