Specificity of the Occlusion Myocardial Infarction-Trained PMcardio “Queen of Hearts” AI-ECG in Inflammatory Myopericardial Syndromes Authors: Wesley Bennar, Dorian Garin, Rayane M Saidi, Lisa Simioni, Thaïs Pittet, Serban Puricel, Pascal Meier, Mario Togni, Stéphane Cook Published: 2026 ABSTRACT AbstractBackgroundInflammatory myopericardial syndromes (IMPS)—including pericarditis, myocarditis, and overlap phenotypes—frequently present with chest pain and electrocardiographic abnormalities that can mimic occlusion myocardial infarction (OMI). High electrocardiographic specificity is essential to avoid unnecessary catheterization laboratory activation.ObjectivesTo evaluate the ECG-level specificity of the OMI-trained PMcardio Queen of Hearts (QoH) AI-ECG model for OMI classification in patients with IMPS.MethodsWe conducted a single-centre retrospective study including consecutive emergency department patients with adjudicated IMPS (2010-2025) at the University & Hospital of Fribourg. Obstructive coronary artery disease was excluded by coronary angiography or cardiac magnetic resonance imaging according to contemporary guidelines. Index ECGs were analysed using QoH. Specificity was compared with European Society of Cardiology (ESC) STEMI ECG criteria, the 4-variable formula (when applicable), and 3 large language models (LLMs; ChatGPT 5.2, Claude Opus 4.5, and Gemini 3). The primary endpoint was ECG-level specificity. Because only adjudicated IMPS cases were included, the analysis was designed to estimate ECG-level specificity rather than overall diagnostic accuracy.ResultsAmong 242 IMPS patients, QoH correctly classified 230 as non-OMI (specificity: 95.0%; 95% CI: 91.5-97.1). ESC STEMI ECG criteria showed 78.1% specificity, and the 4-variable formula demonstrated 73.3% specificity in anterior presentations (n = 30). All LLMs exhibited substantially lower specificity (26.9% to 70.2%; all P < 0.001 vs QoH). Discordant ESC-positive/QoH-negative cases (n = 43) predominantly showed diffuse, nonterritorial inflammatory ECG patterns. Specificity was higher in (myo)pericarditis than in (peri)myocarditis (97.7% vs 87.7%, P = 0.004).ConclusionsIn IMPS, the OMI-trained QoH model demonstrated high specificity and fewer false-positive OMI classifications than guideline ECG-STEMI criteria, ECG-based scores, and general-purpose LLMs.Key words: AI-ECG, inflammatory myopericardial syndromes, myocarditis, occlusion myocardial infarction, pericarditis, PMcardio, Queen of Hearts Inflammatory myopericardial syndromes (IMPS)—including pericarditis, myocarditis, and overlap entities such as myopericarditis or perimyocarditis—are frequent non-ischemic causes of acute chest pain. Their ECG manifestations (diffuse ST-segment elevation [STE], PR depression, repolarization abnormalities) can resemble acute coronary occlusion, creating a trade-off between rapid reperfusion and diagnostic accuracy. The 2025 European Society of Cardiology (ESC) guidelines consolidate myocarditis and pericarditis under the IMPS umbrella and advocates a structured, multimodal approach—clinical context, biomarkers, echocardiography, and cardiac magnetic resonance imaging (MRI)—rather than reliance on ECG alone.1 Over the past decade, the occlusion myocardial infarction (OMI) paradigm has reframed acute myocardial infarction triage: a meaningful subset of angiographic occlusions lack classic ST-segment elevation myocardial infarction (STEMI) thresholds, while some STE activations are nonischemic.2, 3, 4, 5, 6, 7, 8 Task-specific artificial intelligence (AI)-ECG models trained on angiographic OMI endpoints, notably PMcardio’s Queen of Hearts (QoH), have outperformed STEMI criteria and expert interpretation across diverse settings.9, 10, 11, 12 In IMPS—where ST-modification is often inflammatory—an important practical question is whether an OMI-focused AI can avoid false OMI alarms. We therefore assessed the specificity of QoH in an IMPS-only cohort and contrasted it with the ESC STEMI ECG criteria, the 4-variable formula when appropriate and 3 general-purpose large language models (LLMs): ChatGPT 5.2, Claude Opus 4.5, and Gemini 3.13,14 Methods Study design and adjudication We performed a retrospective cohort study of consecutive emergency-department presentations between January 2010 and June 2025 at the University & Hospital of Fribourg. Eligible patients presented with chest pain and received a final adjudicated diagnosis of IMPS, including pericarditis, myocarditis, myopericarditis, or perimyocarditis. Final diagnoses were established by 2 experienced cardiologists who independently reviewed all available clinical, laboratory, ECG, and imaging data and were blinded to QoH and LLM outputs. Only cases fulfilling contemporary ESC 2025 diagnostic criteria with sufficient documentation were included, while cases with persistent diagnostic uncertainty, incomplete documentation, or alternative competing diagnoses were excluded (Figure 1).1 Because this study aimed to estimate ECG-level specificity for non-OMI classification in an IMPS-only cohort, particular attention was given to adjudication of noncoronary diagnoses using a guideline-consistent hierarchical strategy.1 All patients with troponin elevation and/or suspected (peri)myocarditis included in the study underwent invasive coronary angiography to exclude obstructive coronary artery disease and/or cardiac MRI supporting the diagnosis of myocarditis through demonstration of a nonischemic injury pattern consistent with IMPS. All invasive coronary angiographies were performed <6 hours from presentation. In pericarditis-only presentations without myocardial injury, systematic invasive coronary angiography or noninvasive cardiac magnetic resonance was not required when the clinical presentation fulfilled diagnostic criteria for acute pericarditis, and no other features suggesting acute coronary occlusion were present. In these cases, adjudication relied on concordant clinical presentation, biomarker profile, ECG evolution, echocardiographic findings, and follow-up supporting a noncoronary inflammatory diagnosis. This approach reflects contemporary ESC-recommended diagnostic pathways in which coronary angiography is performed selectively rather than systematically in patients with suspected inflammatory pericardial syndromes (Supplemental Table 1).1 The study was approved by the Cantonal Research Ethics Committee of Fribourg (003-REP-CER-FR), and written informed consent was obtained. Study Flow Chart and Patient Selection Flow diagram showing the screening, exclusion, and inclusion process of patients presenting with suspected IMPS. Among 393 screened patients, 151 were excluded for predefined criteria, including missing data, lack of confirmatory imaging, alternative diagnoses, incomplete ECGs, or missing informed consent. A total of 242 IMPS-only patients were included and categorized into (myo)pericarditis and (peri)myocarditis subgroups according to diagnostic work-up with cardiac magnetic resonance imaging and/or coronary angiography. Subgroup distributions and diagnostic modalities are detailed. ESC = European Society of Cardiology; IMPS = inflammatory myopericardial syndromes; MINOCA = myocardial infarction with non-obstructive coronary arteries; SCAD = spontaneous coronary artery dissection. ECG processing and index tests The first available 12-lead ECG at presentation was analysed using the QoH application (PM Cardio App, current version 3.4.1). All ECGs were analysed using the QoH algorithm, which is specifically developed to detect OMI (STEMI and STEMI equivalents) from a single 12-lead ECG. Outputs were recorded as OMI positive (STEMI/STEMI equivalent), low-confidence, or negative. For comparison, ESC STEMI ECG criteria were applied, defined as a STE at the J-point in at least 2 contiguous leads ≥2.5 mm in men <40 years, ≥2 mm in men ≥40 years, or ≥1.5 mm in women regardless of age in leads V2–V3 and/or ≥1 mm in the other leads.13 These criteria are consistent with those outlined in the Fourth Universal Definition of Myocardial Infarction.15 All ECGs were independently reviewed by 2 other experienced interventional cardiologists with specific expertise in ischemia detection and ECG interpretation. Reviewers were blinded to clinical data, final diagnosis, and study outcomes, and ECGs were anonymized prior to analysis. In cases of discrepant classification, consensus was reached after joint review. Territory involvement was defined according to the distribution of STE on the index ECG in patients meeting STE criteria following standard lead conventions (V1–V4 anterior; I, aVL, V5–V6 lateral; II, III, aVF inferior). Multiple territories could be assigned when applicable, whereas ECGs without STE criteria were categorized as STE criteria negative. The 4-variable formula was calculated in ECGs meeting its predefined applicability criteria, using four criteria: QTcB duration (ms), STE at 60 ms in lead V3 (mm), R-wave amplitude in lead V4 (mm), QRS amplitude in lead V2 (mm). This formula is applicable when ECG shows ≥1 mm STE in ≥1 of the precordial leads V2-V4 with the following exclusion criteria: >5 mm STE; nonconcave STE; inferior reciprocal changes; anterior ST-segment depression; terminal QRS distortion in lead V2 or V3; Q waves in any of lead V2-V4; and any T-wave inversion from lead V2-V6. The threshold ≥18.2 defined positivity. The equation of the 4-variable formula is provided in the Supplementary Appendix.14,16 This formula was originally developed to differentiate subtle anterior occlusion from early repolarization and was included for contextual comparison. ESC STEMI ECG criteria and the 4-variable formula were applied as ECG-only comparators to reflect early triage decision-making based on electrocardiographic findings rather than as comprehensive diagnostic standards for IMPS. In parallel, all index ECGs were independently analysed using 3 commercially available web-based LLMs chatbot interfaces: ChatGPT 5.2 (OpenAI), Claude Opus 4.5 (Anthropic), and Gemini 3 (Google DeepMind). Analyses were performed on December 11, 2025, using Google Chrome (browser v. 142.0.7444.234) to ensure consistent access conditions and via the publicly available web-based interfaces of ChatGPT 5.2 (OpenAI), Claude Opus 4.5 (Anthropic), and Gemini 3 (Google DeepMind). Exact model build snapshot identifiers were not exposed through the public interfaces and were therefore unavailable for documentation. ECGs were exported from the institutional electronic health record system as anonymized PNG images without preprocessing. Images were reviewed at their native resolution using a standard desktop display environment without resizing, normalization, enhancement, or contrast adjustment prior to submission to the LLM interfaces. Exact pixel dimensions and display scaling parameters were not standardized across cases because ECGs were analyzed as exported from the clinical system. ECGs were analyzed in a new, independent conversation to prevent contextual carryover between cases. Each ECG was evaluated once per model without repeated inference runs. Each LLM was instructed, via a standardized prompt, to return a binary classification (OMI/no OMI) based on ECG interpretation with the sole additional information that the patient presented to the emergency department with chest pain as the presenting complaint (Supplemental Appendix). Accordingly, the LLM evaluation reflected ECG interpretation with minimal clinical context. Responses that unambiguously endorsed OMI (eg, “OMI”, “STEMI-equivalent”, “acute coronary occlusion”) were counted as positive; responses that explicitly denied OMI (“no OMI”, “no evidence of acute occlusion”) were counted as negative. Equivocal or hedged responses (eg, “possible”, “cannot exclude”, “indeterminate”) were, in line with the clinical convention used for QoH low-confidence outputs, counted as negative in the primary analysis. Parameters such as temperature and top-p were not user-configurable through the public chatbot interfaces. Endpoint The primary endpoint was specificity (true-negative rate) for correctly classifying IMPS ECGs as non-OMI. False-positive rate was reported as the complement of specificity (1 − specificity). Secondary endpoints included the effect of reclassifying low-confidence outputs as positive, territory-based performance (anterior/lateral/inferior), stratification by IMPS phenotype ((myo)pericarditis vs (peri)myocarditis), and comparative performance between QoH, LLMs, ESC STEMI ECG criteria, and the 4-variable formula. An exploratory qualitative analysis of discordant ESC-positive/QoH-negative cases was performed to identify recurrent electrocardiographic patterns associated with benign classification. Because this IMPS-only cohort included no OMI cases, analyses were specifically designed to estimate ECG-level specificity and false-positive classification rates rather than overall diagnostic accuracy. Statistical analysis Specificity was the primary performance metric, representing the proportion of IMPS patients correctly classified as non-OMI. Binomial proportions (specificity, false-positive rate) were reported with Wilson score 95% confidence intervals. The Wilson method was preferred over the Wald (normal approximation) interval because specificity values in this cohort approach the upper boundary of the 0 to 1 interval (≥95%), and several subgroup cell counts are small, conditions under which the Wald interval is known to undercover. The Wilson interval provides more accurate nominal coverage and prevents impossible interval bounds (<0 or >1). The specificity comparison of QoH vs ESC STEMI ECG criteria for non-OMI classification in the IMPS cohort was prespecified as the primary analysis. All other comparisons, including pairwise comparisons with LLMs, analyses across ECG territories, and phenotype-based subgroup analyses ((myo)pericarditis vs (peri)myocarditis), were considered secondary and exploratory. No correction for multiplicity was applied; accordingly, P-values outside the primary comparison are reported as descriptive and should be interpreted as hypothesis-generating. All paired comparisons were performed at the patient-ECG level, with one index ECG per patient analysed by all tests. QoH low-confidence outputs were classified as negative (non-OMI) in the primary analysis, and all paired comparisons used this dichotomization; a prespecified sensitivity analysis reclassifying low-confidence outputs as positive is reported separately. This approach reflects real-world triage logic in which catheterization laboratory activation typically requires an explicit OMI-positive classification, whereas low-confidence outputs would not be expected to trigger activation without additional investigations. McNemar's test with continuity correction was used for paired comparisons; the exact McNemar test (binomial method) was substituted when the total number of discordant pairs was <25. Territory-level specificity is reported descriptively with 95% confidence intervals given limited subgroup sizes; no formal cross-territory statistical comparison was performed because territories are not mutually exclusive at the ECG level. Continuous data are presented as mean ± SD or median (interquartile range); categorical data as n (%). A two-sided P < 0.05 was considered statistically significant. Analyses were conducted with Python 3.13.0 (NumPy, pandas, statsmodels, scikit-learn, SciPy) in Visual Studio Code (version 1.95) connected to a local PostgreSQL database (version 14.15). Results Between January 2010 and June 2025, 242 patients with confirmed IMPS were included in the study (Figure 1, Central Illustration). Specificity of Queen of Hearts AI-ECG in Inflammatory Myopericardial Syndromes This retrospective single-center study evaluated the ECG-level specificity of the PMcardio Queen of Hearts (QoH) artificial intelligence (AI) ECG model for correctly classifying inflammatory myopericardial syndromes (IMPS) ECGs as nonocclusion myocardial infarction (OMI) in patients with acute chest pain. A total of 242 IMPS patients were included. When applied as an ECG-only tool, QoH demonstrated high overall specificity (95.0%), with fewer false-positive classifications than ESC STEMI ECG criteria, the 4-variable formula, and general-purpose large language models. Specificity was higher in (myo)pericarditis than in (peri)myocarditis and showed modest variation across ECG territories. Prospective studies are warranted to define sensitivity and to evaluate the potential impact of QoH on clinical decision-making and outcomes in patients with suspected IMPS. LLM = large language model. Cohort characteristics (Myo)pericarditis was present in 176 patients (72.7%), while (peri)myocarditis was diagnosed in 65 patients (26.9%), and 1 patient had neither specific diagnosis documented. A total of 177 (73.1%) patients were male. Electrocardiographic territory involvement was documented in 30 patients (12.4%) for the anterior territory, 29 patients (12.0%) for the lateral territory, and 25 patients (10.3%) for the inferior territory. The 4-variable score was available for 48 patients (19.8%), of whom 13 (27.1%) tested positive. Baseline and clinical characteristics are presented in Table 1. Table 1Baseline and Clinical Characteristics Stratified by Condition TypeOverall (N = 242)(Peri)myocarditis (n = 65)(Myo)pericarditis (n = 176)P ValueMale, n (%)177 (73.1)52 (80.0)124 (70.5)0.19QoH false positive for OMI, n (%)12 (5.0)8 (12.3)4 (2.3)0.004QoH low confidence, n (%)3 (1.2)2 (3.1)1 (0.6)0.374-variable score positive, n/N (%)13/48 (27.1)3/15 (20.0)10/33 (30.3)0.69ESC STEMI ECG criteria positive, n/N (%)53/242 (21.9)19/65 (29.2)34/176 (19.3)0.14Anterior territory STE, n (%)30 (12.4)6 (9.2)24 (13.6)0.48Lateral territory STE, n (%)29 (12.0)11 (16.9)18 (10.2)0.23Inferior territory STE, n (%)25 (10.3)9 (13.8)16 (9.1)0.40Values are n (%) unless otherwise specified.P values calculated using chi-square test or Fisher’s exact test for categorical variables. One patient with adjudicated IMPS could not be assigned to a specific phenotypic subgroup due to insufficient documentation and was therefore excluded from phenotype-stratified subgroup analyses.ESC = European Society of Cardiology; QoH = Queen of Hearts; STE = ST-segment elevation; STEMI = ST-segment elevation myocardial infarction. Baseline and Clinical Characteristics Stratified by Condition Type Values are n (%) unless otherwise specified. P values calculated using chi-square test or Fisher’s exact test for categorical variables. One patient with adjudicated IMPS could not be assigned to a specific phenotypic subgroup due to insufficient documentation and was therefore excluded from phenotype-stratified subgroup analyses. ESC = European Society of Cardiology; QoH = Queen of Hearts; STE = ST-segment elevation; STEMI = ST-segment elevation myocardial infarction. QoH specificity The QoH demonstrated a specificity of 95.0% (95% CI: 91.5%-97.1%) in correctly identifying IMPS patients as not having OMI, with a false-positive rate of 5.0% (95% CI: 2.9%-8.5%). Among the 242 patients, 230 were correctly classified as not having OMI, while 12 patients were incorrectly flagged as potential OMI cases (Table 2). Three additional patients were classified with low confidence by the QoH (1.2%). When these low-confidence results were reclassified as false positives, the specificity decreased to 93.8% (95% CI: 90.0%-96.2%) with a false-positive rate of 6.2% (95% CI: 3.8%-10.0%), representing 3 additional false-positive cases. In the complementary sensitivity analysis restricted to patients with invasive coronary angiography and/or cardiac MRI, QoH demonstrated a specificity of 87.5% (95% CI: 77.2-93.5) in the primary analysis and of 84.4% (73.6-91.3) when the low-confidence results were counted as positive. In the complementary sensitivity analysis restricted to patients with invasive coronary angiography only, QoH demonstrated a specificity of 77.3% (95% CI: 56.6-89.9) in the primary analysis (Supplemental Tables 2 and 3). Table 2Queen of Hearts Specificity for Non-OMI Classification in an IMPS Cohortn%95% CICorrectly identified as NOT OMI (true negative)23095.091.5–97.1Incorrectly classified as OMI (false positive)125.02.9–8.5Low confidence classificationa31.2—Values are number and percentage with 95% confidence intervals calculated using Wilson score method. STEMI denotes ST-elevation myocardial infarction; IMPS = inflammatory myopericardial syndromes.aIn the primary analysis, low-confidence outputs were classified as negative (non-OMI) and therefore included within the true-negative category. Queen of Hearts Specificity for Non-OMI Classification in an IMPS Cohort Values are number and percentage with 95% confidence intervals calculated using Wilson score method. STEMI denotes ST-elevation myocardial infarction; IMPS = inflammatory myopericardial syndromes. In the primary analysis, low-confidence outputs were classified as negative (non-OMI) and therefore included within the true-negative category. QoH specificity comparison with alternative diagnostic criteria The QoH demonstrated significantly superior specificity compared to the ESC STEMI ECG criteria. The ESC criteria showed a specificity of 78.1% (95% CI: 72.5%-82.9%) with a false-positive rate of 21.9% (95% CI: 17.1%-27.5%), incorrectly classifying 53 of 242 IMPS patients as OMI (Table 3). McNemar’s test for paired comparisons revealed a highly significant difference between the 2 diagnostic approaches (P < 0.001). Except for the prespecified primary comparison between QoH and ESC STEMI ECG criteria in the full IMPS cohort, all additional comparisons were exploratory, and the associated P-values should be interpreted as descriptive and hypothesis-generating. The 4-variable score was available for 48 patients and demonstrated a specificity of 72.9% (95% CI: 58.7%-83.6%) with a false-positive rate of 27.1% (95% CI: 16.4%-41.3%), incorrectly classifying 13 patients as STEMI. In the anterior territory subgroup, where both the QoH and 4-variable score were available for 30 patients, the QoH demonstrated a specificity of 96.7% (95% CI: 83.3%-99.4%) compared to 73.3% (95% CI: 55.6%-85.8%) for the 4-variable score (Table 3). The QoH incorrectly classified one patient (3.3%), while the 4-variable score incorrectly classified 8 patients (26.7%). McNemar’s test confirmed a statistically significant difference favoring the QoH score (P = 0.02). Figure 2 shows the comparative specificity and false-positive rates for the QoH (with and without low-confidence reclassification), ESC STEMI ECG criteria, and the 4-variable formula. Table 3QoH, LLMs, ESC STEMI ECG Criteria, and 4-Variable Score Specificity Comparison for Non-OMI Classification in an IMPS CohortDiagnostic TestnTrue NegativesFalse PositivesSpecificity (95% CI)False Positive Rate (95% CI)P ValueQoH2422301295.0% (91.5-97.1)5.0% (2.9-8.5)ReferenceQoH including low confidence2422271593.8% (90.0-96.2)6.2% (3.8-10.0)—Claude2421707270.2% (64.2-75.7)29.8% (24.3-35.8)<0.001ChatGPT2421519162.4% (56.1-68.3)37.6% (31.7-43.9)<0.001Gemini2426517726.9% (21.7-32.8)73.1% (67.2-78.3)<0.001ESC STEMI ECG criteria2421895378.1% (72.5-82.9)21.9% (17.1-27.5)<0.0014-variable score (all patients)48351372.9% (58.7-83.6)27.1% (16.4-41.3)—4-variable score (anterior territory)3022873.3% (55.6-85.8)26.7% (14.2-44.4)0.02QoH (anterior territory)3029196.7% (83.3-99.4)3.3% (0.6-16.7)ReferenceP values calculated using McNemar’s test for paired comparisons vs Queen of Hearts score. 95% confidence intervals calculated using Wilson score method.ESC = European Society of Cardiology; IMPS = inflammatory myopericardial syndromes; QoH = Queen of Hearts; STEMI = ST-elevation myocardial infarction. QoH, LLMs, ESC STEMI ECG Criteria, and 4-Variable Score Specificity Comparison for Non-OMI Classification in an IMPS Cohort P values calculated using McNemar’s test for paired comparisons vs Queen of Hearts score. 95% confidence intervals calculated using Wilson score method. ESC = European Society of Cardiology; IMPS = inflammatory myopericardial syndromes; QoH = Queen of Hearts; STEMI = ST-elevation myocardial infarction. Specificity of Queen of Hearts versus Alternative Criteria Comparative specificity of Queen of Hearts vs alternative criteria for non-OMI classification in the IMPS cohort. The left panel displays specificity, representing the proportion of IMPS patients correctly identified as not having OMI. The right panel shows false-positive rates, representing the proportion of IMPS patients incorrectly classified as OMI. Dashed lines indicate clinically relevant thresholds (green: 95% specificity/5% false-positive rate; orange: 90% specificity/10% false-positive rate). 4-Var = 4-variable score; ESC = European Society of Cardiology; IMPS = inflammatory myopericardial syndromes; OMI = occlusion myocardial infarction (encompassing ST-elevation myocardial infarction and ST-elevation myocardial infarction-like pattern); QoH denotes Queen of Hearts. QoH comparison with LLMs (ChatGPT, Claude, and Gemini) We additionally compared QoH performance with 3 general-purpose LLMs: Claude (Anthropic), ChatGPT (OpenAI), and Gemini (Google). All 3 LLMs exhibited substantially lower specificity than QoH for classifying ECGs as non-OMI in this IMPS cohort (Figure 3). Among LLMs, Claude demonstrated the highest specificity at 70.2% (95% CI: 64.1%-75.7%), followed by ChatGPT at 62.4% (95% CI: 56.0%-68.4%), and Gemini at 26.9% (95% CI: 21.6%-32.9%). Compared with QoH, all LLMs yielded significantly higher false-positive rates for OMI classification (29.8% to 73.1% vs 5.0%; all P < 0.001). The absolute specificity difference between QoH and the best-performing LLM (Claude) was 24.8%, corresponding to 60 additional false-positive OMI classifications per 242 IMPS patients. Specificity of Queen of Hearts versus Large Language Model The left panel displays specificity, representing the proportion of IMPS patients correctly identified as not having OMI. The right panel shows false-positive rates, representing the proportion of IMPS patients incorrectly classified as OMI. Dashed lines indicate clinically relevant thresholds (green: 95% specificity/5% false-positive rate; orange: 90% specificity/10% false-positive rate). IMPS = inflammatory myopericardial syndromes; OMI = occlusion myocardial infarction (encompassing ST-elevation myocardial infarction and ST-elevation myocardial infarction-like pattern); QoH = Queen of Hearts. Qualitative analysis of discordant ESC-positive/QoH-negative cases Among patients classified as STEMI-positive by ESC criteria but correctly classified as non-OMI by QoH (n = 43), recurrent inflammatory-type ECG patterns were observed (Table 4, Figure 4). Notched J-points were present in 32 of 43 cases (74.4%), and ST-segment morphology was uniformly concave in all leads meeting STE criteria in 32 of 43 cases (74.4%). STE involved ≥2 territories in 29 of 43 patients (67.4%). Spodick’s sign was identified in 16 of 43 cases (37.2%), and PR-segment depression in 13 of 43 cases (30.2%). An ST/T ratio >0.25 was observed in 22 of 43 patients (51.2%). Reciprocal ST-segment depression (outside leads aVR and V1) was only present in 3 of 43 cases (7.0%) and was limited to the aVL lead. Table 4Qualitative Analysis of Discordant ESC-Positive/QoH-Negative Cases (N = 43)Notched J point (≥1 derivation)32 (74.4%)Concave ST-segment morphology in all STE leads32 (74.4%)Spodick sign16 (37.2%)STE in ≥2 territories29 (67.4%)PR-segment depression13 (30.2%)Reciprocal changes (except in aVR and V1):aVL only3 (7.0%)3 (7.0%)ST/T ratio >0.2522 (51.2%)Values are n (%) unless otherwise specified.ESC = European Society of Cardiology; QoH = Queen of Hearts; STE = ST-segment elevation; STEMI = ST-elevation myocardial infarction. Qualitative Analysis of Discordant ESC-Positive/QoH-Negative Cases (N = 43) Values are n (%) unless otherwise specified. ESC = European Society of Cardiology; QoH = Queen of Hearts; STE = ST-segment elevation; STEMI = ST-elevation myocardial infarction. Representative Discordant ESC-Positive/QoH-Negative ECGs Representative ECGs from discordant cases meeting ESC STEMI ECG criteria but classified as non-OMI by QoH. Despite extensive STE involving anterior, inferior, and/or lateral leads (A to D), straight or partially convex ST morphology (A and C), and bifid ST morphology in V3-V4 (C), QoH likely favored benign inflammatory features, including predominantly concave ST morphology (A to D), terminal-QRS notching (A, B, and D), preserved R-wave progression (A to D), absence of clear territorial distribution (A, B, and D), subtle PR-segment depression and positive Spodick sign (B), and lack of reciprocal ST-segment depression (A to D). ESC = European Society of Cardiology; QoH = Queen of Hearts; STEMI = ST-segment elevation myocardial infarction; OMI = occlusion myocardial infarction; STE = ST-segment elevation. Qualitative analysis of discordant QoH-positive/ESC-negative cases QoH yielded false-positive OMI classifications while ESC STEMI criteria remained negative in 3 ECG cases (Figure 5). None of these ECGs met ESC-defined ST-elevation thresholds. On qualitative analysis with heatmap visualization, these discordant cases were characterized by subtle repolarization abnormalities. These included localized inframillimetric anterior STE with concave morphology, inferior territory submillimetric ST elevation associated with J-point notching, and lateral J-point notching with accompanying inferior ST-segment depression and T-wave inversion. Across cases, additional features included relative T-wave prominence and low QRS-to-T amplitude contrast. Representative Discordant QoH-Positive/ESC-Negative ECGs Representative ECGs from discordant cases classified as OMI by QoH but not meeting ESC STEMI ECG criteria. Despite the absence of ESC STEMI ECG criteria (A to C), heatmap activation highlighted subtle or borderline repolarization abnormalities, including localized inframillimetric anterior STE with concave morphology (A), inferior territory submillimetric STE with J-point notching (B), and lateral J-point notching with associated inferior ST-segment depression and T-wave inversion (C). Additional features included relative T-wave prominence (A to C) and low QRS-to-T amplitude contrast (A and B). ESC = European Society of Cardiology; QoH = Queen of Hearts. STE territory-based subgroup analysis Diagnostic performance of the QoH varied across electrocardiographic territories involvement. In the anterior territory (n = 30), the specificity was 96.7% (95% CI: 83.3%-99.4%) with a false-positive rate of 3.3%. In the lateral territory (n = 29), the specificity was 86.2% (95% CI: 69.4%-94.5%) with a false-positive rate of 13.8%. In the inferior territory (n = 25), the specificity was 80.0% (95% CI: 60.9%-91.1%) with a false-positive rate of 20.0% (Table 5). The inferior territory showed the highest observed false-positive rate of OMI classifications, with 5 of 25 patients incorrectly classified, compared to 1 of 30 in the anterior territory and 4 of 29 in the lateral territory. The STE criteria negative group demonstrated a specificity of 97.9% (95% CI: 94.7%-99.2%) with a false-positive rate of 2.1% (Table 5). These territory-based subgroup analyses were exploratory and should be interpreted descriptively given limited subgroup sample sizes and wide confidence intervals. Table 5ST-Segment Elevation Territory-Based Subgroup Analysis of Queen of Hearts Specificity (n = 84)ECG TerritorynTrue NegativesFalse PositivesSpecificity (95% CI)False Positive Rate (95% CI)ECG TerritoryAnterior STE3029196.7% (83.3-99.4)3.3% (0.6-16.7)AnteriorLateral STE2925486.2% (69.4-94.5)13.8% (5.5-30.6)LateralInferior STE2520580.0% (60.9-91.1)20.0% (8.9-39.1)InferiorSTE criteria negative189185497.9% (94.7-99.2)2.1% (0.8-5.3)95% CIs calculated using Wilson score method.ECG = electrocardiographic. ST-Segment Elevation Territory-Based Subgroup Analysis of Queen of Hearts Specificity (n = 84) 95% CIs calculated using Wilson score method. ECG = electrocardiographic. Condition type analysis Differences in QoH performance were observed between (myo)pericarditis and (peri)myocarditis patients (P = 0.004), although this phenotype-based comparison should be interpreted as exploratory. (Myo)pericarditis patients (n = 176) demonstrated higher specificity at 97.7% (95% CI: 94.4%-99.1%) with a false-positive rate of 2.3% (95% CI: 0.9%-5.6%). By contrast, (speri)myocarditis patients (n = 65) showed lower specificity at 87.7% (95% CI: 78.0%-93.5%) with a false-positive rate of 12.3% (95% CI: 6.5%-22.0%) (Table 6). The absolute difference in false-positive rates between the 2 conditions was 10.0 percentage points. Gender distribution was similar ((myo)pericarditis 73.3% male vs (peri)myocarditis 72.3% male, P = 0.88) between the 2 condition types. (Peri)myocarditis patients demonstrated higher rates of positive ESC STEMI ECG criteria (29.2% vs 19.3%, P = 0.09) and numerically higher rates of anterior territory involvement (16.9% vs 10.8%, P = 0.19), though these differences did not reach statistical significance (Table 1). One patient with adjudicated IMPS could not be assigned to a specific phenotypic subgroup because of insufficient documentation and was therefore excluded from phenotype-stratified subgroup analyses. Table 6Queen of Hearts Specificity Stratified by Condition TypeCondition TypenTrue NegativesFalse PositivesSpecificity (95% CI)False Positive Rate (95% CI)P Value(Myo)pericarditis176172497.7% (94.3-99.1)2.3% (0.9-5.7)Reference(Peri)myocarditis6557887.7% (77.5-93.6)12.3% (6.4-22.5)0.00495% confidence intervals calculated using Wilson score method. One patient with adjudicated IMPS could not be assigned to a specific phenotypic subgroup due to insufficient documentation and was therefore excluded from phenotype-stratified subgroup analyses. Queen of Hearts Specificity Stratified by Condition Type 95% confidence intervals calculated using Wilson score method. One patient with adjudicated IMPS could not be assigned to a specific phenotypic subgroup due to insufficient documentation and was therefore excluded from phenotype-stratified subgroup analyses. Discussion Main findings and clinical context Patients presenting with acute chest pain can be difficult to manage, and distinguishing OMI from IMPS may be challenging. Because the ECG is the initial test performed for chest pain in the emergency setting, its accurate interpretation is essential to guide ongoing management. The OMI paradigm urges clinicians to look beyond ST-segment elevation thresholds and recognize STEMI equivalents and subtler occlusion patterns. This requires a higher-level expertise to avoid missing OMI and delay appropriate care.2, 3, 4, 5, 6, 7 In that light, the QoH, trained on angiographic OMI, appears particularly useful when IMPS is suspected. In our IMPS-only cohort, QoH demonstrated high ECG-level specificity for OMI classification and behaved conservatively in the presence of IMPS (Central Illustration). These observations are consistent with prior work suggesting that QoH may reduce false-positive OMI classifications compared with conventional ECG criteria.9, 10, 11, 12,17 Absolute specificity values are lower in subgroups of patients with invasive coronary angiography and/or cardiac MRI, which is anticipated given the enrichment for higher-risk troponin-positive presentations and ECG abnormalities that prompted invasive workup. Critically, however, the relative ranking and the direction of effect are preserved, supporting the robustness of the primary finding against residual diagnostic uncertainty in the pericarditis-only adjudication pathway. QoH vs conventional criteria Compared with ESC STEMI ECG criteria and the 4-variable score, QoH demonstrated higher specificity and fewer false-positive classifications, when applied as an ECG only tool. This likely reflects its training on angiographic OMI endpoints rather than reliance on fixed ST-elevation thresholds and its greater robustness to inflammatory “mimic” patterns. By contrast, the 4-variable formula was developed to distinguish subtle anterior left anterior descending coronary artery occlusion from early repolarization and was not designed for broader inflammatory populations.14,16 These findings align with the broader OMI literature showing that threshold-based STE rules and ECG scores may be insufficient when applied outside their validated contexts.2, 3, 4, 5, 6, 7 QoH vs general-purpose LLMs The markedly lower specificity observed with general-purpose LLMs reflects fundamental differences in architecture and intended use. QoH is a task-specific AI optimized for quantitative ECG waveform analysis, whereas LLMs rely on probabilistic language-based pattern recognition and are not designed for high-stakes biomedical signal interpretation.10,11 In IMPS—where inflammatory repolarization frequently mimics ischemia—LLMs showed a pronounced tendency toward OMI overclassification. General-purpose LLMs lack calibrated diagnostic thresholds, structured waveform modeling, and uncertainty control mechanisms required for reliable ECG-based decision support. Their outputs are susceptible to image quality variation, formatting differences, lead labeling inconsistencies, and prompt sensitivity, and they do not inherently integrate structured clinical context known to influence ECG interpretation.18, 19, 20, 21, 22 Contextualization has been shown to modify LLM diagnostic performance and reduce hallucination and out-of-distribution errors.18,20,22,23 Although emerging vision-language ECG models may partially address these limitations, they remain distinct from general-purpose systems.19, 20, 21, 22, 23 Our ECG image-based evaluation with minimal clinical context therefore represents an upper-bound estimate of false-positive risk when LLMs are used without clinical integration. These findings underscore the need for rigorous validation and governance before deployment in acute cardiovascular care. Because each ECG was evaluated only once per model using publicly available interfaces, intramodel variability and uncertainty calibration could not be assessed; accordingly, comparative specificity estimates for LLMs should be interpreted with appropriate caution. Conversely, analysis of QoH-positive/ESC-negative cases highlights potential failure modes of the model. These ECGs showed subtle repolarization abnormalities, including localized inframillimetric STE with J-point notching. Heatmap analysis suggests that QoH may also overemphasize selected features such as relative T-wave prominence and low QRS-to-T amplitude differences. These findings delineate conditions under which false-positive classifications may occur and support cautious interpretation within a broader clinical framework. Mechanistic insights from discordant cases Qualitative analysis of ESC-positive/QoH-negative cases suggests that QoH integrates distributed morphological and spatial ECG features beyond millimeter-based ST-segment thresholds. Discordant tracings frequently exhibited diffuse concave STE, terminal-QRS notching, multiterritorial involvement, PR-segment depression, and elevated ST/T ratios—hallmarks of inflammatory syndromes—whereas strict territorial confinement and reciprocal changes were uncommon.24, 25, 26 This supports the hypothesis that QoH captures global repolarization coherence and waveform patterning not encompassed by rigid amplitude-based criteria. In contrast, QoH-positive/ESC-negative cases highlight potential limitations of the model. These ECGs were characterized by subtle repolarization abnormalities with heatmap analysis suggesting that QoH may emphasize selected repolarization features while not incorporating other characteristics, delineating conditions under which false-positive classifications may occur. Explainability and clinical implementation As with other deep learning–based AI-ECG systems, QoH retains elements of a “black box” model. Although visual attention maps provide partial insight into regions contributing to positive classifications, they do not offer causal interpretability and are unavailable for negative outputs.10,11 Limited transparency may affect clinician trust, particularly when AI outputs diverge from guideline thresholds. Prior studies have shown that contextual information substantially influences AI-based ECG interpretation and diagnostic confidence.18 Accordingly, QoH could be considered as a decision-support tool within guideline-consistent, multimodal workflows under human oversight rather than as an autonomous decision-maker. Continued advances in interpretability, calibration, and prospective validation will be essential to mitigate black box limitations in high-stakes triage environments. Subgroup considerations QoH performance showed a numerical trend toward STE territory dependence, with inferior patterns appearing more prone to misclassification than anterior patterns. This observation is consistent with contemporary data demonstrating that IMPS may produce regional STE within heterogeneous inflammatory ECG backgrounds24,25,27,28 and supports territory-aware interpretation of AI outputs. The observed performance gap between (myo)pericarditis and (peri)myocarditis warrants specific consideration. Higher specificity in pericarditis-predominant presentations likely reflects more homogeneous and characteristic inflammatory ECG patterns, whereas myocardial involvement introduces greater repolarization complexity, spatial heterogeneity, and temporal variability, increasing the likelihood of misclassification.1,25,26,29 Nonetheless, these subgroup analyses were not powered for formal hypothesis testing and should therefore be interpreted primarily in terms of effect-size patterns and confidence-interval widths. This difference underscores an important limitation of ECG-only approaches in inflammatory syndromes with myocardial injury. Accordingly, QoH should be viewed as a task-specific, ECG-level classification tool that may serve as an initial “entry point” for OMI exclusion, but not as a standalone decision-making system. Its outputs require integration within a broader multimodal clinical framework, including biomarkers and imaging, to ensure appropriate interpretation and diagnostic relevance.1,9, 10, 11, 12 Limitations This single-center observational study has several limitations. First, its retrospective, single-center design may introduce selection and information bias. Second, analyses relied on a single index ECG per patient and did not capture dynamic electrocardiographic changes. Third, because only IMPS patients were included, OMI sensitivity could not be assessed, limiting generalizability and extrapolation to triage effectiveness in unselected chest pain populations. Fourth, qualitative analysis of discordant cases was exploratory and subject to observer interpretation. Fifth, LLM performance was evaluated using publicly available chatbot interfaces under real-world usage conditions. Model build identifiers and inference parameters (eg, temperature, top-p) were not user-configurable, and each ECG was evaluated once per model without an assessment of interrun variability. Accordingly, response variability and uncertainty calibration could not be quantified. Furthermore, LLM performance was evaluated using standardized image-based inputs and fixed prompts, which may not reflect real-world clinical integration. Sixth, ECG-level performance could not be directly linked to downstream clinical decisions, catheterization laboratory activation, or patient outcomes. Seventh, because multiple secondary and exploratory comparisons were performed without adjustment for multiplicity, associated P-values should be interpreted as descriptive and hypothesis-generating. Finally, QoH performance may evolve with future model updates. Nevertheless, the magnitude of the specificity advantage, consistent subgroup findings, and concordant mechanistic analyses support the robustness and clinical relevance of our conclusions. Conclusions Among patients with IMPS, the QoH demonstrated high ECG-level specificity for non-OMI classification, with fewer false-positive classifications than ESC STEMI ECG criteria, the 4-variable score, and LLMs. Prospective studies are warranted to define sensitivity and to evaluate the potential impact of QoH on clinical decision-making and outcomes in patients with suspected IMPS. COMPETENCY IN PATIENT CARE AND MEDICAL KNOWLEDGE: In patients presenting with acute chest pain and inflammatory myopericardial syndromes, electrocardiographic patterns frequently mimic occlusion myocardial infarction. An OMI-trained AI-ECG model (Queen of Hearts) demonstrated high specificity in this setting. COMPETENCY IN SYSTEMS-BASED PRACTICE AND PRACTICE-BASED LEARNING: By demonstrating higher ECG-level specificity for OMI classification in an IMPS-only cohort compared with guideline STEMI criteria and general-purpose language models, task-specific AI-ECG may represent a complementary ECG-level classification approach. TRANSLATIONAL OUTLOOK: Translation of OMI-focused AI-ECG tools into routine care requires prospective validation in unselected chest pain populations and assessment of integration with clinical evaluation, biomarkers, and cardiac imaging. Key challenges include defining safe implementation thresholds, ensuring interpretability, and evaluating real-world impact on clinical decision-making and outcomes. Data availability statement The raw data supporting the conclusions of this article will be made available by the authors on reasonable request. Funding support and author disclosures This study was funded by the Fond Scientifique Cardiovasculaire of the University and Hospital Fribourg. The 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 an expanded Methods section and supplemental tables, please see the online version of this paper. Supplementary material Supplementary appendix References Associated Data Supplementary Materials Supplementary appendix Data Availability Statement The raw data supporting the conclusions of this article will be made available by the authors on reasonable request.