Our study population is comprised of more than half a million volunteers aged 40–69 years who enrolled in the UK Biobank between the years 2006 and 2010 (14). These participants lived within 25 miles of one of the 22 assessment centers located throughout England, Wales, and Scotland. On enrollment in the UK Biobank, participants provided signed consent and answered questions about sociodemographic, lifestyle, environmental, and health-related factors. 12-lead (at-rest) ECGs were performed at an imaging assessment center for UK Biobank. Exact details of the procedure can be found at https://biobank.ctsu.ox.ac.uk/crystal/refer.cgi?id=510. ECG acquisition was approximately 10 s per individual. Participants who had an ECG done during their clinic visit (2014 or later) or subsequent clinic visit (2019 or later) were included in our study. For participants with two ECGs, only the first valid ECG was accounted. We excluded individuals with prevalent stroke (n = 808), which left 67,757 individuals for the association analysis of ECG-age and incident stroke.

All the lifestyle and clinical risk factors of stroke were measured at the same time as the ECG measurement. Lifestyle factors (smoking, alcohol consumption, physical activity, and diet) were self-reported and were dichotomized as previous publications presented (15). For blood pressure, two measures were taken one minute apart using a digital blood pressure monitor. Body mass index was calculated as weight in kilograms divided by height in meters squared. Particulate matter (PM) air pollution (PM_2.5, unit micro-g/m3) was collected from local environment air pollution data in the year 2010 (Data Field 24,006). Another air pollutant, PM₁₀ (PM with an aerodynamic diameter of <10 micro-g), was measured in 2007 (Data Field 24,019) and 2010 (Data Field 24,005). Health-related outcomes including diabetes mellitus (DM), chronic kidney diseases (CKD), and dyslipidemia were followed up based on the primary care data (READ code), hospital inpatient data (ICD-9/ICD-10 codes), death registry (ICD-10 code), and self-reported medical conditions. Their first occurrence dates were derived and provided by the UK Biobank: DM included type 1, type 2, and other diabetes (Data Fields 130,706, 130,708, 130,710, 130,712, 130,714), CKD Data Field was 132,032, and dyslipidemia was 130,814.

Ischemic stroke and hemorrhagic (intracerebral and subarachnoid subtypes) stroke were the outcomes of interest. These outcomes were defined by self-reported medical conditions, ICD-9, or ICD-10 codes (16, 17). The Data Field related to the stroke outcomes were 42,006, 42,008, 42,010, and 42,012. The included participants were followed up since they took the first ECG until whichever came first: newly diagnosed stroke, or their last follow-up time in the UK Biobank, in which they were free of stroke, death, or end of August 2023.

The ECG-age was obtained using a DNN that applies the raw ECG waveform in an end-to-end approach as opposed to classical automatic analysis methods (18). The model was trained to predict a person's age by extracting features directly from the data instead of relying on traditional ECG interpretation (18, 19). Through this learning process, the DNN captured how aging affects ECG waveform (11). For this study, the raw ECG signals used to generate the ECG-age model were derived from the CODE study, which was further refined using the UK Biobank ECG data. The CODE study is part of the Telehealth Network of Minas Gerais (TNMG) with a database containing ECGs obtained from 2010 to 2017 in Brazilian primary care settings (20). To date, the CODE data set has been recognized as the world's largest ECG database used to develop artificial intelligence (AI)-ECG applications with 1,558, 415 participants (21). The ECG-age model in the CODE study and its development have been previously described (12). The model uses a DNN to make predictions like the residual network proposed for image classification except with unidimensional signals (11). More information on the code used in the DNN-estimated ECG-age model training, evaluation, and statistical analysis can be found at https://github.com/antonior92/ecg-age-prediction.

To assess the DNN-estimated ECG-age as a predictor for stroke risk, we used the Δage (ECG-age minus the chronological age) as the independent variable and thereby capturing the excess risk caused by a greater decline in cardiovascular health than expected by chronological aging. We also separated participants into three aging categories based on Δage mean absolute error (MAE) (12). Those with an ECG-age older than the chronological age by ≥MAE years were considered accelerated aging. Those with an ECG-age younger than the chronological age by ≥MAE years were considered decelerated aging. Those with an ECG-age within the MAE were considered normal aging. We used Cox proportional hazards regression models to assess the association of Δage and incident stroke: the first model was chronological age and sex adjusted, and the second multivariable model was additionally adjusted for SBP, BMI, particulate matter air pollution, smoking status, diet, physical activity level, alcohol consumption, diabetes, CKD, and dyslipidemia. The results were expressed as stroke risk of Δage per 10 years.

For the secondary analysis, we assessed the association of Δage with prevalent stroke by logistic regression models and adjusted for chronological age, sex, and additional clinical risk factors. Several sensitivity analyses included: (A) we excluded participants with prevalent atrial fibrillation at the time of ECG measurement, (B) we assessed the 5-year stroke risk by the Cox model, (C) we assessed the ability of Δage to predict incident stroke using the C-statistics where we calculated the change in the predictive capacity with and without the Δage in the multivariable Cox models, and the net reclassification index was also generated to assess the prediction ability (22), (D) we explored the non-linear relationship of Δage and the hazard of incident stroke by using the fractional polynomials and splines.

In our sensitivity analysis, we performed stratified analysis in two age groups (chronological age <60, or ≥ 60), and in men and women separately to assess the associations of Δage and incident stroke. We used the multivariable-adjusted Kaplan–Meier curves which were achieved by the inverse probability weighting method to illustrate the cumulative risk of stroke across decelerated, normal, and accelerated aging groups. We used P values less than 0.05 (two-sided) as the statistical significance level. All the analyses were performed using R software package version 4.2.1 (https://www.r-project.org/).

Table 2 shows the association of Δage with incident stroke in the chronological age- and sex-adjusted, and multivariable-adjusted models. Every 10-year increase in Δage was associated with a 22% increase in the risk of incident stroke [HR = 1.22 (95% CI, 1.00–1.49)] after adjusting for multiple risk factors. The association remained significant if we restricted to 5 years risk of incident stroke in both age- and sex- adjusted and multivariable-adjusted models (Supplementary Table S2). The association of Δage with incident stroke in the multivariable-adjusted model was attenuated when we excluded prevalent atrial fibrillation cases (Supplementary Table S3). However, the inclusion of Δage did not significantly improve the prediction of incident stroke with a change of C-statistics of 0.003 [95% CI, −0.003–0.009] in the multivariable model. There was also no significant change in terms of net reclassification improvement (NRI) (P > 0.05).

Figure 1 shows the multivariable-adjusted cumulative risk of incident stroke across three aging groups. Participants in the accelerated aging group had a higher probability of developing stroke compared to the normal and decelerated aging groups (log-rank test P value = 0.02). We also fit a smoothing spline fit for the relative hazard of incident stroke as a function of Δage (Supplementary Figure S1). The graph did not support a nonlinear relationship between Δage and hazard of incident stroke (P = 0.07 for the nonlinear term).

Two additional analyses on comparisons of the stroke risk models based on PCE with or without ECG Δage and revised FSRS (R-FSRS) with or without ECG Δage were conducted. Although there was subtle improvement in the C statistics when Δage was added to both models, these findings were not statistically significant (Supplementary Table S8).

In the age and sex-stratified analysis, Δage was positively associated with stroke incidence in individuals older than 60 years and in men (HR = 1.31 [95% CI, 1.05–1.63] and 1.34 [95% CI, 1.03–1.74] respectively). However, the difference did not reach the pre-defined significance threshold (both with interaction P values greater than 0.05). Similar association results were seen in the incidence of ischemic stroke (Supplementary Table S4). Because we had very limited incident hemorrhagic stroke cases, we did not find any association of Δage and hemorrhagic stroke (Supplementary Table S1). Moreover, in Supplementary Table S5, accelerated aging was associated with a 42% increase in incident stroke [HR = 1.42 (95% CI, 1.12–1.80)] in the age- and sex-adjusted model. Decelerated aging was associated with a 10% decrease in incident stroke, but this finding was not statistically significant.

As shown in Supplementary Table S6, participants with prevalent stroke were more likely to be male and have higher Δage, BMI, and systolic blood pressure. Like their incident stroke counterparts, they were more likely to have prevalent DM, CKD, and dyslipidemia. Supplementary Table S7 shows the association between Δage and prevalent stroke in the age- and sex-adjusted, and multivariable-adjusted models. There was a significant association between prevalent stroke and Δage [OR, 1.28 (95% CI, 1.11–1.49)], suggesting that those with a history of stroke were 28% more likely to exhibit accelerated aging.