Introduction

Atrial fibrillation (AF) is a serious public health concern because of its increasing prevalence in the aging population. Rhythm control for AF reduces the risk of adverse cardiovascular events1,2,3, and AF catheter ablation has become an important treatment strategy in all cases4,5. However, long-term maintenance of sinus rhythm after AF catheter ablation has been unsatisfactory6,7; and those at risk of AF recurrence after catheter ablation allow for additional substrate modification or targeted ablation strategies to reduce the risk of procedure-related complications, repeat procedures, and mortality8. Therefore, novel markers that might improve risk prediction of AF recurrence after catheter ablation are of clinical significance.

Aging is the single most important contributor in AF pathogenesis9 and is associated with AF recurrence and procedure-related complications after catheter ablation10,11. Therefore, it is reasonable to expect that the ECG-derived surrogate for biological aging could have a role in recurrence after AF catheter ablations. Preliminary results on the application of artificial intelligence (AI) to 12-lead ECGs to determine biological age, distinct from chronological age, have shown excellent accuracy12. In addition, AI-estimated ECG age (AI-ECG age) has been shown to accurately predict age-related cardiovascular disease and mortality13,14,15; and the difference between AI-ECG age and chronological age (AI-ECG age gap) has shown to be greatest among those with established cardiovascular disease12,16. These findings suggest the potential of AI-ECG age gap as a novel marker that reflects the electroanatomic changes in the heart that might predict recurrence after AF catheter ablation.

In this study, we hypothesized that the AI-ECG age gap would predict recurrence after AF catheter ablation. To test this hypothesis, we validated a pre-trained residual network (ResNet)-based AI-ECG age prediction model14 on four multinational datasets and then applied the AI-ECG age prediction model on two independent AF ablation cohorts in South Korea.

Results

ECG age prediction model validation

The performance of the AI-ECG age prediction model was validated on four independent multinational datasets (Table 1 and Fig. 1). The mean absolute error (MAE) and the Pearson correlation between the predicted AI-ECG age and chronological age were as follows; MAE of 8.53 (standard deviation [SD] 7.2), 9.57 (SD 8.7), 9.96 (SD 7.6), and 8.79 (SD 7.1) and Pearson correlation of 0.83, 0.74, 0.60, and 0.53 on the CODE-15%, PhysioNet, SaMi-Trop, and UK Biobank cohort, respectively (Table 1). The average MAE of the four validation cohorts was 9.2 years and an AI-ECG age gap of ≥ 10 years was considered aged-ECG.

Table 1 AI-ECG age prediction model validation cohorts
Full size table
Fig. 1: Study flow chart.
figure 1

AI-ECG prediction model was validated in four independent multinational datasets and was then tested on two independent AF ablation cohorts. AI artificial intelligence, AF atrial fibrillation, UKB UK Biobank.

Full size image

Supplementary Fig. 1 shows a pairwise comparison of the predicted AI-ECG age and chronological age in the four validation cohorts using scatterplots and confusion matrices grouped by 10-year intervals to visually demonstrate the accuracy of the AI-ECG age prediction model. The accuracy by age interval was as follows: CODE-15%: 84.5% (95% CI, 79.5–89.4%), PhysioNet: 83.5% (95% CI, 76.6–90.5), Sami-Trop: 82.3% (95% CI, 73.4–91.1), and UK Biobank: 74.3% (95% CI, 62.6–86.0).

ECG age gap and recurrent AF after catheter ablation in the YUHS discovery cohort

In the YUHS discovery cohort, the AI-ECG age prediction model significantly overestimated chronological age (Wilcoxon test, p < 0.001). Participants were median 60 (52–67) years, 25.3% female, 63.6% had paroxysmal AF, and had a median CHA2DS2-VASc score of 1 (1-2) (Table 2). The median AI-ECG age gap in the aged-ECG group was 17.2 (13.1 to 22.9) years and that in the normal ECG age group was −0.6 (−7.5 to 4.9) years (Table 3). Participants with aged-ECG had more persistent AF, more heart failure, and larger LA diameter than the normal ECG age group. Anti-arrhythmic drug usage pre-procedure, post-procedure during the blanking period, and post-procedure after the blanking period are provided in Supplementary Table 1.

Table 2 Baseline characteristics of the study population in the YUHS discovery and KUAH evaluation cohort
Full size table
Table 3 Risk of early and clinical recurrences of AF stratified by ECG age gap in the YUHS discovery and KUAH evaluation cohort
Full size table

1724 recurrent AF cases developed over 5 years of follow-up. The 5-year cumulative event rate of AF recurrence was significantly higher in the aged-ECG compared with normal ECG age group (log rank p < 0.001; Fig. 2). After adjusting for chronological age and sex, the hazard ratio for 5-year risk of AF recurrence in the aged ECG compared with normal ECG age group was 1.44 (95% CI, 1.31–1.59). These associations remained statistically significant after accounting for clinical risk factors and echocardiographic parameters (HR 1.21; 95% CI, 1.09–1.35). Each year increase in AI-ECG age gap was associated with a 1% (95% CI 1.01–1.02) increase in the 5-year risk of recurrent AF (Table 3). The aged-ECG group was independently associated with early recurrence after AF ablation with consistency (adjusted OR 1.39; 95% CI, 1.21–1.59; Table 3). Sensitivity analysis using different cutoff thresholds for the aged-ECG versus normal ECG age showed consistent findings (Supplementary Table 2 and Fig. 2).

Fig. 2: 5-year cumulative event rate of clinical recurrence in the Aged ECG and normal ECG age group in the YUHS discovery and KUAH evaluation cohort is presented.
figure 2

Aged ECG was defined as AI-ECG age gap of ≥10 years and normal ECG age was defined as AI-ECG age gap of < 10 years. 5-year cumulative event rate of clinical recurrence stratified by AI-ECG age gap in a YUHS discovery and b KUAH evaluation cohorts. Abbreviations are the same as in Table 2.

Full size image

The higher cumulative event rate of 5-year AF recurrence in the aged ECG compared with the normal ECG age group was consistently present regardless of median chronological age or median LA diameter (Fig. 3). Consistently, there was no significant interaction between chronological age or LA diameter and the higher risk of 5-year AF recurrence in the aged ECG group (Fig. 4). In contrast, the 5-year cumulative event rate and the risk of AF recurrence were significantly higher in the aged-ECG among patients with paroxysmal AF but not those with non-paroxysmal AF (p for interaction = 0.037, Fig. 4; Supplementary Fig. 3).

Fig. 3: 5-year cumulative event rate of clinical recurrence in the Aged ECG and normal ECG age group according to chronological age and LA diameter in the YUHS discovery and KUAH evaluation cohort is presented.
figure 3

Aged ECG was defined as AI-ECG age gap of ≥10 years and normal ECG age was defined as AI-ECG age gap of <10 years. 5-year cumulative event rate of clinical recurrence stratified by AI-ECG age gap and a median chronological age or b median LA diameter in the YUHS discovery and KUAH evaluation cohorts Abbreviations are the same as in Table 2.

Full size image
Fig. 4: Subgroups analysis of the 5-year risk of clinical recurrence stratified by AI-ECG age gap in the YUHS discovery and KUAH evaluation cohorts.
figure 4

Hazard ratios for 5-year risk of clinical recurrence in the Aged ECG compared with the normal ECG age group in the YUHS discovery and KUAH evaluation cohort is presented. Hazard ratios were adjusted for AF type, age, sex, body mass index, hypertension, diabetes, vascular disease, heart failure, LA diameter, E/e’, and LV ejection fraction. CI confidence interval. Other abbreviations are the same as in Table 2.

Full size image

The C-index increased for recurrent AF when AI-ECG age gap was added to a model including multiple clinical risk factors and the APPLE score was 0.041 (95% CI 0.034 to 0.047) and 0.085 (95% CI, 0.072 to 0.098), respectively (Supplementary Table 3).

Reproducibility of the ECG age gap and recurrent AF after catheter ablation in the KUAH evaluation cohort

In the KUAH evaluation cohort, 424 recurrent AF cases developed over 5 years of follow-up. The 5-year cumulative event rate (log-rank, p < 0.001; Fig. 2), and the 5-year risk (HR 1.27; 95% CI, 1.03–1.57; Table 3) of recurrent AF was higher in the aged ECG compared with the normal ECG age group. Sensitivity analysis using different cutoff thresholds for the aged-ECG versus normal ECG age showed similar findings (Supplementary Table 2 and Fig. 2).

Consistent with the YUHS discovery cohort, there was no significant interaction with age, sex, and LA diameter (Fig. 4). However, a significant interaction with AF type for the 5-year risk of AF recurrence was identified (p for interaction 0.042). The C-index increase with the addition of AI-ECG age gap in predicting recurrent AF showed consistent findings with the YUHS discovery cohort.

AI-ECG age gap and ECG Parameters

As the AI-ECG age gap increased, a weak positive correlation with heart rate (r = 0.29), PR interval (r = 0.19), and QTc interval (r = 0.26, Fig. 5a) was identified. AI-ECG age showed similar weak positive correlations with conventional ECG parameters (Fig. 5b) but not for chronological age (Fig. 5c). Participants with aged-ECG were associated with significantly higher heart rates (p < 0.001), longer PR interval (p < 0.001), and longer QTc interval (p < 0.001) compared with normal ECG age group (Fig. 5d). ECG signatures captured by the Grad-CAM were not recognizable by cardiologists (Supplementary Fig. 4).

Fig. 5: Correlation between conventional ECG parameters and (a) AI-ECG age gap, (b) AI-ECG age, (c) chronological age, and (d) AI-ECG age gap group.
figure 5

bpm beats per minute, ms milliseconds. Other abbreviations are the same as in Table 2.

Full size image

Discussion

In this study, we validated a ResNet-based AI-ECG age prediction model on four independent multinational datasets (total ECG no. = 414,804) and then tested the model on two independent AF catheter ablation cohorts to examine whether the AI-ECG age gap can predict recurrent AF after catheter ablation. We showed that the gap between pre-procedural AI-ECG age and chronological age predicts recurrent AF after catheter ablation. The predictive performance of the ECG age gap was robust across geographically distant, independent AF ablation cohorts and preserved across different chronological ages and LA diameters. These findings suggest that AI-ECG age gap might be used as an interpretable, simple risk marker for recurrence after AF catheter ablation.

A single raw ECG tracing is a natural summary of the electroanatomic changes accumulated in the heart14, and several lines of evidence suggested that AI-ECG is useful in identifying hidden disease states from a single 12-lead ECG14,17,18,19,20,21,22,23,24,25,26,27. For example, AI-ECG has been reported to detect LV systolic dysfunction27, valvular heart disease25, coronary artery disease28, hypertrophic cardiomyopathy26, and even anemia22. AI-ECG is now expanding its role from not only disease prediction but also disease management18. Examples include predicting those at risk of sudden cardiac death for potential implantation of an implantable cardioverter defibrillator29 or predicting those who would respond to cardiac resynchronization therapy30.

In AF, AI-ECG has been reported to accurately predict AF from a single sinus rhythm ECG19 and subsequent studies demonstrated predictions comparable to that of the CHARGE-AF model17. In a recent non-randomized interventional trial, a higher detection rate of AF was reported using the AI-guided approach31. Beyond AF prediction, AI-ECG also predicted AF-related strokes21, and the potential application of AI-ECG for targeted AF screening to prevent AF-related adverse events has become an active area of investigation.

To our knowledge, no study has assessed the potential application of AI-ECG age in AF ablation. In this study, we show that evaluating a single pre-procedural sinus rhythm ECG could be used as a simple risk marker that could be easily adopted in clinical practice for risk prediction of recurrence after AF catheter ablation. Of note, the applicability of AI-ECG age was extended to Asian race/ethnicity given that the AI-ECG age prediction model was predominantly developed and tested on Caucasians in previous studies, which has been pointed out as a major challenge in a recent review article regarding the clinical application of AI-ECG age32.

AI-ECG age was capable of capturing at least or more than what has traditionally been suggested given the significant association with recurrence after accounting for multiple clinical risk factors and echocardiographic parameters. In addition, when AI-ECG age gap was added to the APPLE score, an established clinical risk score for recurrence, the increase in discrimination was statistically significant. Although summarizing ECG changes into a single parameter might be an oversimplification, disease and related outcome prediction through AI-ECG age is useful because AI-ECG age transmits the idea of the complex mechanism of AF recurrence into a simple language that does not require medical expertise and is easily understood by the patients for shared decision making14.

Interestingly, the AI-ECG age gap predicted recurrent AF more significantly among patients with paroxysmal AF compared to those with persistent AF. The reason is unclear but AF is a progressive disease and patients with persistent AF include a wide range of AF progression and structural remodeling stages5, in which substrate heterogeneity might have weakened the association with AI-ECG age gap. In addition, AI-ECG age prediction model might suggest advanced heart age in patients with persistent AF regardless given the associated advanced electroanatomic changes of the heart including lower LA voltage, slowed conduction, and complex fractionated signals33. Conversely, AI-ECG age prediction model might be more useful in capturing subtle ECG abnormalities with greater discrimination among patients with less advanced electroanatomic changes, that is, paroxysmal AF. The AI-ECG age in this study also predicted early recurrences after AF catheter ablation, and might potentially complement the clinical decision whether to maintain antiarrhythmic drugs in immediate post-ablation periods34.

Despite a weak correlation between chronological age and conventional ECG parameters, the AI-ECG age prediction model showed remarkable performance in predicting chronological age. This suggests that AI-ECG age prediction model is leveraging complex ECG patterns (capturing P, QRS, and T segments across the board) rather than relying on specific ECG parameters predominant at certain ages. This complexity of AI-ECG age prediction is akin to the intuitive acceptance that when predicting age using portrait photos, a person perceives specific facial regions (e.g. eyes, mouth wrinkles, forehead) simultaneously combined and integrated for both young and elderly. Attia et al.12 also argued that the explainability of the AI-ECG model would be limited because it does not rely on a single measure or simple score metric but rather involves multiple levels of nonlinear interactions, possibly including subtle variations beyond human intuition or conventional ECG features.

Several future applications of AI-ECG in AF ablation can be considered. First, because AF ablation strategy other than pulmonary vein isolation is uncertain5, a more extensive ablation lesion set among those with advanced electroanatomic changes detected by the AI-ECG age can be considered. Second, a potential application in the maintenance of anti-arrhythmic drugs after AF ablation in patients with advanced electroanatomic changes detected by the AI-ECG age is possible. In particular, the AF guidelines weakly recommend short-term maintenance of anti-arrhythmic drugs after AF ablation to prevent early recurrences5, patients with advanced electroanatomic changes detected by AI-ECG age could benefit from maintaining extended periods of arti-arrhythmic drugs. These potential future directions of the AI-ECG age in AF ablation require further investigation.

This study has several limitations. First, not all patients with AF planned for catheter ablation may benefit from pre-procedural AI-ECG age prediction because AI-ECG age in this study was predicted from pre-procedural sinus rhythm ECGs. Second, extra-PV ablations other than circumferential PVI were left to each operator’s discretion and might have influenced the results of this study. Third, the study participants analyzed were included from large-volume tertiary medical centers in South Korea, and might not represent the general AF patients who are planned for AF catheter ablation. Fourth, participants who had no pre-procedural sinus rhythm ECGs were excluded from the analysis. This requires attention given the possibility of a selection bias that might have influenced the results of this study. In this respect, pre-procedural AF rhythm ECGs to predict recurrence after AF catheter ablation might be an interesting area of future research. Fifth, the complexity of the end-to-end DNN model applied in this study makes it hard to interpret the mechanism of our results. Sixth, future randomized clinical trials are needed before the clinical implementation of AI-ECG age gap in risk prediction of recurrence after AF catheter ablations. Seventh, future research on the utility of AI-ECG age gap in predicting recurrent AF after catheter ablation across different geographical locations and demographics might be useful given that the discovery and evaluation of AF ablation cohorts in this study were from South Korea.

Although a single 12-lead ECG in patients with AF is part of the routine evaluation in every outpatient clinic visit, its role is mostly limited to identifying the rhythm status of the patients with little prognostic information on their future rhythm control outcomes. We believe that our study is a further step towards the practical application of AI-ECG in AF ablation. Overall, a pre-procedural ECG age predicts recurrent AF after catheter ablation independent of clinical risk factors and echocardiographic parameters.

Methods

ResNet-based ECG age prediction model

A pre-trained residual network (ResNet)-based model developed by Lima et al.14 was used for ECG age prediction. The ResNet-based model architecture included residual blocks composed of 1D convolutional layers, each followed by batch normalization, rectified linear unit (ReLU) activation, and dropout layers (Supplementary Fig. 5). The performance of the ResNet for detecting ECG abnormalities and ECG-based age prediction has demonstrated in previous studies14,35,36. We then validated the ResNet-based model for ECG age prediction on four independent multinational datasets that were unseen during pretraining. The four model validation cohorts consisted of a total of 414,804 ECGs that included the CODE-15% (n = 345,779), PhysioNet (n = 21,799), Sami-Trop (n = 1631), and UK Biobank (n = 45,595) cohorts. Detailed information on each of the validation cohorts is provided elsewhere37,38,39,40.

Pre-processing for 12-Lead ECG

We pre-processed 12-lead ECG tracings to create a uniform dataset for input into the pre-trained ResNet-based AI model. Each ECG tracing was standardized to a fixed length of 4096 samples to accommodate the model’s input requirement.

  • Standardization of sample length: For ECG tracings originally lasting 10 seconds and sampled at 500 Hz (resulting in 5000 samples), we resampled these to 400 Hz, reducing the count to 4000 samples. We then applied zero-padding to the tracings, prefixed and suffixed with 48 zeros on both ends, to expand them to 4096 samples.

  • ECG lead order: Each ECG dataset was structured to include tracings across 12 standard leads in the following order: {DI, DII, DIII, AVR, AVL, AVF, V1, V2, V3, V4, V5, V6

The pre-processing steps ensured consistent compatibility of each ECG tracing with the input specifications of the AI model.

ResNet model architecture

The ResNet-based model architecture released and developed by Lima et al. 14. incorporates residual blocks composed of 1D convolutional layers, each followed by batch normalization, ReLU activation, and dropout layers to prevent overfitting (Supplementary Fig. 5). The architecture is described as follows:

  • Initial Convolution and Residual Block: The input undergoes a 1D convolution with 64 filters, followed by batch normalization. A subsequent residual block includes two sequential 64-filter convolutional layers, each followed by batch normalization, ReLU activation, and dropout.

  • Down-sampling and Expanding Filters: A max pooling layer reduces dimensionality before a series of deeper layers with increasing filters (128, 196, and 256), each paired with a residual block and max pooling.

  • Final Stages and Output: The network expands to 320 filters in the final stages, each including a convolutional layer with batch normalization, ReLU, and dropout. The output is processed through a linear layer that generates a single value per input.

The total number of trainable parameters in the model is 6,924,705. The estimated memory requirement for running this model is approximately 11.12 GB, which includes 0.19 MB for input size, 11.10 GB for forward/backward pass size, and 26.42 MB for parameter storage. This indicates that substantial memory resources are necessary to deploy this pre-trained model effectively.

Study population

After validation, the ResNet-based AI-ECG age prediction model was tested on two independent AF catheter ablation cohorts from the Yonsei University Health System (YUHS) and Korea University Anam Hospital (KUAH). The YUHS and KUAH are geographically distant, tertiary medical centers in South Korea. The YUHS discovery and KUAH evaluation cohorts are retrospective cohorts that included patients diagnosed with AF who provided baseline information, medical history, physical measures, lifestyle factors, and biological samples and underwent AF catheter ablation with detailed pre- and post-procedural follow-ups. AI-ECG age was predicted from pre-procedural 12-lead sinus rhythm ECGs sampled at 500 Hz and a length of ten seconds. The exclusion criteria of this study were as follows; 1) patients who underwent prior AF catheter ablation, 2) patients with a history of mitral valve surgery or mitral stenosis, and 3) patients without a pre-procedural sinus rhythm ECG (e.g. patients with persistent AF who did not underwent cardioversion or failed to restore sinus rhythm after cardioversion before the index procedure). Finally, a total of 5,466 and 1,564 patients were analyzed from the YUHS discovery and KUAH evaluation cohorts, respectively.

All participants provided written informed consent. The study protocol adhered to the principles of the Declaration of Helsinki and was approved by the Institutional Review Board of the Yonsei University Health System.

AI-ECG age gap

AI-ECG age gap was defined as follows:

$${AI}{\rm{\mbox{-}}}{ECG\; age\; gap}={AI}{\rm{\mbox{-}}}{ECG\; age}-{chronological\; age}$$
(1)

Study participants in the YUHS discovery and KUAH evaluation cohorts were categorized into two groups as follows:

$${Aged\; ECG}={AI}{\rm{\mbox{-}}}{ECG\; age\; gap}\ge 10{years}$$
(2)
$${Normal\; ECG\; age}={AI}{{\mbox{-}}}{ECG\; age\; gap} \,<\, 10{years}$$
(3)

where the cut-off of 10 years was based on the average of the mean absolute error (MAE) of the AI-ECG age calculated from the AI model validation cohorts (CODE-15%, PhysioNet, Sami-Trop, and UK Biobank) (Table 1).

Echocardiographic evaluation

All participants underwent pre-procedural transthoracic echocardiography. The left atrium (LA) diameter, peak trans-mitral inflow velocity to tissue Doppler echocardiography of the peak septal mitral annular velocity (E/e’), and left ventricular (LV) ejection fraction were assessed in accordance with the American Society of Echocardiography guidelines41.

Electrophysiological mapping and atrial fibrillation catheter ablation (AFCA)

AFCA protocols were largely consistent in the YUHS discovery and KUAH evaluation cohorts. Prucka CardioLab Electrophysiology system (General Electric Medical Systems, Inc, Milwaukee, WI) was used for intracardiac electrogram recordings. After a trans-septal puncture, a multi-view pulmonary venogram was obtained; 350 to 400 seconds of activated clotting time was maintained via intravenous heparin injection. A three-dimensional (3D) electro-anatomical mapping system (NavX; Abbott or CARTO; Biosense Webster) merged with 3D spiral computed tomography was used to guide the AFCA in all patients. The luminal esophageal temperature was monitored and maintained below 38.4 °C for the radiofrequency-PVI and 15 °C for the Cryo-PVI. For the radiofrequency-PVI, we used an open irrigated, 3.5 mm tip deflectable catheter (Celsius and Smart-Touch; Johnson & Johnson, Inc., Diamond Bar, CA; Coolflex and FlexAbility, Abbott, Inc., Minnetonka, MN). RF power of 30–35 W and 20–25 W was applied for the anterior and posterior area of the LA, respectively. Cavotricuspid isthmus (CTI) ablation was performed in most of the patients except for those with AV conduction disease. All patients underwent CPVI that encircled the right and left PVs.

Post-ablation management and follow-up

Patients were regularly followed up in the outpatient clinic at one, three, six, and 12 months and every six months thereafter or whenever patients experienced symptoms after the AF catheter ablation. We obtained an ECG in all patients at every outpatient visit and 24-h Holter recordings at three and six months and every six months thereafter according to the 2017 Heart Rhythm Society/European Heart Rhythm Association/European Cardiac Arrhythmia Society Expert Consensus Statement guidelines42. We examined patients who experienced symptoms of palpitations representing an arrhythmia recurrence using Holter monitor or event monitor recordings. A researcher whose assignment was independent of the study group conducted the Holter analysis and adjudication. AF recurrence was defined as any episode of AF or atrial tachycardia (AT) of at least 30 seconds. Any ECG documentation of AF recurrence less than three months after the procedure was considered an early recurrence, and an AF recurrence more than three months after the procedure was considered a clinical recurrence. The post-ablation management and follow-up were consistent in the YUHS discovery cohort and KUAH evaluation cohort.

ECG pattern recognition of the AI-ECG age prediction model

Gradient-Weighted Class Activation Mapping (Grad-CAM)43 was used to explore the ECG signatures that AI-ECG age prediction model captures in predicting chronological age. In addition, we calculated the Pearson correlation coefficients between conventional ECG parameters (heart rate, PR interval, QRS duration, and QTc interval) and the predicted AI-ECG age as well as the AI-ECG age gap.

Statistical analysis

Categorical variables are reported as numbers (percentages) and compared using the Chi-square or Fisher’s exact test. Continuous variables are reported as medians (interquartile range) and compared using the Kruskal-Wallis test. Event rates were calculated as the number of events divided by the total follow-up time per 100 person-years.

The performance of the AI-ECG age gap for prediction of AF recurrence after catheter ablation was assessed using multiple analytical methods. First, the hazard ratios and the associated 95% confidence intervals were calculated using Cox proportional hazard models. The hazard ratios were calculated for the aged-ECG compared with the normal ECG age group as well as using AI-ECG age gap as a continuous variable (per 1 AI-ECG age gap increase). The proportional hazard assumption was not violated as determined by Schoenfeld residual plots. Second, the 5-year cumulative incidence curves for AF recurrence after catheter ablation of the aged-ECG and normal ECG age groups were assessed. Third, we carried out multiple subgroup analyses according to median chronological age, sex, median LA diameter, and AF type (paroxysmal vs non-paroxysmal) to assess whether the AI-ECG age gap predicts AF recurrence after catheter ablation differently for different patient characteristics. Fourth, the odds ratios and the associated 95% confidence intervals were assessed to examine whether the AI-ECG age gap predicted early recurrences (AF recurrence within 3 months of the procedure). Fifth, Harrell’s C-index for AF recurrence when AI-ECG age gap was added to 1) a model including multiple clinical risk factors and 2) the APPLE score44, an established risk score for prediction of AF recurrence, was calculated. Bootstrapping was used to estimate the confidence intervals of C-indices and their differences among different models. Sixth, we examined the hazard ratios and its associated 95% confidence intervals as well as the 5-year cumulative incidence curves for AF recurrence after catheter ablation using different cut-off thresholds for the aged ECG versus normal ECG age as a sensitivity analysis.

All analyses were performed using R statistics, version 4.0.2 software (R Foundation for Statistical Computing); and a two-sided p-value < 0.05 was considered statistically significant.