The Department of Anesthesiology at Severance Hospital in Seoul, South Korea, has established a cardiac surgery registry that compiles extensive perioperative clinical data. The hospital’s institutional review board approved this study (approval numbers: 4-2022-1506 and 4-2024-1515), waived the requirement for informed consent because only anonymized data were used retrospectively, and limited data usage to the approved timeframe. All data were de-identified before analysis in accordance with institutional anonymization protocols, and no personally identifiable information was accessible to the investigators. As this study analyzed retrospective anonymized registry data without direct participant involvement, no financial or other compensation was provided to participants. No images or supplementary materials contain identifiable individuals.

Adult patients (age ≥19 y) who underwent cardiac surgery at Severance Hospital between December 2018 and December 2023 were included (Figure 1). For patients who underwent multiple surgeries during this period, only the first surgery was included. Preoperative clinical data reflecting patient status, including demographics, medical history, laboratory results, physiological measurements, medications, and type of surgery were collected from the institutional cardiac surgery registry (Table S1 in Multimedia Appendix 1). Variables with missing rates exceeding 5% in the entire cohort were excluded, and the remaining missing data were handled using mean imputation.

Standard 10-second 12-lead ECGs of these patients were also extracted. AF ECGs were defined as those with an automatic interpretation indicating atrial fibrillation or atrial flutter. We excluded patients who had an AF ECG recorded before surgery, those with a known history of AF, or those undergoing the MAZE procedure (a surgical treatment for AF). For the remaining patients, only ECGs taken within 30 days before the surgery were included in the analysis. ECGs with automatic interpretations indicating paced rhythms, poor quality, or lead reversals were also excluded. While there is no consensus on the precise definition of postoperative AF [1,16,17], we defined it as AF documented by ECG within 30 days after surgery, with this 30-day cutoff being widely used across multiple studies. Patients with first AF observed after this 30-day postoperative period were excluded. A sensitivity analysis was conducted using a 7-day definition for postoperative AF.

Our primary objective was to evaluate the AI-ECG-AF model from a clinical perspective, focusing on its utility as a novel predictive biomarker for postoperative AF in patients undergoing cardiac surgery. Accordingly, we developed the AI-ECG-AF model using a methodology similar to that of a previous study [13], with some modifications, and concentrated our efforts on its clinical evaluation. The details of the model development are described below.

The standard 12-lead ECG database from Severance Hospital contained 5.1 million ECGs from 1.2 million patients from 1993 to 2022 (Figure 1). The ECGs were utilized in their original form as raw 1D waveform time-series data. All non-AF ECGs were used for modeling. We classified patients into two groups: patients positive for AF, who had at least one AF ECG recorded in the database, and patients negative for AF, who had no AF ECG recorded in the database and had no diagnostic codes for AF in their electronic medical record. Patients who had an AF diagnosis code but lacked ECG confirmation were excluded from the analysis to avoid ambiguity (referred to as unverified AF). ECGs with automatic interpretations indicating paced rhythms, poor quality, or lead reversals were excluded. ECGs from patients included in the cardiac surgery cohort were also excluded to prevent data leakage.

For patients positive for AF, we used non-AF ECGs recorded starting from 31 days before their first AF ECG and onward. Structural changes associated with AF would be present before the first recorded AF episode, so while we could use ECGs from before the first AF ECG, we chose a relatively short time interval as a conservative measure to avoid using ECGs before any structural changes developed, as done in the previous study [13]. For patients negative for AF, all their non-AF ECGs were used for modeling.

We used the 1D EfficientNet-B0 architecture (Figure S1 in Multimedia Appendix 1) [18]. Given the extensive size of the dataset, we utilized only the middle 5 seconds of the 10-second ECG (ie, from 2.5 to 7.5 s). The ECGs, originally at 250 or 500 Hz, were standardized to 250 Hz by downsampling the 500 Hz. Each lead was z-normalized (mean=0, standard deviation=1). No additional signal filtering was applied. Since four of the six limb leads can be linearly derived from any two leads using the Einthoven law and the Goldberger equation [19,20], we used eight leads (I, II, V1-V6) as input. We randomly split the dataset at a 6:2:2 ratio (stratified) into training, validation, and test sets while ensuring no individual overlap. During each training epoch, we balanced classes by randomly oversampling the minority class and undersampling the majority class to equal sizes while maintaining the original dataset size. No such resampling was applied to the validation or test sets, which were kept in their original distributions. Hyperparameter optimization was performed through a grid search based on learning rate, batch size, and weight decay. We used the Adam optimizer, and a learning rate of 0.003, batch size of 512, and weight decay of 0 were ultimately selected. A learning rate halving strategy was applied, reducing the learning rate by half if the validation loss from the validation set did not improve for 5 consecutive epochs. The model was trained on the training set, with early stopping applied if the validation loss from the validation set did not improve for up to 15 consecutive epochs. The final evaluation was conducted on the test set. The model produces an output score ranging from 0 to 1, where higher scores indicate a greater likelihood of AF as predicted by the model.

Our first objective was to verify whether the AI-ECG-AF model score is an independent risk factor for postoperative AF in cardiac surgery patients. We applied the AI-ECG-AF model to the ECGs from our cardiac surgery cohort to obtain prediction outputs. We performed multivariable logistic regression with postoperative AF as the dependent variable and the AI-ECG-AF model score along with clinical variables as independent variables. To ensure the robustness of results, we conducted the logistic regression analyses with two sets of clinical variables.

In the first logistic regression, we included both the AI-ECG-AF model score and clinical variables from the POAF score [7], an existing postoperative AF prediction tool, which comprised age category (<60, ≥60 and <70, ≥70 and <80, and ≥80), chronic obstructive pulmonary disease (COPD), estimated glomerular filtration rate <15 mL/min, emergency surgery, preoperative intra-aortic balloon pump/extracorporeal membrane oxygenation/ventricular assist device, left ventricular ejection fraction <30%, and valve surgery as independent variables.

In the second logistic regression, we utilized all preoperative clinical data from the institutional cardiac surgery registry (Table S1 in Multimedia Appendix 1) and performed stepwise logistic regression with backward elimination, incorporating the AI-ECG-AF model score alongside clinical variables as independent variables. The resulting postoperative AF prediction score was termed the AI-enhanced post-cardiac-operative AF (AI-COAF) score.

Subgroup analyses by sex, age (<65 and ≥65 y), and type of surgery (valve and nonvalve) were additionally performed for both logistic regression analyses. For the second logistic regression, the subgroup analyses utilized the predetermined dependent variables identified during the backward selection process to ensure consistency.

Our second objective was to evaluate the additive or synergistic predictive value of the AI-ECG-AF model score when combined with other risk factors or the existing POAF score [7]. We compared the AI-ECG-AF’s predictive ability (via the area under the receiver operating characteristic curve [AUROC] and average precision) with the POAF score and tested if adding the AI-ECG-AF score to the POAF score (AI-POAF [artificial-intelligence–augmented postoperative atrial fibrillation] score=POAF score+α×AI-ECG-AF score) improved predictions. We found the optimal weighting factor (α=4) using data from December 2018 to December 2021 (derivation dataset). All model comparisons used data from January 2022 to December 2023 (comparison dataset). We also retrained the AI-POAF score using the predetermined dependent variables on the derivation dataset and compared its performance with other models on the comparison dataset. Performance metrics, including accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and the F₁-score, were calculated in the comparison dataset using the threshold that maximized the Youden J index in the derivation dataset.

Normality of continuous variables was assessed with the Shapiro-Wilk test. Since no continuous variables were normally distributed, Mann-Whitney U tests were used for comparisons. Categorical variables were compared using the χ² test. The determination of 95% CI of AUROC and AUROC comparisons used the Delong’s method (paired and two-sided) [21]. Statistical significance was set at P<.05.

Model training and evaluation were performed in Python (version 3.10.6) using Pytorch (version 2.0.1) and Scikit-learn (version 1.3.0), while statistical analyses were conducted with the TableOne package (version 0.8.0). Regression analyses were performed in R (version 4.0.4).

For AI-ECG-AF development, 4.05 million non-AF ECGs from 1.13 million patients were used for modeling (Figure 1). Within this dataset, a subset of 0.61 million ECGs, corresponding to 0.04 million patients, was labeled as positive for AF, while the remaining 3.45 million ECGs, derived from 1.09 million patients, were categorized as negative for AF. Among all the ECGs included in the AI-ECG-AF development process, the average age at the time of measurement was 60.3 years, with a standard deviation of 10.3 years. Additionally, 55.3% of the ECGs in the dataset were recorded from male patients. In terms of performance, the AI-ECG-AF model achieved an AUROC of 0.901, with a 95% CI ranging from 0.900 to 0.902, for identifying the electrocardiographic signature of AF in non-AF ECGs in the test set of the AI-ECG-AF development cohort (Figure 2). The fully developed AI-ECG-AF model, along with comprehensive instructions detailing its usage, has been made entirely accessible to the public through our dedicated online repository [22].

A total of 5204 ECGs obtained from a cohort of 2266 patients who had undergone cardiac surgery were included in the final analysis (Table 1 and Figure 1). Patients with postoperative AF (1815 ECGs from 763 patients) had a lower proportion of males (66.7% [1210/1815] vs 70.5% [2390/3389], P=.005), were older (≥80 y: 8.8% [159/1815] vs 3.2% [109/3389]; 70‐79 y: 40.1% [727/1815] vs 27.1% [919/3389]; P<.001), had a higher prevalence of COPD (7.6% [138/1815] vs 3.3% [112/3389], P<.001), a higher rate of valve surgery (47.7% [865/1815] vs 41.5% [1405/3389], P<.001), and higher POAF score (median [IQR]: 2 [1-3] vs 2 [1-2], P<.001), and AI-ECG-AF model scores (median [IQR]: 0.553 [0.313‐0.763] vs 0.346 [0.154‐0.601], P<.001) compared to those who were negative for postoperative AF (3389 ECGs from 1503 patients). The remaining preoperative clinical characteristics from the institutional cardiac surgery registry are summarized in Table S2 in Multimedia Appendix 1. Figure S2 in Multimedia Appendix 1 provides a visual representation of the incidence of POAF across different postoperative days, showing that the highest incidence occurred between postoperative days 1 and 4.

Figure 3 shows the results of the logistic regression analyses. To facilitate a more intuitive interpretation, the AI-ECG-AF model scores, originally on a scale from 0 to 1, were rescaled to a range of 0 to 10 by multiplying by ten. This adjustment allows the odds ratios displayed in Figure 3 to represent the change in odds associated with a 10% absolute increase in the AI-ECG-AF model score after accounting for other relevant clinical variables.

After adjusting for the variables included in the POAF score, a 10% absolute increase in the AI-ECG-AF model score was found to be associated with a 19.7% rise in the odds of developing postoperative AF (95% CI 16.9%‐22.6%). The AI-ECG-AF model score was identified as an independent risk factor for postoperative AF, not only when analyzing the entire patient cohort but also consistently across all demographic and surgical subgroups. A comprehensive list of odds ratios for all analyzed variables is provided in Table S3 in Multimedia Appendix 1.

Similarly, after adjusting for variables extracted from the institutional cardiac surgery registry, which were selected using stepwise logistic regression with backward elimination, a 10% absolute increase in the AI-ECG-AF model score was associated with a 20.9% rise in the odds of developing postoperative AF (95% CI 17.8%‐24.2%). The AI-ECG-AF model score remained an independent risk factor for postoperative AF in the entire cohort as well as across all demographic and surgical subgroups. The odds ratios for all variables are provided in Table S4 in Multimedia Appendix 1.

To further assess the robustness of these findings, a sensitivity analysis was performed using a more stringent 7-day definition for postoperative AF. In this analysis, an additional 188 ECGs from 72 patients, who had their first recorded AF episode occurring beyond the initial 7-day postoperative period but within 30 days, were excluded. The AI-ECG-AF model score remained a significant independent risk factor for postoperative AF across both regression models, with adjusted odds ratios for AI-ECG-AF model score×10 of 1.200 (95% CI 1.171‐1.231) and 1.221 (95% CI 1.188‐1.254), respectively (Figure S3 in Multimedia Appendix 1). Furthermore, the association remained consistent across all demographic and surgical subgroups.

Figure 4 shows the ROC and precision–recall curves of the models for predicting postoperative AF. The AUROCs were 0.643, 0.644, 0.680, and 0.710 for the POAF score, AI-ECG-AF model score, AI-POAF score, and AI-COAF score, respectively. According to the Delong test, both the AI-POAF score and AI-COAF score demonstrated significantly higher AUROCs than the POAF score (both P<.001), whereas there was no statistically significant difference between the POAF score and the AI-ECG-AF model score (P=.927). The respective average precisions were 0.429, 0.430, 0.489, and 0.513. Table 2 shows the performance at the threshold determined at maximum Youden J index.

For the sensitivity analysis using a 7-day definition for postoperative AF, the AUROCs were 0.647, 0.641, 0.678, and 0.706 for the POAF score, AI-ECG-AF model score, AI-POAF score, and AI-COAF score, respectively (Figure S4 in Multimedia Appendix 1). Consistent with the primary analysis, the Delong test indicated that the AUROCs of the AI-POAF score and AI-COAF score remained significantly higher than that of the POAF score (both P<.001), while no significant difference was observed between the POAF score and the AI-ECG-AF model score (P=.692).

In this study, we evaluated the potential of the AI-ECG-AF model as a predictor for postoperative AF in cardiac surgery patients. After adjusting for other clinical variables, a 10% absolute increase in the AI-ECG-AF model score was associated with a 1.197- to 1.209-fold increase in the odds of developing postoperative AF. We demonstrated that the AI-ECG-AF model score provides additive or synergistic predictive value when integrated with an existing postoperative AF prediction tool or other risk factors: the AUROC of the existing POAF score was 0.643; adding the AI-ECG-AF model score increased it to 0.680 (P<.001), and combining the AI-ECG-AF model score with other risk factors raised it to 0.710 (P<.001).

Attempts to predict postoperative AF have faced significant challenges. Various studies have evaluated various scoring systems, originally developed for different purposes, for predicting postoperative AF in cardiac surgery patients [6,23]. For example, the European System for Cardiac Operative Risk Evaluation (for predicting mortality after cardiac surgery) [24], the Congestive Heart Failure, Hypertension, Age≥75 years, Diabetes Mellitus, Stroke–Vascular Disease, Age 65-74 years, and Sex Category Score (for predicting stroke risk in AF patients) [25], and the Hypertension, Age≥75 years, Transient Ischemic Attack or Stroke, Chronic Obstructive Pulmonary Disease, and Heart Failure Score (for predicting progression from paroxysmal to sustained AF) [26] have been evaluated for predicting postoperative AF in cardiac surgery patients. However, these scoring systems have only demonstrated moderate performance at best [6,23]. Although the POAF score was specifically developed for predicting postoperative AF in cardiac surgery patients [7], it similarly showed only moderate performance and has not been widely adopted in clinical practice or subsequent research [6,23]. There remains a clear unmet need for more effective postoperative AF risk stratification in cardiac surgery patients, and modern approaches such as AI may offer better solutions.

AI applications to ECG, the electrical fingerprint of the heart, are opening up new possibilities. There is growing evidence that advanced AI techniques with deep convolutional neural networks are capable of detecting subtle signals and patterns from ECGs that do not fit traditional knowledge and are unrecognizable by the human eye, enabling prediction of diseases that were previously unpredictable with ECGs [8-12,27,28]. The AI-ECG-AF model exemplifies this advancement, as it detects the electrocardiographic signature of AF present in non-AF ECGs [13]. This has particularly important implications because AF is often paroxysmal and therefore difficult to capture during ECG screening, yet it remains a major risk factor for stroke that requires anticoagulation [13,29]. There is robust evidence supporting the real-world effectiveness of this model, with a prospective trial demonstrating that AI-ECG-AF model-guided targeted screening of AF with ECGs resulted in a significant increase in AF detection rates, particularly among those classified as high risk by the model [30].

AI-ECG-AF can be understood as a model that identifies subtle ECG patterns potentially associated with structural substrates for AF, such as atrial fibrotic changes related to aging [13,31]. Cardiac surgery induces significant stress on the heart, including inflammation, fluid shifts, and changes in the autonomic nervous system, all of which can further exacerbate preexisting atrial vulnerabilities and increase the likelihood of developing postoperative AF [3,14,15,32]. The findings from our study suggest that the AI-ECG-AF model is capable of identifying these underlying predispositions in patients undergoing cardiac surgery, allowing for the recognition of individuals at heightened risk for postoperative AF.

We demonstrated that the AI-ECG-AF model was an independent risk factor for postoperative AF and enhanced the predictive performance of conventional clinical scores. Unlike existing tools such as the POAF score—which are derived from clinical risk factors—the AI-ECG-AF model offers complementary information by capturing atrial electrophysiological vulnerability embedded in the ECG. When used in combination, the two scores provided significantly improved prediction, underscoring the potential of the AI-ECG-AF score as a novel biomarker that complements traditional clinical models and supports the implementation of more targeted prophylactic interventions and perioperative monitoring strategies. Notably, the AI-ECG-AF model, though originally developed in a general patient population, demonstrated clinical utility as a predictive biomarker even in the high-risk setting of cardiac surgery, highlighting its potential generalizability. This suggests its broader applicability, including possible integration into preoperative risk assessment protocols across various cardiac conditions. Of note, our results remained consistent irrespective of whether the broader 30-day definition or the more restrictive 7-day definition for postoperative AF was applied.

In contrast to the previous study that limited its AI-ECG-AF model development to normal sinus rhythm ECGs [13], our study incorporated all non-AF ECGs, excluding only those with lead misplacements, artifacts, or artificial pacemakers. We successfully demonstrated that the AI-ECG-AF model could maintain high performance even when processing diverse, heterogeneous ECGs, thus expanding its clinical applicability. Our model was trained on a comprehensive dataset of 4.05 million ECGs from 1.13 million patients, including 0.61 million AF-positive ECGs from 0.04 million patients. This extensive dataset helps minimize overfitting, accounts for broader ECG variability, and enhances the model’s generalizability in real-world clinical applications [33]. The developed AI-ECG-AF model, along with detailed usage instructions, has been made fully accessible to the public through our online repository [22].

Although multiple studies have demonstrated that amiodarone or beta-blockers can provide a protective effect against postoperative AF following cardiac surgery [34,35], there remains a lack of conclusive evidence supporting the routine use of these medications in all patients [36]. Furthermore, patients who develop postoperative AF are often advised to undergo long-term oral anticoagulation therapy, which has been shown to offer potential benefits in reducing the risk of thromboembolic events [37,38]. Given these considerations, accurately identifying individuals at an elevated risk of developing postoperative AF is of critical importance, as these patients require close rhythm monitoring not only in the immediate postoperative period but also throughout long-term follow-up. The AI-ECG-AF model may serve as a novel biomarker in recognizing high-risk patients who could potentially benefit from preoperative administration of amiodarone or beta-blockers, in addition to targeted strategies for the prevention of thromboembolic complications. Future studies should also aim to define and validate clinically meaningful thresholds for the AI-based scores evaluated in this study (ie, the AI-ECG-AF, AI-POAF, and AI-COAF scores), incorporating health-economic analyses and expert consensus to balance the benefits of intervention against potential harms, thereby facilitating clinical implementation.

This study’s findings should be interpreted in light of the following limitations. The study is subject to limitations inherent in its single-center retrospective design. First, as the data were derived from a single institution, the generalizability of our results to other populations or health care settings may be limited. External validation across diverse institutions is warranted to confirm the robustness of our findings. AI-ECG-AF has the advantage of being applicable using ECGs alone, but subtle waveform variations between different ECG machines may affect model performance, potentially necessitating institution-specific fine-tuning. Additionally, differences in population characteristics may require local calibration of composite scores such as AI-POAF or AI-COAF to ensure predictive accuracy. Second, our analysis was limited to cardiac surgery, yet approximately 85% of surgeries are noncardiac, which still show postoperative AF rates of 2%‐30% [39,40]. Although the mechanisms underlying POAF after noncardiac surgery are multifactorial and not fully defined, systemic inflammation, perioperative sympathetic surges, and other stressors may exacerbate atrial vulnerability and trigger AF [39,40]. Accordingly, AI-ECG-AF may also serve as a risk factor for POAF in this broader surgical population, warranting future investigation. Third, AF was defined based solely on documented episodes in stored standard 12-lead ECGs, so transient AF episodes could have been missed. In particular, AF episodes not recorded in the database—such as those detected only on continuous telemetry monitoring—or paroxysmal AF episodes occurring transiently in daily life may not have been captured. Such omissions could have influenced both the performance of the AI-ECG-AF model and the evaluation of postoperative AF. Fourth, while perioperative antiarrhythmic medications likely influence postoperative AF, we could not adjust for all such medications due to the retrospective nature of the study and the heterogeneity in medication protocols. Some antiarrhythmic medications, such as preoperative beta-blockers, were systematically recorded in our institutional registry and included in certain analyses. However, comprehensive data on all antiarrhythmic drugs were unavailable, and accurate retrospective collection was infeasible. Furthermore, detailed information such as dosage, timing, and route of administration was not reliably captured. Fifth, although AI-ECG-AF was first introduced several years ago [13], our model development did not pursue novel advances in architecture or feature extraction. Instead, we focused on clinical evaluation, assessing its utility as a predictive biomarker for POAF in cardiac surgery patients. Despite using an older architecture (EfficientNet-B0) and a standard pipeline, the model achieved an AUROC exceeding 0.9 in the development test set—deemed sufficient for clinical validation. Future studies could explore more advanced architectures or innovative techniques to enhance model performance, potentially leading to even stronger results in clinical validation. Sixth, our logistic regression analyses included multiple ECGs per patient, but we treated each as an independent observation, which may have led to underestimated standard errors and overstated statistical significance. Statistical techniques that explicitly account for within-patient correlation were not applied. Nonetheless, the large sample size and consistent subgroup findings support the robustness of our results. Finally, the AI-ECG-AF model was not trained to distinguish between different AF phenotypes (eg, paroxysmal vs persistent), and future models may be improved by incorporating more granular AF phenotype data.

In conclusion, the AI-ECG-AF model serves as a novel, robust, and independent risk factor for postoperative AF following cardiac surgery and provides additive or synergistic predictive value when integrated with existing postoperative AF prediction tools or with other risk factors. By capturing atrial electrophysiological vulnerability not reflected in conventional clinical scores, the AI-ECG-AF model may function as a noninvasive biomarker for preoperative risk stratification for postoperative AF prediction in cardiac surgery patients, potentially enabling health care providers to initiate tailored prophylactic therapy and implement targeted monitoring during the perioperative period.