Abstract
Background
Women with suspected acute myocardial infarction (AMI) are at increased risk of delayed or missed diagnosis. Artificial intelligence–enabled electrocardiogram (AI-ECG) may support earlier AMI detection, but prospective evidence for sex-consistent performance and rule-out safety is limited.
Objectives
The purpose of this study was to evaluate sex-stratified diagnostic performance, rule-out safety, and phenotype robustness of an AI-ECG for AMI in a prospective, multicenter emergency department cohort.
Methods
ROMIAE (Rule-Out Acute Myocardial Infarction Using Artificial Intelligence Electrocardiogram Analysis) was a prospective external validation study at 18 centers in South Korea. Adults with suspected AMI were enrolled. AI-ECG analyzed initial ECGs using prespecified risk thresholds. Diagnostic performance was assessed using the area under the receiver-operating characteristic curve. Rule-out safety was evaluated by sensitivity, negative predictive value, and missed AMI rate at the low-risk cutoff. All analyses were sex-stratified, with prespecified subgroup analyses.
Results
Among 8,493 patients, 3,186 were women. AI-ECG demonstrated comparable discrimination in women and men (area under the receiver-operating characteristic curve 0.875 [95% CI: 0.853-0.896] vs 0.871 [95% CI: 0.858-0.883]). At the prespecified low-risk cutoff, rule-out sensitivity was 98.8% (95% CI: 97.0-99.5) in women and 99.8% (95% CI: 99.4-100.0) in men, with high negative predictive value (99.2% [95% CI: 97.9-99.7] in women and 99.1% [95% CI: 96.7-99.7] in men) and low missed AMI rates (<1% in both sexes). Performance remained stable across ST-segment elevation and non–ST-segment elevation myocardial infarction, without degradation in women.
Conclusions
In a prospective, multicenter emergency department cohort, AI-ECG demonstrated sex-consistent diagnostic performance and preserved rule-out safety for AMI. Further validation and implementation studies are warranted. (ROMIAE [Rule-Out Acute Myocardial Infarction Using Artificial Intelligence Electrocardiogram Analysis] Trial; NCT05435391)
Key words: acute myocardial infarction, AI-ECG, artificial intelligence, electrocardiogram
Central Illustration
Acute myocardial infarction (AMI) remains a leading cause of morbidity and mortality worldwide.1 Despite advances in diagnostic strategies, women remain at increased risk of delayed or missed AMI diagnoses, which has been associated with worse clinical outcomes.2,3 Compared with men, women more frequently present with atypical symptoms, non–ST-elevation phenotypes, and less overt electrocardiographic abnormalities at initial evaluation, particularly on early electrocardiogram (ECG) assessment, factors that increase vulnerability to underrecognition in acute care settings.
Current diagnostic pathways for suspected AMI rely largely on physician interpretation, rule-based clinical risk scores, and biomarker testing.4,5 Although effective in many clinical contexts, these approaches may be influenced by cognitive bias and delays inherent to biomarker-based confirmation. Such limitations may disproportionately affect populations already at risk of underdiagnosis, highlighting the need for diagnostic tools that are objective and rapidly deployable.6,7
Artificial intelligence (AI)-ECG enables automated AMI risk assessment independent of presenting symptoms or provider judgment.8, 9, 10, 11 In particular, prospective evidence is limited regarding its ability to preserve rule-out safety—such as high sensitivity and negative predictive value (NPV)—and consistent performance across sexes in clinically challenging phenotypes.
In this prespecified secondary analysis of the ROMIAE (Rule-Out Acute Myocardial Infarction Using Artificial Intelligence Electrocardiogram Analysis) cohort, we evaluated sex-stratified diagnostic performance and rule-out safety of AI-ECG for AMI, extending the findings of the primary ROMIAE publication.12
Methods
Study design and data source
The ROMIAE study was a prospective, multicenter external validation study designed to evaluate the diagnostic performance of an AI-ECG model for AMI.12 The study was conducted across 18 emergency departments (EDs) in South Korea between March 2022 and October 2023. The study protocol was approved by the Institutional Review Board at each participating site, and all participants provided written informed consent prior to enrollment. The trial was registered at ClinicalTrials.gov (NCT05435391). The present study represents a prespecified secondary analysis of the ROMIAE cohort focusing on sex-stratified diagnostic performance and subgroup robustness of AI-ECG.13
Study population
Adult patients (age ≥18 years) presenting to the ED with clinically suspected AMI were eligible for inclusion. Patients were enrolled if they presented within 24 hours of symptom onset or experienced worsening or recurrent chest pain within 24 hours prior to or after ED arrival. Patients were excluded if they presented with out-of-hospital cardiac arrest, traumatic chest pain, declined participation, or had a clear noncardiac diagnosis such as pneumothorax. Inclusion and exclusion criteria were applied uniformly across sexes. Biological sex was defined according to electronic medical record documentation at the time of presentation.
AI-ECG model
The AI-ECG model evaluated in this study was based on a previously developed deep learning architecture (AiTiAMI, version 1.00.00; Medical AI Co, Ltd). The model processes raw 12-lead ECG signals sampled at 500 Hz and generates a continuous probability score (0-100) representing the likelihood of AMI.12
The model was trained using multicenter ECG data sets from institutions not participating in the present validation cohort, ensuring complete separation between training and validation data. No clinical variables, laboratory values, or demographic characteristics—including sex—were incorporated as explicit model inputs. Detailed descriptions of the model architecture and training procedures have been reported previously in the original ROMIAE study.12 Prespecified thresholds were used to categorize patients into low-, intermediate-, and high-risk groups. A score <3.0 defined the low-risk category, selected a priori to achieve approximately 99% sensitivity, while a score ≥48.5 defined the high-risk category, corresponding to approximately 90% specificity in the development data set.14,15
Data collection and outcome adjudication
Clinical care was not influenced by study participation. All participating centers followed contemporary international guidelines for the evaluation of suspected AMI using high-sensitivity troponin–based diagnostic algorithms.4,5 Prospectively collected data included the initial 12-lead ECG, high-sensitivity troponin levels, presenting symptoms, and physician-estimated clinical probability of AMI after initial evaluation (Physician AMI score, 0-10 scale). Follow-up ECGs and serial biomarker testing were performed at the treating team’s discretion.
The primary outcome was a diagnosis of AMI during the index hospitalization, including type 1 and type 2 MI, according to the Fourth Universal Definition of Myocardial Infarction.15 Outcome adjudication was performed by 2 board-certified emergency physicians at each site, with a third adjudicator resolving disagreements. Adjudicators were blinded to AI-ECG results.
Statistical analysis
Baseline characteristics were summarized using descriptive statistics. Continuous variables were reported as medians with IQRs and compared between groups using the Mann-Whitney U test. Categorical variables were reported as frequencies and percentages and compared using the chi-square test or Fisher exact test, as appropriate.
The diagnostic performance of the AI-ECG model was evaluated for the primary outcome of AMI diagnosis. All analyses were performed using prospectively collected data according to prespecified analytic plans.13 Discrimination was assessed using the area under the receiver-operating characteristic curve (AUROC) and the area under the precision–recall curve. Sensitivity, specificity, positive predictive value (PPV), and NPV were calculated at prespecified thresholds or optimal cutoffs determined by Youden’s J statistic. CIs for all performance metrics were estimated using bootstrap resampling. Rule-out safety at the prespecified low-risk cutoff was assessed using sensitivity, NPV, and the missed AMI rate, defined as the proportion of patients with a final diagnosis of AMI among those classified as low risk by AI-ECG.
To contextualize AI-ECG performance within established diagnostic strategies, prespecified comparative analyses were performed against commonly used clinical assessment tools. These included the Physician AMI score, History, ECG, Age, Risk factors, and Troponin (HEART) score, GRACE (Global Registry of Acute Coronary Events) 2.0 score, and initial high-sensitivity troponin levels, reflecting routine ED practice.16,17 For ST-segment elevation myocardial infarction (STEMI), where early diagnostic decisions are primarily driven by electrocardiographic interpretation and physician assessment, AI-ECG performance was compared directly with the Physician AMI score. For non–STEMI (NSTEMI), a diagnostically more heterogeneous phenotype, AI-ECG performance was compared with the Physician AMI score, HEART score, GRACE score, and initial high-sensitivity troponin levels. Comparisons of AUROC between the AI-ECG model and each competing diagnostic tool were performed using DeLong's test for correlated receiver-operating characteristic curves, applied separately within each sex and MI phenotype subgroup. Comparisons of other performance metrics, including sensitivity, specificity, PPV, and NPV, were conducted using bootstrap resampling with 1,000 iterations.
To assess diagnostic equity and robustness, all primary analyses were conducted in a sex-stratified manner. Prespecified subgroup analyses were performed across MI phenotype, symptom type, ECG rhythm, MI subtype, and culprit coronary artery, as detailed in the study protocol. Within each subgroup, diagnostic performance metrics were estimated using identical analytic methods. Comparisons of AUROC between women and men within each prespecified clinical subgroup were performed using DeLong’s test. For MI subtype and culprit artery subgroups, heterogeneity of AUROC across categories was assessed using likelihood ratio tests comparing models with and without subgroup indicator variables.
To formally assess risk prediction performance beyond discrimination, sex-stratified calibration analyses and decision curve analysis were performed. Calibration was evaluated using smoothed calibration curves estimated by locally weighted scatterplot smoothing with 95% CIs. Calibration-in-the-large (intercept) and calibration slope were estimated by fitting a logistic regression model of observed outcomes on the logit-transformed predicted probabilities, with the intercept assessed after fixing the slope to 1.0 and the slope estimated in an unconstrained model. Overall probabilistic accuracy was quantified using the Brier score. Decision curve analysis was performed to assess the net clinical benefit of AI-ECG across a range of threshold probabilities, with comparators including physician-estimated AMI probability, HEART score, GRACE score, and initial high-sensitivity troponin. All calibration and decision curve analyses were conducted sex-stratified.
All statistical analyses were performed using R software (version 4.3.2) and Python (version 3.9.7). Statistical significance was assessed using 2-sided tests with a threshold of P < 0.05. Multiple-comparison adjustment was not performed because subgroup analyses were intended to evaluate performance robustness rather than formal hypothesis testing. This study was reported in accordance with the European Heart Rhythm Association AI reporting checklist for AI–based medical research and the TRIPOD + AI (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence) reporting guidelines18 (Supplemental Tables 1 and 2).
Results
Study population and baseline characteristics by sex
Baseline demographic and clinical characteristics stratified by sex are summarized in Table 1. Women were older at presentation and more frequently reported atypical chest pain compared with men (35.1% [1,118/3,186] vs 30.8% [1,632/5,307]). Men more commonly had diabetes mellitus, dyslipidemia, coronary artery disease, and prior MI, whereas congestive heart failure was more prevalent among women. The index AMI diagnosis was less prevalent among women than men (10.5% [335/3,186] vs 23.6% [1,251/5,307]). Among AMI cases, STEMI was more common in men, whereas NSTEMI accounted for a relatively larger proportion of cases among women.
Table 1.
Baseline Characteristics by Sex
| ROMIAE |
P Value | ||
|---|---|---|---|
| Women (n = 3,186) | Men (n = 5,307) | ||
| Demographics | |||
| Age, y | 64.0 (52.0-75.0) | 61.0 (50.0-70.0) | <0.001 |
| Past history | |||
| Hypertension | 1,456 (45.7) | 2,467 (46.5) | 0.496 |
| Diabetes mellitus | 740 (23.2) | 1,377 (25.9) | 0.005 |
| Hyperlipidemia | 931 (29.2) | 1,382 (26.0) | 0.002 |
| Coronary artery disease | 576 (18.1) | 1,186 (22.3) | <0.001 |
| Acute myocardial infarction | 198 (6.2) | 628 (11.8) | <0.001 |
| Congestive heart failure | 146 (4.6) | 156 (2.9) | <0.001 |
| Chronic kidney disease | 201 (6.3) | 347 (6.5) | 0.710 |
| ER chief complaints | <0.001 | ||
| Typical chest pain | 946 (29.7) | 2,192 (41.3) | |
| Atypical chest pain | 1,118 (35.1) | 1,632 (30.8) | |
| Noncardiac chest pain | 318 (10.0) | 441 (8.3) | |
| Other symptoms | 804 (25.2) | 1,042 (19.6) | |
| Chest pain onset ≤3 h | 1,596 (50.1) | 2,760 (52.0) | 0.092 |
| Physician ECG score | 2.0 (2.0-4.0) | 3.0 (2.0-6.0) | <0.001 |
| HEART score | 4.0 (3.0-5.0) | 5.0 (3.0-6.0) | <0.001 |
| GRACE score | 84.0 (60.0-111.0) | 84.0 (61.0-107.0) | 0.202 |
| AI-ECG score | 12.1 (4.5-31.8) | 22.7 (9.9-50.4) | <0.001 |
| Type 1 | 65.2 (39.6-86.2) | 77.0 (46.6-95.0) | <0.001 |
| Type 2 | 20.1 (11.2-66.9) | 47.3 (16.1-72.4) | 0.287 |
| STEMI | 85.3 (67.6-97.1) | 95.7 (79.1-98.0) | <0.001 |
| NSTEMI | 53.8 (26.5-72.6) | 54.6 (27.6-78.5) | 0.168 |
| AMI type | <0.001 | ||
| Non-AMI | 2,851 (89.5) | 4,056 (76.4) | |
| Type 1 | 312 (9.8) | 1,192 (22.5) | |
| Type 2 | 23 (0.7) | 59 (1.1) | |
| ACS type | <0.001 | ||
| Non-ACS | 2,685 (84.3) | 3,646 (68.7) | |
| Unstable angina | 166 (5.2) | 410 (7.7) | |
| STEMI | 104 (3.3) | 538 (10.1) | |
| NSTEMI | 231 (7.3) | 713 (13.4) | |
Values are median (IQR) or n (%).
ACS = acute coronary syndrome; AI-ECG = artificial intelligence–enabled electrocardiogram; AMI = acute myocardial infarction; ER = emergency room; HEART = History, ECG, Age, Risk factors, and Troponin; GRACE = Global Registry of Acute Coronary Events; NSTEMI = non–ST-segment elevation myocardial infarction; ROMIAE = Rule-Out Acute Myocardial Infarction Using Artificial Intelligence Electrocardiogram Analysis; STEMI = ST-segment elevation myocardial infarction.
Sex-stratified discrimination and risk stratification performance of AI-ECG for AMI
AI-ECG demonstrated comparable discrimination performance for AMI in women and men (Table 2, Figure 1). The AUROC was 0.875 (95% CI: 0.853-0.896) in women and 0.871 (95% CI: 0.858-0.883) in men, with no significant difference between groups (DeLong’s test, P = 0.71). The area under the precision–recall curve was 0.593 (95% CI: 0.542-0.646) in women and 0.763 (95% CI: 0.741-0.784) in men.
Table 2.
Sex-Stratified Diagnostic Performance of AI-ECG
| Women (n = 3,186) | Men (n = 5,307) | |
|---|---|---|
| AMI events, n (%) | 335 (10.5%) | 1,251 (23.6%) |
| Discrimination | ||
| AUROC (95% CI) | 0.875 (0.853-0.896) | 0.871 (0.858-0.883) |
| AUPRC (95% CI) | 0.593 (0.542-0.646) | 0.763 (0.741-0.784) |
| Rule-out safety (low-risk cutoff)a | ||
| Sensitivity (%) | 98.8% (97.0-99.5) | 99.8% (99.4-100.0) |
| Negative predictive value (%) | 99.2% (97.9-99.7) | 99.1% (96.7-99.7) |
| Missed AMI rate (%)b | 0.83% (4/480) | 0.92% (2/217) |
| Rule-in context (high-risk cutoff)a | ||
| Specificity (%) | 91.3% (90.3-92.3) | 87.9% (86.9-88.8) |
| Positive predictive value (%) | 47.6% (43.3-52.1) | 64.8% (62.4-67.2) |
Performance metrics were calculated at the prespecified low-risk cutoff defined in the study protocol.
AUPRC = area under the precision–recall curve; AUROC = area under the receiver operating characteristic curve; other abbreviations as in Table 1.
The prespecified low- and high-risk cutoffs (<3.0 and ≥48.5) were defined a priori during model development to achieve approximately 99% sensitivity and 90% specificity, respectively, in the development data set.
Missed AMI rate was defined as the proportion of patients with a final diagnosis of acute myocardial infarction among those classified as low risk by AI-ECG. The denominator represents the total number of patients classified as low risk within each sex.
Figure 1.
Sex-Stratified Discrimination Performance of Artificial Intelligence–Enabled Electrocardiogram for Acute Myocardial Infarction
Receiver-operating characteristic curves with 95% CIs comparing the discrimination performance of artificial intelligence–enabled electrocardiogram for acute myocardial infarction in women and men. AUC = area under the receiver operating characteristic curve.
Beyond discrimination, AI-ECG provided effective risk stratification across predefined risk categories (Figure 2, Central Illustration). In women, the prevalence of AMI increased stepwise across AI-ECG risk categories (0.8%, 4.8%, and 47.5% in the low-, intermediate-, and high-risk groups, respectively). A similar enrichment pattern was observed in men (0.9%, 9.4%, and 64.7%). Among AMI-positive cases, NSTEMI accounted for a larger proportion of total AMI in women than in men in both the intermediate- (93.5% vs 88.4%) and high-risk (56.5% vs 44.8%) categories.
Figure 2.
Sex-Specific Risk Stratification Gradient and Acute Myocardial Infarction Subtype Distribution of Artificial Intelligence–Enabled Electrocardiogram
(A) Women. (B) Men. Stacked bar plots show the distribution of non–acute myocardial infarction, non–ST-segment elevation myocardial infarction, and ST-segment elevation myocardial infarction across artificial intelligence–enabled electrocardiogram–defined low-, intermediate-, and high-risk categories, presented separately for women and men. In both sexes, the prevalence of acute myocardial infarction increases stepwise from the low-risk to the high-risk category. Within the intermediate- and high-risk groups, non–ST-segment elevation myocardial infarction accounts for a larger proportion of total acute myocardial infarction in women than in men. AMI = acute myocardial infarction; NSTEMI = non–ST-segment elevation myocardial infarction; STEMI = ST-segment elevation myocardial infarction.
Central Illustration.
Artificial Intelligence–Enabled Electrocardiogram for Early Triage of Suspected Acute Myocardial Infarction
This schematic contrasts the conventional diagnostic pathway for suspected acute myocardial infarction with an artificial intelligence–enabled electrocardiogram–based triage approach. Traditional evaluation relies on symptom interpretation, clinician assessment, and serial cardiac biomarker testing, which can introduce diagnostic delay and uncertainty—particularly in women, who more often present with atypical symptoms or less overt electrocardiographic findings. In contrast, artificial intelligence–enabled electrocardiogram provides an immediate, objective, time-zero assessment based solely on electrocardiographic signals at the time of presentation, independent of symptom interpretation or biomarker availability. In the prospective multicenter ROMIAE cohort, artificial intelligence–enabled electrocardiogram demonstrated sex-consistent discrimination and preserved rule-out safety, with low missed acute myocardial infarction rates at the prespecified low-risk threshold. Performance remained stable across myocardial infarction phenotypes, including ST-segment elevation myocardial infarction and non–ST-segment elevation myocardial infarction. These findings support artificial intelligence–enabled electrocardiogram as a phenotype-agnostic early triage adjunct that may improve diagnostic efficiency and equity at first clinical contact. AI-ECG = artificial intelligence–enabled electrocardiogram; AUROC = area under the receiver operating characteristic curve; ED = emergency department; MI = myocardial infarction; other abbreviations as in Figure 2.
To formally assess risk prediction performance beyond discrimination, calibration analyses and decision curve analysis were performed sex-stratified. Calibration results and decision curve analysis are presented in Supplemental Figures 1 and 2.
Rule-out safety of AI-ECG at the prespecified low-risk cutoff
As shown in Table 2, at the prespecified low-risk cutoff, AI-ECG demonstrated preserved rule-out safety in both women and men. Sensitivity was 98.8% (95% CI: 97.0-99.5) in women and 99.8% (95% CI: 99.4-100.0) in men, indicating a low false-negative rate. NPV was similarly high in both sexes, at 99.2% (95% CI: 97.9-99.7) in women and 99.1% (95% CI: 96.7-99.7) in men. In addition, at the same low-risk cutoff, AI-ECG classified a larger proportion of women as low risk compared with men (15.1% [95% CI: 13.9%-16.4%; 480/3186] vs 4.1% [95% CI: 3.6%-4.7%; 217/5307]).
The miss rate of AI-ECG was <1% in both women and men, occurring in 0.83% (95% CI: 0.23-2.12%; 4/480) of women and 0.92% (95% CI: 0.11-3.29%; 2/217) of men. These results demonstrate rule-out safety of AI-ECG at the prespecified low-risk threshold, supporting its potential utility as an early triage adjunct (Figure 3).
Figure 3.
Rule-Out Safety of Artificial Intelligence–Enabled Electrocardiogram by Sex at the Prespecified Low-Risk Cutoff
The figure shows sensitivity and negative predictive value for women and men at the prespecified low-risk cutoff. Points indicate point estimates and horizontal lines represent 95% CIs. The right panel shows the missed acute myocardial infarction rate by sex, with bars representing point estimates and error bars indicating 95% CIs. Missed acute myocardial infarction rates were below 1% in both women and men, indicating preserved rule-out safety across sexes. Abbreviation as in Figure 2.
Robustness of AI-ECG performance across prespecified clinical subgroups
The robustness of AI-ECG performance was evaluated across AMI phenotypes, with a particular focus on differences between STEMI and NSTEMI, as summarized in Table 3. In patients with STEMI, AI-ECG demonstrated high discrimination performance and preserved safety metrics and showed numerically comparable or higher performance across multiple metrics compared with the physician-estimated clinical probability of AMI. In contrast, within patients with NSTEMI, AI-ECG maintained stable discrimination and demonstrated overall comparable performance relative to the HEART score, physician-estimated probability of AMI, and initial high-sensitivity troponin I or T. No systematic degradation of AI-ECG performance was observed across discrimination, sensitivity, specificity, or NPV.
Table 3.
Performance for STEMI and NSTEMI
| AUROC (95% CI) | P Value | Sensitivity (95% CI) | P Value | Specificity (95% CI) | P Value | PPV (95% CI) | P Value | NPV (95% CI) | P Value | AUPRC (95% CI) | P Value | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Women | ||||||||||||
| STEMI | ||||||||||||
| AI-ECG | 0.973 (0.960-0.983) | – | 0.931 (0.877-0.974) | – | 0.914 (0.904-0.924) | – | 0.281 (0.233-0.330) | – | 0.997 (0.995-0.999) | – | 0.691 (0.596-0.777) | – |
| Physician AMI score | 0.971 (0.949-0.986) | 0.834 | 0.874 (0.804-0.938) | 0.157 | 0.969 (0.963-0.976) | <0.001 | 0.506 (0.430-0.585) | <0.001 | 0.995 (0.992-0.998) | 0.198 | 0.753 (0.677-0.830) | 0.654 |
| NSTEMI | ||||||||||||
| AI-ECG | 0.831 (0.801-0.859) | – | 0.547 (0.480-0.609) | – | 0.914 (0.904-0.924) | – | 0.339 (0.290-0.389) | – | 0.961 (0.954-0.968) | – | 0.372 (0.311-0.435) | – |
| HEART score | 0.880 (0.858-0.898) | 0.001 | 0.568 (0.502-0.628) | 0.622 | 0.927 (0.918-0.936) | 0.037 | 0.386 (0.333-0.436) | 0.067 | 0.964 (0.956-0.970) | 0.524 | 0.387 (0.324-0.455) | 0.840 |
| Physician AMI score | 0.781 (0.748-0.814) | 0.01 | 0.192 (0.141-0.244) | <0.001 | 0.969 (0.963-0.975) | <0.001 | 0.334 (0.252-0.415) | 0.892 | 0.937 (0.928-0.945) | <0.001 | 0.264 (0.210-0.318) | <0.001 |
| hs-Troponin I or T | 0.851 (0.827-0.873) | 0.231 | 0.857 (0.810-0.900) | <0.001 | 0.844 (0.831-0.857) | <0.001 | 0.308 (0.271-0.343) | 0.130 | 0.987 (0.982-0.991) | <0.001 | 0.588 (0.555-0.619) | <0.001 |
| GRACE score | 0.755 (0.728-0.782) | <0.001 | 0.200 (0.148-0.250) | <0.001 | 0.931 (0.921-0.940) | 0.007 | 0.189 (0.141-0.235) | <0.001 | 0.935 (0.926-0.944) | <0.001 | 0.196 (0.152-0.240) | <0.001 |
| Men | ||||||||||||
| STEMI | ||||||||||||
| AI-ECG | 0.968 (0.960-0.975) | – | 0.926 (0.904-0.947) | – | 0.879 (0.869-0.888) | – | 0.503 (0.470-0.535) | – | 0.989 (0.985-0.992) | – | 0.871 (0.845-0.893) | – |
| Physician AMI score | 0.960 (0.950-0.969) | 0.159 | 0.862 (0.832-0.890) | <0.001 | 0.945 (0.938-0.952) | <0.001 | 0.675 (0.640-0.710) | <0.001 | 0.981 (0.977-0.985) | 0.001 | 0.845 (0.816-0.871) | 0.003 |
| NSTEMI | ||||||||||||
| AI-ECG | 0.797 (0.778-0.815) | – | 0.568 (0.531-0.606) | – | 0.879 (0.868-0.889) | – | 0.451 (0.417-0.485) | – | 0.921 (0.912-0.929) | – | 0.469 (0.426-0.510) | – |
| HEART score | 0.835 (0.821-0.850) | <0.001 | 0.483 (0.447-0.519) | <0.001 | 0.922 (0.913-0.929) | <0.001 | 0.519 (0.481-0.557) | <0.001 | 0.910 (0.902-0.919) | 0.014 | 0.511 (0.471-0.550) | 0.252 |
| Physician AMI score | 0.751 (0.731-0.769) | <0.001 | 0.271 (0.238-0.303) | <0.001 | 0.945 (0.938-0.952) | <0.001 | 0.465 (0.417-0.509) | 0.602 | 0.881 (0.871-0.891) | <0.001 | 0.373 (0.338-0.411) | <0.001 |
| hs-Troponin I or T | 0.804 (0.787-0.819) | 0.545 | 0.772 (0.740-0.802) | <0.001 | 0.835 (0.825-0.846) | <0.001 | 0.452 (0.424-0.479) | 0.955 | 0.954 (0.947-0.961) | <0.001 | 0.629 (0.605-0.650) | <0.001 |
| GRACE score | 0.667 (0.646-0.688) | <0.001 | 0.098 (0.078-0.121) | <0.001 | 0.959 (0.953-0.965) | <0.001 | 0.297 (0.238-0.355) | <0.001 | 0.858 (0.848-0.869) | <0.001 | 0.242 (0.217-0.269) | <0.001 |
Threshold-dependent metrics were calculated at predefined high-risk cutoffs: AI-ECG at the prespecified cutoff; HEART score, physician AMI score, and GRACE score at the high-risk cutoffs used in the ROMIAE study; and hs-troponin at the 99th-percentile URL. Performance metrics are shown for completeness; interpretation focuses on discrimination and rule-out safety.
P values indicate comparisons with AI-ECG within the same sex and myocardial infarction phenotype.
Beyond MI phenotype, AI-ECG performance remained generally consistent across prespecified clinical subgroups, including demographic characteristics, symptom presentation, electrocardiographic rhythm, MI subtype, and culprit coronary artery, as shown in Supplemental Tables 3 and 4. Although numerical differences were observed in selected subgroups, overall discrimination remained stable across most subgroups in both sexes. Heterogeneity analyses demonstrated statistically significant differences across selected MI phenotypes and culprit artery subgroups, with a general trend toward lower discrimination in more complex or non–ST-elevation presentations. Additionally, performance was attenuated in patients with prior coronary artery disease, prior AMI, or congestive heart failure, consistent with the known impact of underlying cardiac disease on ECG-based interpretation.
Discussion
In this prespecified secondary analysis of the prospective multicenter ROMIAE cohort, we demonstrate that AI-ECG provides sex-consistent diagnostic performance for AMI, extending the findings of the primary ROMIAE study. AI-ECG showed comparable discrimination in women and men and preserved rule-out safety in both sexes. In addition, diagnostic performance remained robust across myocardial infarction phenotypes, including STEMI and NSTEMI, supporting a phenotype-agnostic diagnostic capability. Together, these findings provide prospective evidence that AI-ECG may support stable diagnostic performance across sexes and could enable more equitable early triage in patients presenting with suspected AMI.
Sex-consistent and safe diagnostic performance
Women with suspected AMI have historically faced structural challenges in timely and accurate diagnosis, driven by higher rates of atypical symptom presentation and less overt ECG findings.2,19,20 Diagnostic pathways relying heavily on symptom interpretation, clinical risk scores, and biomarker confirmation may contribute to delayed recognition in women. In this context, our findings demonstrate that AI-ECG provides comparable discrimination and safety in women and men, with preserved rule-out safety and similarly low missed AMI rates. In women, this was accompanied by meaningful rule-out efficiency, allowing a larger proportion to be safely classified as low risk at presentation. Rule-in performance showed sex-related numerical differences, which can be interpreted in the context of differing AMI prevalence.
Notably, the ROMIAE cohort comprised a higher proportion of male patients (62%), with a correspondingly higher AMI prevalence in men. This sex-related difference in prevalence is the primary driver of the observed PPV disparity between sexes, as PPV is inherently dependent on disease prevalence rather than model performance per se. As a result, the safety estimates derived in women—including sensitivity and NPV—should be interpreted with this context in mind, as they should be interpreted in the context of the cohort's sex composition, as performance metrics are inherently influenced by AMI prevalence within the study population.
From a workflow perspective, these results suggest that AI-ECG may serve as an early triage tool in the ED, enabling risk stratification based on ECG alone at an early stage of evaluation.8 By operating independently of biomarker availability, AI-ECG has the potential to support safe rule-out decisions before laboratory confirmation, which may be particularly relevant in women presenting with suspected AMI.21
When evaluating rule-out strategies in the ED, the missed AMI rate is an important safety metric. In the present study, AI-ECG demonstrated a missed AMI rate below 1% in both sexes at the prespecified low-risk threshold. However, certain presentations—such as ischemic injury without overt biomarker elevation, including mental stress–induced myocardial ischemia, which has been reported to be more prevalent in women—may not be reliably captured by ECG-based assessment alone.22 These findings underscore that AI-ECG should be interpreted as a supportive early triage tool, and that clinical judgment and serial biomarker assessment remain essential. Furthermore, the relatively low prevalence of type 2 MI in this cohort, representing approximately 5% of AMI cases, should be considered when interpreting safety estimates, as the observed results are largely driven by type 1 MI events.
Robustness across phenotypes in the context of prior diagnostic strategies
NSTEMI represents a diagnostically challenging phenotype, especially in women, given their more frequent atypical presentations and less overt ECG findings.19,23,24 Importantly, AI-ECG performance in women with NSTEMI remained clinically robust, with meaningful discrimination and high NPV. In addition, AI-ECG performance remained consistent across MI subtypes and symptom presentation patterns, including typical vs atypical symptoms and timing of symptom onset, as well as across different culprit coronary arteries.
High-sensitivity troponin–based algorithms and clinical risk scores such as GRACE and HEART have been shown to exhibit sex-related classification differences.25, 26, 27 In contrast, AI-ECG in the present study operated independently of biomarker thresholds and clinical scoring systems, while maintaining consistent performance across sexes and MI phenotypes. Notably, this sex-consistent performance was achieved without sex-specific model optimization, suggesting that objective ECG-based assessment may reduce interpretive variability inherent to symptom-driven diagnostic pathways.
Clinical implications
The present analysis extends the primary ROMIAE study by providing several clinically relevant insights. First, AI-ECG preserved rule-out safety specifically in women, with a sensitivity of 98.8% and a missed AMI rate below 1%, supporting its safety as an early triage adjunct in a population historically at risk for underdiagnosis. Second, AI-ECG classified a substantially larger proportion of women as low risk compared with men (15.1% vs 4.1%), suggesting potential workflow efficiency gains in this population. Third, diagnostic performance remained robust across MI phenotypes, including NSTEMI—a clinically challenging presentation disproportionately affecting women—without systematic degradation.
Despite these findings, AI-ECG should be interpreted as a supportive triage tool requiring integration with the full clinical context, not as a standalone diagnostic criterion. Importantly, ischemic presentations from nonatherosclerotic etiologies—including myocardial infarction with nonobstructive coronary arteries, spontaneous coronary artery dissection, and vasospastic angina—may not be reliably captured by ECG-based algorithms, as these entities frequently lack characteristic electrocardiographic abnormalities.28 Women are disproportionately represented among these etiologies, and clinicians should maintain a high index of suspicion for such phenotypes, even when AI-ECG risk scores are low.
Acknowledging these contextual limitations, several prior studies have demonstrated that AI-ECG can meaningfully augment acute coronary syndrome workflows. The ARISE (Artificial Intelligence–Powered Rapid Identification of ST-Elevation Myocardial Infarction via electrocardiogram) trial evaluated AI-ECG as a real-time clinical decision support system for STEMI and showed that AI-ECG-assisted triage reduced treatment delays.29 In parallel, occlusion MI-focused approaches have emphasized detecting acute coronary occlusion among patients without classic ST elevation and have highlighted the value of ECG-only risk scores for both rule-in and rule-out classification and workflow reclassification.9, 10, 11
When considered alongside prior studies, the present analysis complements existing evidence by providing prospective multicenter evidence of sex-consistent diagnostic performance and rule-out safety across MI phenotypes in ED settings.
Study Limitations
Several limitations of this study should be acknowledged. First, this investigation was conducted in a single country, which may limit the generalizability of the findings to health care systems with different patient populations, clinical workflows, or ECG acquisition practices. Second, biological sex was defined based on electronic medical record documentation and does not capture gender identity or sociocultural factors, which may also influence symptom presentation and health care–seeking behavior. Third, the physician-estimated probability of AMI was inherently subjective and may vary across clinicians and institutions, potentially affecting comparative performance assessments. Fourth, prespecified AI-ECG risk thresholds were fixed based on prior model development and were not optimized separately for each subgroup, which may limit flexibility across different clinical contexts. Finally, although outcome adjudication was performed during the index hospitalization, certain clinical presentations without overt biomarker elevation may remain diagnostically challenging. Therefore, the observed missed AMI rate should be interpreted with caution, particularly in women who more frequently present with atypical phenotypes. These findings should be interpreted as supportive evidence, and further external validation across diverse health care systems and prospective implementation studies will be important to determine the real-world clinical impact of AI-ECG-guided triage strategies.30
Conclusions
In this prospective, multicenter external validation study, AI-ECG demonstrated sex-consistent diagnostic performance for AMI, with preserved discrimination and rule-out safety in women. AI-ECG performance remained robust across MI phenotypes, including STEMI and NSTEMI, supporting a phenotype-agnostic diagnostic capability. These findings suggest that AI-ECG may serve as an early triage adjunct in ED settings, with potential to improve diagnostic equity during the initial evaluation of suspected AMI, particularly in women.
Perspectives.
COMPETENCY IN MEDICAL KNOWLEDGE: Women with suspected AMI face an increased risk of diagnostic delay and underrecognition due to atypical presentations and less overt ECG findings. In a prospective, multicenter ED cohort, AI-ECG demonstrated sex-consistent diagnostic performance and preserved rule-out safety.
TRANSLATIONAL OUTLOOK: AI-ECG may serve as an early triage adjunct when biomarker results are pending, with sex-consistent performance supporting integration into ED workflows to improve diagnostic equity. Future studies should evaluate real-world implementation across diverse health care systems and assess impacts on clinical decision-making, downstream testing, and outcomes—particularly among women and other high-risk populations.
Funding support and author disclosures
This work has been supported by a nonconditional grant from Medical AI Co, Ltd. This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (RS-2024-00335935). Drs HS Lee, S Kang, JM Son, KS Kim, JM Kwon, and MS Lee are employees of Medical AI Co, Ltd and hold stocks in the company. Dr Kyuseok Kim is a stakeholder in Medical AI Co., Ltd. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.
Footnotes
The authors attest they are in compliance with human studies committees and animal welfare regulations of the authors’ institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the Author Center.
Appendix
For a list of the ROMIAE Study Group as well as supplemental tables and figures, please see the online version of this paper.
Contributor Information
Min Sung Lee, Email: lylm85@gmail.com.
ROMIAE Study Group:
Jong Eun Park, Heajin Chung, Jin Hee Jeong, Sung-Jun Park, Sejoong Ahn, Haewon Jung, Seok Hoon Ko, Han Sung Choi, June-sung Kim, Bangshill Rhee, Keon Kim, So Eun Kim, Ji Hwan Bu, Jae Hwan Kim, Dong Won Kim, Yongil Cho, Yonghee Lee, Kihwan Choi, Sangchun Choi, Jeong Min Son, Sora Kang, Hak Seung Lee, Min-yeong Kim, Nuri Shin, Jong-Hwan Jang, and Yong-Yeon Jo
Supplementary material
References
- 1.Martin S.S., Aday A.W., Almarzooq Z.I., et al. 2024 Heart disease and stroke statistics: a report of US and Global data from the American Heart Association. Circulation. 2024;149:e347–e913. doi: 10.1161/CIR.0000000000001209. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.van Oosterhout R.E., de Boer A.R., Maas A.H., Rutten F.H., Bots M.L., Peters S.A. Sex differences in symptom presentation in acute coronary syndromes: a systematic review and meta-analysis. J Am Heart Assoc. 2020;9 doi: 10.1161/JAHA.119.014733. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.DeFilippis E.M., Collins B.L., Singh A., et al. Women who experience a myocardial infarction at a young age have worse outcomes compared with men: the Mass General Brigham YOUNG-MI registry. Eur Heart J. 2020;41:4127–4137. doi: 10.1093/eurheartj/ehaa662. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Rao S.V., O’Donoghue M.L., Ruel M., et al. 2025 ACC/AHA/ACEP/NAEMSP/SCAI guideline for the management of patients with acute coronary syndromes: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. J Am Coll Cardiol. 2025;85:2135–2237. doi: 10.1016/j.jacc.2024.11.009. [DOI] [PubMed] [Google Scholar]
- 5.Byrne R.A., Rossello X., Coughlan J., et al. 2023 ESC guidelines for the management of acute coronary syndromes. Eur Heart J. 2023;44:3720–3826. doi: 10.1093/eurheartj/ehad191. [DOI] [PubMed] [Google Scholar]
- 6.Kwok C.S., Bennett S., Azam Z., et al. Misdiagnosis of acute myocardial infarction: a systematic review of the literature. Crit Pathw Cardiol. 2021;20:155–162. doi: 10.1097/HPC.0000000000000256. [DOI] [PubMed] [Google Scholar]
- 7.Than M., Herbert M., Flaws D., et al. What is an acceptable risk of major adverse cardiac event in chest pain patients soon after discharge from the Emergency Department?: a clinical survey. Int J Cardiol. 2013;166:752–754. doi: 10.1016/j.ijcard.2012.09.171. [DOI] [PubMed] [Google Scholar]
- 8.Khera R., Oikonomou E.K., Nadkarni G.N., et al. Transforming cardiovascular care with artificial intelligence: from discovery to practice: JACC state-of-the-art review. J Am Coll Cardiol. 2024;84:97–114. doi: 10.1016/j.jacc.2024.05.003. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Herman R., Meyers H.P., Smith S.W., et al. International evaluation of an artificial intelligence–powered electrocardiogram model detecting acute coronary occlusion myocardial infarction. Eur Heart J Digit Health. 2024;5:123–133. doi: 10.1093/ehjdh/ztad074. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Al-Zaiti S.S., Martin-Gill C., Zègre-Hemsey J.K., et al. Machine learning for ECG diagnosis and risk stratification of occlusion myocardial infarction. Nat Med. 2023;29:1804–1813. doi: 10.1038/s41591-023-02396-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Liu W.-C., Lin C.-S., Tsai C.-S., et al. A deep learning algorithm for detecting acute myocardial infarction: deep learning model to detect AMI. EuroIntervention. 2021;17:765. doi: 10.4244/EIJ-D-20-01155. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Lee M.S., Shin T.G., Lee Y., et al. Artificial intelligence applied to electrocardiogram to rule out acute myocardial infarction: the ROMIAE multicentre study. Eur Heart J. 2025;46:1917–1929. doi: 10.1093/eurheartj/ehaf004. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Shin T.G., Lee Y., Kim K., Lee M.S., Kwon J.M., ROMIAE study group ROMIAE (Rule-Out Acute Myocardial Infarction using Artificial Intelligence Electrocardiogram Analysis) trial study protocol: a prospective multicenter observational study for validation of a deep learning–based 12-lead electrocardiogram analysis model for detecting acute myocardial infarction in patients visiting the emergency department. Clin Exp Emerg Med. 2023;10:438. doi: 10.15441/ceem.22.360. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Doudesis D., Lee K.K., Yang J., et al. Validation of the myocardial-ischaemic-injury-index machine learning algorithm to guide the diagnosis of myocardial infarction in a heterogenous population: a prespecified exploratory analysis. Lancet Digit Health. 2022;4:e300–e308. doi: 10.1016/S2589-7500(22)00025-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Lopez-Ayala P., Boeddinghaus J., Nestelberger T., et al. External validation of the myocardial-ischaemic-injury-index machine learning algorithm for the early diagnosis of myocardial infarction: a multicentre cohort study. Lancet Digit Health. 2024;6:e480–e488. doi: 10.1016/S2589-7500(24)00088-8. [DOI] [PubMed] [Google Scholar]
- 16.Backus B., Six A., Kelder J., et al. A prospective validation of the HEART score for chest pain patients at the emergency department. Int J Cardiol. 2013;168:2153–2158. doi: 10.1016/j.ijcard.2013.01.255. [DOI] [PubMed] [Google Scholar]
- 17.Hung J., Roos A., Kadesjö E., et al. Performance of the GRACE 2.0 score in patients with type 1 and type 2 myocardial infarction. Eur Heart J. 2021;42:2552–2561. doi: 10.1093/eurheartj/ehaa375. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Svennberg E., Han J.K., Caiani E.G., et al. State of the art of artificial intelligence in clinical electrophysiology in 2025: a scientific statement of the European Heart Rhythm Association (EHRA) of the ESC, the Heart Rhythm Society (HRS), and the ESC Working Group on E-Cardiology. Europace. 2025;27:euaf071. doi: 10.1093/europace/euaf071. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Khraishah H., Daher R., Garelnabi M., Karere G., Welty F.K. Sex, racial, and ethnic disparities in acute coronary syndrome: novel risk factors and recommendations for earlier diagnosis to improve outcomes. Arterioscler Thromb Vasc Biol. 2023;43:1369–1383. doi: 10.1161/ATVBAHA.123.319370. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Redfors B., AngerAas O., RAamunddal T., et al. Trends in gender differences in cardiac care and outcome after acute myocardial infarction in Western Sweden: a report from the Swedish web system for enhancement of evidence-based care in heart disease evaluated according to recommended therapies (SWEDEHEART) J Am Heart Assoc. 2015;4 doi: 10.1161/JAHA.115.001995. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Giordano V., Guillari A., Sansone V., Catone M., Rea T. Women acute myocardial infarction—identifying and understanding the gender gap (WAMy-GAP): a study protocol. Healthcare MDPI. 2024;12:972. doi: 10.3390/healthcare12100972. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Vaccarino V., Almuwaqqat Z., Kim J.H., et al. Association of mental stress–induced myocardial ischemia with cardiovascular events in patients with coronary heart disease. JAMA. 2021;326:1818–1828. doi: 10.1001/jama.2021.17649. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Lichtman J.H., Leifheit E.C., Safdar B., et al. Sex differences in the presentation and perception of symptoms among young patients with myocardial infarction: evidence from the VIRGO study (variation in recovery: role of gender on outcomes of young AMI patients) Circulation. 2018;137:781–790. doi: 10.1161/CIRCULATIONAHA.117.031650. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Safdar B., Spatz E.S., Dreyer R.P., et al. Presentation, clinical profile, and prognosis of young patients with myocardial infarction with nonobstructive coronary arteries (MINOCA): results from the VIRGO study. J Am Heart Assoc. 2018;7 doi: 10.1161/JAHA.118.009174. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Supples M.W., Snavely A.C., O'Neill J.C., et al. Sex and race differences in the performance of the European Society of Cardiology 0/1-h algorithm with high-sensitivity troponin T. Clin Cardiol. 2024;47 doi: 10.1002/clc.24199. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Wenzl F.A., Kraler S., Ambler G., et al. Sex-specific evaluation and redevelopment of the GRACE score in non-ST-segment elevation acute coronary syndromes in populations from the UK and Switzerland: a multinational analysis with external cohort validation. Lancet. 2022;400:744–756. doi: 10.1016/S0140-6736(22)01483-0. [DOI] [PubMed] [Google Scholar]
- 27.Lee K.K., Ferry A.V., Anand A., et al. Sex-specific thresholds of high-sensitivity troponin in patients with suspected acute coronary syndrome. J Am Coll Cardiol. 2019;74:2032–2043. doi: 10.1016/j.jacc.2019.07.082. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Tamis-Holland J.E., Jneid H., Reynolds H.R., et al. Contemporary diagnosis and management of patients with myocardial infarction in the absence of obstructive coronary artery disease: a scientific statement from the American Heart Association. Circulation. 2019;139:e891–e908. doi: 10.1161/CIR.0000000000000670. [DOI] [PubMed] [Google Scholar]
- 29.Lin C., Liu W.-T., Chang C.-H., et al. Artificial intelligence–powered rapid identification of ST-elevation myocardial infarction via electrocardiogram (ARISE)—A pragmatic randomized controlled trial. NEJM AI. 2024;1 [Google Scholar]
- 30.Cha H., Kim J. Ethical considerations of artificial intelligence in emergency medicine for triage and resource allocation: a scoping review. Clin Exp Emerg Med. 2025;12:306. doi: 10.15441/ceem.25.199. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.