Real-world application of deep learning for ECG-based prediction of coronary artery disease and revascularization needs

Abstract

Aims

Early detection of the need for coronary revascularization and timely intervention may reduce fatal events, but limited screening tools often leads to underdiagnosis. The aim of this study is to use a deep learning model (DLM) that utilizes electrocardiography (ECG) and the eXtreme Gradient Boosting (XGBoost) model to predict risk of coronary revascularization in the general population.

Methods and results

This study included patients with at least one ECG per patient. The development set comprised 113 451 patients for training a DLM. After excluding patients with elevated troponin I levels and those without follow-up records, the internal validation set consisted of 66 680 patients. The external validation was conducted using data from a community hospital. XGBoost predicted events based on demographic data and ECG features. The primary endpoint was coronary revascularization within 1 year. Model performance was evaluated using the C-index. The DLM stratified patients by risk of coronary revascularization within 1 year. The study included 51% males with a mean age of 53 years, 10% with diabetes, and a revascularization rate of 2.6%. High-risk patients had a hazard ratio of 9.77 (95% CI: 7.63–12.51) compared with low-risk patients. The C-index was 0.825 (95% CI: 0.81–0.84). Combining demographic and AI-ECG data, XGBoost achieved a C-index of 0.884 (95% CI: 0.87–0.89). Comparative C-index analysis revealed significantly different discriminative performance between models (P = 1.110223e−15).

Conclusions

The DLM demonstrates ECG's potential as a screening tool for coronary revascularization, enabling opportunistic detection and prompting further evaluation of high-risk patients.

Graphical Abstract

Graphical Abstract.

Introduction

Cardiovascular disease remains a leading cause of mortality and morbidity worldwide,¹ with acute myocardial infarction (AMI) being a critical condition that affects millions of patients annually, imposing significant burdens on individuals, health care systems and society.² At its foundation is coronary artery disease (CAD), a progressive condition characterized by arterial plaque build-up, which, while potentially devastating, is preventable through early identification and appropriate interventions.

CAD presents with diverse symptoms, including chest pain, tightness, dyspnea, cold sweating and dizziness, often leading to misdiagnosis due to symptomatic overlap with other conditions.^3,4 Despite its fatal nature and widespread recognition, CAD remains underdiagnosed; a mere 34.8% of heart failure patients, for whom CAD is a major cause, undergo appropriate testing (stress testing, cardiac magnetic resonance imaging or coronary angiography), with only 9.3% receiving revascularization.⁵ While coronary angiography represents the diagnostic gold standard, its invasive nature renders it impractical for population screening. Current clinical guidelines emphasize precise risk stratification given the prevalence of similar symptom presentations;^3,4 however, both medical and economic resource constraints limit existing screening modalities, which often involve prolonged wait times or radiation exposure, while offering suboptimal sensitivity and specificity (∼75–90%).^4,6 In remote areas with limited medical resources, CAD is often underdiagnosed.⁷ If a straightforward and consistent daily practice workflow is implemented, it can assist physicians in identifying patients with severe CAD and facilitate further evaluation. This approach would be highly beneficial.

Electrocardiography (ECG) is an extensively utilized examination, with ∼3 million ECGs recorded daily worldwide.⁸ Although ECGs may reveal ischemic changes, such as Q waves, poor R progression or T wave inversion that might indicate potential CAD,^9,10 daily clinical practice lacks consensus or convincing criteria for CAD diagnosis using ECG alone. In the current era of artificial intelligence integration, AI-enabled ECG (AI-ECG) algorithms employing deep learning can extract meaningful patterns from complex ECG information and have demonstrated effectiveness in identifying heart failure and acute coronary syndrome (ACS), predicting paroxysmal atrial fibrillation from sinus rhythm, and even predicting mortality.^11–14 Preliminary research suggests that AI-ECG trained on small sample sizes has potential as a CAD screening tool.^15,16 A recent large-scale study using ECG from the 90 days preceding percutaneous coronary intervention (PCI) as cases, in contrast with ECG from patients without evidence of stenosis as controls, achieved an area under the receiver operating characteristic curve of 0.7^15,16 however, this methodology presents notable limitations. It fails to account for temporal electrocardiographic variations preceding PCI, whereas both training and control cohorts comprised high-risk patients, introducing potential selection bias. Additionally, the exclusion of moderate-risk cases (CAD-RADS 2–3) in comparable studies disregards clinically significant information, potentially inflating model accuracy by eliminating challenging borderline cases.¹⁷

Our objective was to develop a model that predicts which patients have CAD requiring coronary revascularization using electronic health records to reconstruct patient information, including coronary interventions or stenosis. By employing deep survival techniques, our research design better reflects real-world scenarios and evaluates the true predictive capabilities of AI-ECG in CAD.

Methods

Study populations

This was a multicenter retrospective cohort study. We collected ECG data from the general population, including outpatient clinics, emergency departments (EDs), and inpatient departments. An event was defined as a patient with angiography-proven significant CAD receiving either PCI or a coronary artery bypass graft (CABG) at the Tri-Service General Hospital and Tri-Service General Hospital Tingjhou Branch between 1 January 2008, and 28 February 2022. All patients received PCI with stent implantation according to Taiwan's National Health Insurance policy for severe stenosis (more than 70% by quantitative coronary angiography assessment) or intermediate lesions (defined as coronary stenosis between 50 and 70% with physiological significance, indicated by coronary fractional flow reserve <0.80 or evaluated by intravascular imaging to identify unstable plaques or severe stenosis). Alternatively, patients may have received CABG according to guideline suggestions. We excluded patients with elevated troponin I levels (>0.5 ng/mL) to rule out ACS in validation set. Our aim was to develop a model for severe chronic coronary syndrome detection.

Data collection and parsing

The algorithm for ECG collection and parsing is shown in Figure 1. We collected ECG data from 1 January 2008, to 28 February 2022. In total, 227 299 patients with at least one ECG during the study period were included from Tri-Service General Hospital to develop the DLM. All patients included in the study received ECG as part of their routine clinical workup. Outpatient and general population screening participants underwent ECG for preventive health evaluations or cardiac risk assessments, whereas EDs and inpatient individuals received ECG due to suspected cardiac symptoms or other clinical indications. We used the first ECG of each patient to develop the model. In the validation set, we excluded patients with elevated troponin I levels to rule out ACS, as well as patients without subsequent medical records. These ECG data were collected from patients in outpatient departments (OPDs), inpatient departments (IPDs), and EDs. These ECG data were divided into 3 datasets randomly. Approximately 50% (113 451) of the ECG data were used to develop the model, and ∼20% (44 771) were used to tune the model. A total of 29.3% (66 680) of the data were used for internal validation. These subsets were randomly selected using stratification on the basis of key demographic and clinical variables (age, sex, comorbidities) to ensure representative samples of the overall population and to minimize selection bias. An additional external validation set, comprising a total of 61 777 patients, was collected from Tri-Service General Hospital Tingjhou Branch.

Figure 1.

Development, tuning, internal validation, and external validation sets. We utilized electrocardiography data from 227 288 patients at an academic medical center to establish development, tuning, and internal validation sets. In these sets, we excluded patients with troponin I levels >0.5 ng/mL and those without follow-up records. The first ECG recorded was used for tuning or internal validation. Additionally, we included ECG data from 61 777 patients at a community hospital as our external validation set for accuracy testing. CR, coronary revascularization; DLM, deep learning model; ECG, electrocardiography.

Variables

Patient characteristic information, past histories, and digital ECG signals were extracted from the electronic medical records of both hospitals. Patient characteristics included sex, age, and the location where the ECG data were collected. Patient disease histories, such as AMI, stroke (STK), CAD, atrial fibrillation (Afib), PCI, and heart failure, were obtained prior to the index date of the ECG using the corresponding International Classification of Diseases, Ninth Revision and Tenth Revision (ICD-9 and ICD-10, respectively) codes, as described previously.

The digital ECG signals were recorded using a Philips 12-lead ECG machine (PH080A, Philips Medical Systems, 3000 Minuteman Road, Andover, MA 01810, United States) in the standard 12-lead format, with a sampling rate of 500 Hz over a 10-s period. We also collected eight ECG measurements from the Philips system as structured variables.

The primary outcome of this study was new-onset coronary revascularization. The incidence time was calculated with reference to the date of the ECG examination.

Model development

During the training process, each original ECG signal was presented as a 12 × 5000 matrix and randomly cropped in the format of a 12 × 4096 matrix as input. We established the survival DLM on the basis of the Cox proportional hazard model. The output of the proposed survival DLM was a continuous value of h(x), which was the output of the last fully connected layer for each ECG. The loss function was also based on the Cox partial likelihood function, and the architecture of the DLM to extract the high-order features has been previously reported.¹⁸ In this survival DLM with binary output, the cases are the patients who received coronary revascularization within the following time, whereas the controls are the patients who did not receive coronary revascularization. We then entered ECG data into the network with a batch size of 32, and Adam optimization was used with an initial learning rate of 0.001 (β1 = 0.9 and β2 = 0.999), which was decayed by a factor of 10 each time the loss occurred after an epoch. A weight decay coefficient of 10⁻⁴ and early stopping were used to prevent overfitting. The DLM with the lowest loss in the tuning set was stored, and the performance in both the internal and external validation sets was evaluated.

We developed additional machine learning models, eXtreme Gradient Boosting (XGBoost) models, and compared three XGBoost models to predict future coronary revascularization, which differ in the input features used for training. Model 1 included demographic and basic clinical characteristics (e.g. age, sex, and comorbidities), Model 2 utilized only ECG features (e.g. T wave axes and QRS wave axes), and Model 3 combined features from both models. This progression assessed the incremental impact of clinically relevant data on predictive performance.

Hyperparameters (e.g. learning rate, maximum depth, and number of trees) were optimized via grid search with cross-validation on the training dataset. XGBoost’s feature importance metric identified the most predictive features, demonstrating the contribution of these features. Model performance was evaluated using the C-index (95% CI) on internal and external validation datasets to ensure robustness. Additionally, XGBoost was employed to integrate AI-ECG data to evaluate its prediction performance.

Statistical analysis

The distribution of patient characteristics was expressed as numbers of patients, percentages, means, and standard deviations where appropriate. We provided two cutoff points: the high-sensitivity cutoff point was selected by maximizing Youden's index in the tuning set, and the highest positive predictive value (PPV) cutoff point was selected by maximizing the F-measure in the tuning set. The entire dataset was divided into three risk groups: low-risk (probability less than the high-sensitivity cutoff point), middle-risk (probability between the high-sensitivity cutoff point and high-PPV cutoff point), and high-risk (probability higher than the high-PPV cutoff point). High-sensitivity ensures that the model identifies as many high-risk patients as possible, minimizing false-negatives. Youden’s index provides a robust method for determining this balance. Although our model revealed great negative predictive value (NPV), we did not use that value as our cutoff point. Because NPV impacts by prevalence, our model was applied in the general population, which mostly belongs to the low-risk group. We selected the high-PPV cutoff point to identify patients with a high likelihood of requiring coronary revascularization. This is particularly relevant for clinical decision-making, where resources and interventions are prioritized for patients at the highest risk.

To measure the performance of coronary revascularization incidence prediction, we used multivariable Cox proportional hazards models with a C-index. Hazard ratios (HRs) and 95% confidence intervals (95% CIs) among these three groups, adjusted for sex and age, were calculated via Kaplan–Meier (KM) analysis. Additionally, we used forest plots to conduct stratification analysis across sex, ECG assessment locations, and patient comorbidities and to evaluate the model performance differences across specific groups. To evaluate whether the two prediction models demonstrated significantly different discriminative performance for survival outcomes, we conducted a statistical comparison of their C-indices using the compare C package in R based on Uno's C with perturbation-resampling method.

Results

Patient population

In total, 289 076 eligible patients were enrolled in this study, and their baseline characteristics are summarized in Table 1. Within the development cohort, 10.3% of patients had diabetes mellitus (DM), 9.2% had CAD, and 13.9% had hyperlipidemia. Notably, the external validation set exhibited a higher prevalence of these conditions, with 17.6% of patients having DM, 14.8% having CAD and 28.5% having hyperlipidemia. The average age of the cohort was 53.1 years, with a median follow-up duration of 2.11 years in development set. The overall mortality rates were similar across the different sets: 5.2% in the development set, 5.3% in the internal validation set, and 5.4% in the external validation set. The rates of coronary revascularization events were ∼3.1% in the development set, 2.6% in the internal validation set, and 2.7% in the external validation set. The average time between ECG and revascularization was 93.91 days. This reflects the interval from the initial ECG examination to the clinical decision and subsequent intervention.

Table 1.

Baseline characteristics

	Development	Tuning	Internal validation	External validation
Demography
Gender (male)	58 117 (51.2%)	23 425 (51.2%)	34 602 (50.8%)	31 255 (50.6%)
Age (years)	53.3 ± 18.4	53.2 ± 18.4	53.3 ± 18.4	55.4 ± 18.8
Disease history
DM	11 707 (10.3%)	4646 (10.2%)	6852 (10.1%)	10 845 (17.6%)
HTN	1222 (1.1%)	478 (1.0%)	734 (1.1%)	1311 (2.1%)
HLP	15 794 (13.9%)	6309 (13.8%)	9223 (13.5%)	17 604 (28.5%)
CKD	3632 (3.2%)	1450 (3.2%)	2187 (3.2%)	2967 (4.8%)
STK	6241 (5.5%)	2528 (5.5%)	3738 (5.5%)	4781 (7.7%)
HF	2468 (2.2%)	972 (2.1%)	1457 (2.1%)	2360 (3.8%)
CAD	10 408 (9.2%)	4240 (9.3%)	6209 (9.1%)	9134 (14.8%)
AMI	865 (0.8%)	364 (0.8%)	559 (0.8%)	441 (0.7%)
Afib	1255 (1.1%)	502 (1.1%)	752 (1.1%)	1162 (1.9%)
COPD	6305 (5.6%)	2526 (5.5%)	3723 (5.5%)	7103 (11.5%)
Electrocardiogram data
ECG [Rate]	77.0 ± 17.1	77.0 ± 17.2	76.9 ± 17.1	76.0 ± 16.6
ECG [PR]	161.7 ± 27.2	161.7 ± 27.1	161.8 ± 28.0	163.2 ± 27.9
ECG [QRSd]	93.8 ± 15.6	93.8 ± 15.8	93.8 ± 15.8	94.2 ± 15.7
ECG [QT]	388.9 ± 39.1	388.7 ± 38.9	388.9 ± 39.1	390.5 ± 39.0
ECG [QTc]	434.4 ± 34.5	434.1 ± 34.9	434.3 ± 34.9	433.4 ± 34.5
ECG [Axes_P]	49.9 ± 28.1	49.9 ± 28.7	49.9 ± 28.0	49.5 ± 28.8
ECG [Axes_QRS]	43.7 ± 43.0	43.8 ± 42.8	44.0 ± 42.8	41.9 ± 42.7
ECG [Axes_T]	38.7 ± 39.0	38.2 ± 38.4	38.7 ± 38.6	39.1 ± 38.0
Follow-up
Median follow-up time	2.11	2.11	2.12	3.35
Mortality event	3242 (5.2%)	3552 (5.2%)	6068 (5.3%)	2468 (5.4%)
Revascularization event	1940 (3.1%)	1747 (2.6%)	2914 (2.6%)	1213 (2.7%)

Afib, atrial fibrillation; AMI, acute myocardial infarction; CAD, coronary artery disease; CKD, chronic kidney disease; DM, diabetes mellitus; HF, heart failure; HLP, hyperlipidemia; HTN, hypertension; STK, stroke.

Model performance

Figure 2 illustrates our model's performance. The cutoff point of the high-risk group was greater than the probability of the high-PPV cutoff point, and the cutoff point of the middle-risk group was between the high-sensitivity cutoff point and the high-PPV cutoff point. This model yielded an area under the curve (AUC) of 0.809 for the middle- and high-risk groups in the internal validation set. In the external validation set, the model demonstrated comparable robustness, with an AUC of 0.799 for both risk categories. Notably, we observed high negative predictive values: 99.5% for the middle-risk group and 98.9% for the high-risk group in the internal validation, with the external validation set yielding similar results. The PPVs for the middle-risk group were 3.9% and 3.4% in the internal and external validation sets, respectively. The high-risk group had improved PPVs, reaching 9.7% in the internal validation set and 10.1% in the external validation set. These results highlight the importance of concentrating targeted interventions in high-risk groups.

Figure 2.

Performance of the deep learning model in detecting coronary artery disease. The figure shows the areas under the receiver operating characteristic and precision–recall curves for the deep learning model's predictions of coronary revascularization. Two key operating points are highlighted: (ⅰ) Medium risk threshold: Selected using the maximum Youden index of the AUC curve [sum of sensitivity (Sens.) and specificity (Spec.)]. (ⅱ) High-risk threshold: The highest positive predictive value cutoff point was selected by maximizing the F-measure in the tuning set. These operating points are marked by circles on their respective curves. The figure also provides associated performance metrics, including the AUC, PRAUC, sensitivity, specificity, positive predictive value, and negative predictive value, for each threshold.

Figure 3 shows the comprehensive subgroup analyses confirming the model's consistent performance across diverse clinical parameters. The sex-specific C-index demonstrated robust discriminative ability: 0.78 (95% CI: 0.76–0.80) for males and 0.82 (95% CI: 0.79–0.85) for females. Examination revealed comparable performance regardless of the clinical context: 0.83 (95% CI: 0.79–0.86) for the EDs cohort, 0.85 (95% CI: 0.82–0.87) for the IPDs cohort, and 0.81 (95% CI: 0.79–0.84) for the OPDs cohort in the internal validation set.

Figure 3.

AI-ECG performance stratified by sex, examination position, and comorbidities. Subgroup analysis was performed with a forest plot to evaluate the impact of sex, the position in which patients underwent ECG examination and common comorbidities. The dashed vertical lines indicate the reference (C-index: 0.8) and the overall diagnostic C-index (here, C-index: 0.825). Afib, atrial fibrillation; CKD, chronic kidney disease; DM, diabetes mellitus; ED, emergent department; HLP, hyperlipidemia; HTN, hypertension; IPD, inpatient department; OPD, outpatient department.

Age stratification revealed notably superior performance in younger populations, with a C-index of 0.91 (95% CI: 0.86–0.96) for patients under 45 years compared with 0.73 (95% CI: 0.70–0.76) for those over 65 years. This finding suggests the utility of early risk identification among younger individuals, where traditional risk factors may be less prevalent.

With respect to comorbidities, the model maintained clinically meaningful discriminative ability across various conditions, although performance modestly declined in patients with established comorbidities. The C-index for patients with Afib, CAD, and CKD was 0.67 (95% CI: 0.59–0.75), 0.71 (95% CI: 0.69–0.74), and 0.71 (95% CI: 0.65–0.78), respectively, compared with 0.83–0.85 in patients without these conditions. Hypertension, DM, and hyperlipidemia had minimal impacts on model performance, with C-index remaining above 0.75 regardless of these comorbidities. Both the internal and external validation cohorts demonstrated consistent performance patterns across all the examined subgroups, further supporting the model's generalizability in diverse clinical scenarios.

Risk stratification

Patients were monitored until they underwent revascularization, which included PCI with balloon dilatation, stenting, or CABG. We determined risk cutoff points on the basis of the maximum F score and Youden's index to define high-, middle-, and low-risk groups. Figure 4 illustrates the incidence of revascularization across these risk categories. In the internal validation set, we observed a C-index of 0.825 (95% CI: 0.81–0.84). After adjusting for sex and age, the hazard ratio was 9.77 (95% CI: 7.63–12.51) for the high-risk group and 3.37 (95% CI: 2.72–4.16) for the middle-risk group, with the low-risk group used as the reference. The external validation set yielded a C-index of 0.814 (95% CI: 0.799–0.829). In this set, the adjusted hazard ratio for the high-risk group was 11.07 (95% CI: 8.85–13.85), and that for the middle-risk group was 2.64 (95% CI: 2.15–3.24), again referencing the low-risk group. We focused on chronic coronary syndrome by excluding patients with possible ACS. Our DLM demonstrated exceptional efficacy in stratifying the risk of coronary revascularization in the general population.

Figure 4.

One-year incidence of developing coronary artery disease outcomes in patients with an initially normal diagnosis. The high-risk and middle-risk cutoff points were selected on the basis of the maximum F score and maximum Youden's index in the tuning set, respectively (excluding cases with troponin I > 0.5 mg/dL). The high-risk group showed a sex- and age-adjusted hazard ratio of 9.77 in the internal validation group. The middle-risk group and the lower-risk group (reference) are also displayed for comparison. Sex- and age-adjusted HR, sex- and age-adjusted hazard ratio.

Figure 4 reveals that, in the high-risk group, the cumulative incidence of revascularization was 5.6% at 6 months and 6.7% at 12 months. In contrast, the low-risk group presented rates of 0.3% at 6 months and 0.4% at 12 months. The external validation set exhibited similar trends, with a cumulative incidence of 7.2% at 6 months and 8.5% at 12 months in the high-risk group and 0.4% at 6 months and 0.5% at 12 months in the low-risk group. Notably, the separation of risk group curves occurred rapidly, within one month, and was consistent across both the internal and the external validation sets. In the external validation set, 4.0% of high-risk patients underwent coronary revascularization within one month, in stark contrast to only 0.2% in the low-risk group.

To evaluate the incidence of medium- and long-term coronary revascularization while minimizing short-term event bias, we excluded revascularizations occurring within 90 days post-ECG. The results showed in Supplementary material online, Figure S1 revealed that, compared with the low-risk cohort, the high-risk cohort presented significant hazard ratios (8.73 and 10.11 in the internal and external validation sets, respectively). The C-index remained robust at 0.831 and 0.818, confirming the model's persistent predictive ability beyond the short-term period.

Figure 4 illustrates an initial sharp increase in revascularization events post-ECG. To address concerns regarding potential upstream diagnostic testing influence, we conducted Schoenfeld residual analysis (see Supplementary material online, Figure S2), confirming the validity of proportional hazards assumptions across risk stratification groups. The P values for the middle-risk (0.442) and high-risk (0.914) groups exceeded the 0.05 threshold, whereas the residual plots exhibited no significant temporal trends, substantiating our Cox model's stability for revascularization event analysis. Another possibility is that the early rise in coronary revascularization rates reflects clinical practice patterns in Taiwan, where patients with significant symptoms and abnormalities on non-invasive tests are recommended for coronary angiography after approximately three months of medical therapy if symptoms persist. In contrast, patients who did not undergo early PCI were generally stable and managed conservatively with medical therapy.

In Supplementary material online, Figure S3, we excluded coronary revascularization occurring within 90 days post-ECG to eliminate potential confounding by ACS events such as unstable angina. Subsequent analysis evaluated future ACS risk stratification capability across the defined risk groups. The results revealed substantial hazard ratios for the high-risk and low-risk cohorts: 10.72 in the internal validation set (C-index 0.832) and 5.51 in the external validation set (C-index 0.821). These findings confirm that the risk stratification model effectively identifies patients with a significantly elevated risk of experiencing ACS events within 1 year of follow-up. The maintenance of robust C-index across both validation cohorts further substantiates the model's discriminative ability for identifying patients at heightened risk of future acute coronary events, even when early revascularization procedures are excluded.

Feature importance and the XGBoost model

As shown in Figure 5, we used another machine learning method, XGBoost, which integrates AI-ECG data and predicts future coronary revascularization. This model depicts feature importance in predicting coronary revascularization, categorized into patient characteristics, ECG features and combinations. Among patient characteristics, age and history of CAD emerged as crucial factors, aligning with clinical determinations. For the ECG features, the T wave axis and QRS wave axis played significant roles. When patient characteristics and ECG features were combined, a history of CAD was the most important predictor, followed by age.

Figure 5.

Feature importance and predictive ability comparison of the DLM and XGBoost models. (A) The components of the AI-ECG identified the high-risk group. We trained three XGBoost models using patient characteristics, ECG features, and their combination to predict coronary revascularization. The bars represent the relative importance of the components in predicting coronary revascularization in the XGBoost models. (B) The prediction ability of all patient data for coronary revascularization within 1 year (C-index). The error bars are the 95% CIs of each C-index. The blue and red bars represent the results of prediction using individual patient characteristics and ECG features, respectively. The green bars represent predictions integrating features from XGBoost and logistic regression. The brown bars represent DLM data combined with XGBoost. All analyses were based on data from the entire population in this trial. DLM, deep learning model.

This machine learning model, XGBoost, demonstrated improved predictive ability when combined with our DLM, patient characteristics, and demographic data. The C-index for this combined model reached 0.884 (95% CI: 0.872–0.896) in the internal validation set and 0.866 (95% CI: 0.853–0.879) in the external validation set. Compared with our DLM, which was trained on ECG of patients, our DLM achieved a C-index of 0.825 (95% CI: 0.81–0.84) in the internal validation set and 0.814 (95% CI: 0.799–0.829) in the external validation set. To evaluate whether the two prediction models demonstrated significantly different discriminative performance for survival outcomes, we conducted a statistical comparison of their C-indices using the compare C package in R based on Uno's C with perturbation-resampling method. This comparison produced a P-value of 1.110223e−15, indicating that the difference in C-index between the two models is highly statistically significant.

In our study, we utilized commonly available data, including age, sex, medical history, and ECG results, to develop a machine learning model. Notably, this model does not require invasive examinations such as blood tests, allowing for straightforward data acquisition. Our findings suggest that this model has significant potential as a screening tool or an opportunistic examination for significant obstructive CAD.

Discussion

This retrospective cohort study simulated real-world scenarios to evaluate AI-ECG performance in predicting future PCI or CABG risk, achieving a C-index exceeding 0.81 in both the internal and external validation cohorts. In our high-risk group, the hazard ratio of coronary revascularization within 1 year was 9.77–11.07 times greater than that in the low-risk group, with this stratification capability expected to enhance CAD management through earlier identification of at-risk patients. Subgroup analysis revealed that the AI-ECG algorithm demonstrated superior predictive accuracy in patients without common comorbidities and in those under 45 years of age. Furthermore, the model maintained consistent performance regardless of the assessment location (ED or OPD). These findings underscore the potential utility of this approach as an opportunistic examination or screening tool in general populations.

According to current guidelines, risk stratification is the first step for treatment decisions in patients with angina or equivalent symptoms without a history of CAD.^3,4 Pre-test probability (PTP), estimated using symptoms, age and sex, facilitates risk stratification and determines the need for non-invasive testing when the PTP exceeds 15%.⁴ Non-invasive tests can confirm the diagnosis and guide therapeutic strategies. Observational studies confirmed that patients with moderate to severe ischemia detected by imaging modalities benefit from early coronary revascularization.^19–21 Randomized controlled trials have demonstrated that revascularization improves outcomes and reduces the need for urgent interventions in high-risk groups.^22,23 However, considering the substantial population in need of diagnostic evaluations via non-invasive tests, the establishment of a screening tool capable of accurately identifying individuals with the most severe conditions would be extremely beneficial.

Previous studies have developed models to detect obstructive CAD using 12-lead ECGs with good results (AUROCs: 0.7) but with significant limitations.^15,16 Two teams used pre-coronary angiography ECGs to construct models. Another recently published study used coronary angiography results and coronary computed tomography angiography results to train the model, creating selection bias toward high-risk patients and lacking low- and moderate-risk groups in the model,^15,16,24 whereas another team achieved impressive results (AUROCs: 0.85) but excluded CAD-RADS 2–3 patients. This model excluded the moderate-risk group.¹⁷ The above models used specific groups to train models, making them unsuitable for real-world use. Another team developed a highly predictive AI-ECG model with a C-index of 0.88, but the model focused on patients with chest pain in the ER, not the general population.²⁵ Our model focuses on chronic coronary syndrome, excluding only patients with elevated troponin I levels to eliminate the possibility of ACS, thereby better representing the general population. However, our model still cannot replace clinical guidelines when doctors make revascularization decisions. We chose coronary revascularization as our training target because these patients have the most severe disease and need to be quickly identified for further non-invasive testing.

Primary CAD prevention in individuals typically relies on 10-year coronary event risk predictions using validated clinical risk scores, including the Framingham risk score,²⁶ AHA/ACC Pooled Cohort Equation,²⁷ and European SCORE2,²⁸ which categorize patients into risk groups to guide therapy. However, these estimators remain suboptimal,^29–31 with known limitations, including risk over- or underestimation in groups with unincluded risk modifiers.^30,32,33 Real-world studies have shown that the C-index of the PCE is 0.78,³⁴ whereas the AUCs of the FRS and SCORE2 are 0.63 and 0.75, respectively.^29,31 Our model's ability to identify patients at significantly high-risk for coronary revascularization with an ∼10% PPV and a C-index over 0.88 potentially enables more precise risk stratification, facilitating risk-guided examinations and providing clinicians with a method to identify high-risk asymptomatic individuals requiring further evaluation.

Furthermore, by incorporating medical history, patient characteristics and AI-ECG data, XGBoost improved the C-index to over 0.88, with most high-risk patients receiving revascularization within 1 year following AI-ECG-positive results. According to our XGBoost findings, the key features of the ECG are the T wave axes and QRS wave axes, which are important for the model to make a prediction. Notably, the predictive ability was reduced in patients with comorbidities (AFib, DM, CKD, CAD, hypertension and hyperlipidemia). A preliminary study also demonstrated similar results. They compared comorbidities and medication prescriptions between false-positive and true negative classifications C-index via an AI-ECG model. The results revealed that aortic valve stenosis, atrial fibrillation, congestive heart failure and essential hypertension increase the false-positive rate in predicting obstructive CAD or elevated CAC. They also reported that medications cause similar results. Aspirin, atorvastatin, and losartan increase false-positive rates.¹⁷ This likely reflects the impact of comorbidities and guideline-directed therapies, suggesting that the great potential of AI-ECG lies in screening populations without known comorbidities.

This study has several limitations. First, its retrospective nature inherently introduces biases, despite efforts to mitigate them. External validation was conducted using data from another hospital, and concerns persist regarding potential overfitting and the limited racial diversity within the dataset. Second, although efforts were made to exclude ACS by filtering out patients with elevated troponin I levels, the exclusion criteria may not have entirely eliminated the possibility of ACS, such as unstable angina. To further validate the model, a prospective study is essential. Third, owing to hospital policies and ethical committee regulations, the current model is not open source, which affects the immediate utility and reproducibility of our algorithm. We are waiting for pending institutional and regulatory approvals to enable other researchers to validate our findings. Interested parties are encouraged to contact us to explore project-based agreements for sharing the model. Last, our AI model, based on a CNN, produces probabilities from 0 to 1, indicating the likelihood of undergoing coronary revascularization. The large dataset and the presence of missing data made it challenging to generate complete pre-test probabilities for direct comparison. By acknowledging and mitigating these constraints through prospective studies and diversifying datasets, we can enhance the model's reliability and applicability in real-world clinical practice.

Conclusion

This study introduces an AI-enhanced ECG algorithm designed to predict the risk of coronary revascularization within 1 year in the general population. The algorithm provides a cost-effective method for cardiovascular risk stratification without altering current clinical practices. By analyzing existing diagnostic data, the model identifies subtle ECG changes, potentially enhancing early detection of severe CAD and improving patient outcomes. Furthermore, it holds promise as an opportunistic screening tool, particularly in remote or resource-limited areas with restricted access to medical care. However, the retrospective nature of this study and the need for prospective validation are acknowledged as limitations. These steps are critical to ensuring the model's reliability and broad applicability. Despite these challenges, the findings underscore the transformative potential of AI in advancing cardiovascular care.

Supplementary material

Supplementary material is available at European Heart Journal – Digital Health.

Author contributions

Chiao-Hsiang Chang (MD), Chin-Sheng Lin (MD, PhD), Chun-Ho Lee (MS), Chin Lin (PhD), Chiao-Chin Lee (MD), Wei-Ting Liu (MD), Yung-Tsai Lee, and Dung-Jang Tsai

Funding

This study was supported by funding from the National Science and Technology Council (NSTC 114-2321-B-016-005 and NSTC 112-2222-E-016-001-MY2 to D.-J.T.) and the Cheng Hsin General Hospital, Taiwan (CHNDMC-113-11205 and CHNDMC-114-11205 to C.L.).

Data availability

The patient data are not publicly available due to privacy. For future research, researchers may access the ethical review from Tri-Service General Hospital upon request of the corresponding author.