JACC Adv. 2026;5(4):102660.

Artificial Intelligence-Driven Electrocardiogram Screening for Asymptomatic Left Ventricular Systolic Dysfunction in the General Population.

, , , , , , , , , , ,

제1저자 이태민, 강소라 · 교신저자 이희선, 이학승

PMID 41849876 PMC13015574 DOI 10.1016/j.jacadv.2026.102660 IF 5.5 순환기내과

소속 (5)

Department of Internal Medicine, Seoul National University College of Medicine, Seoul, Republic of Korea; Division of Cardiology, Seoul National University Hospital Healthcare System Gangnam Centre, Seoul, Republic of Korea.
Digital Healthcare Institute, Sejong Medical Research Institute, Bucheon, Republic of Korea; Medical AI Co, Ltd, Seoul, Republic of Korea.
Medical AI Co, Ltd, Seoul, Republic of Korea.
Digital Healthcare Institute, Sejong Medical Research Institute, Bucheon, Republic of Korea; Medical AI Co, Ltd, Seoul, Republic of Korea. Electronic address: cardiolee@gmail.com.
Department of Internal Medicine, Seoul National University College of Medicine, Seoul, Republic of Korea; Division of Cardiology, Seoul National University Hospital Healthcare System Gangnam Centre, Seoul, Republic of Korea. Electronic address: 65847@snuh.org.

Abstract

BACKGROUND Asymptomatic left ventricular systolic dysfunction (LVSD) is a well-established precursor of overt heart failure (HF), yet it often remains undiagnosed in the general population. Artificial intelligence-enabled electrocardiogram (ECG) analysis offers a scalable approach for early detection.

OBJECTIVES The purpose of this study was to evaluate the diagnostic performance of an artificial intelligence-enabled ECG model (AiTiALVSD) for identifying asymptomatic LVSD in a large health screening population.

METHODS In this retrospective, single-center study, we evaluated the AiTiALVSD model among 40,713 self-referred adults who underwent a total of 60,711 ECG-transthoracic echocardiography (TTE) pairs between 2011 and 2023. LVSD was defined as a left ventricular ejection fraction ≤40%. Model discrimination was assessed using the area under the receiver-operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC), and diagnostic performance metrics were compared with established HF risk scores.

RESULTS Among 60,711 ECG-TTE pairs, 32 cases (0.054%) met the criteria for LVSD. The AiTiALVSD model demonstrated excellent discrimination (AUROC 0.973; AUPRC 0.328), with a sensitivity of 90.6%, specificity of 99.4%, positive predictive value of 7.7%, and a negative predictive value of 100%. Established HF risk scores, including the MESA (Multi-Ethnic Study of Atherosclerosis) 5-year HF score and Pooled Cohort Equations to Prevent HF score, showed inferior discrimination (AUROC: 0.696 and 0.672, respectively). The MESA score was not designed to detect prevalent LVSD and was calculated without natriuretic peptide data, which may have disadvantaged its performance in this comparison. Simulation analyses suggested that approximately 1,841 ECGs and 13 confirmatory TTEs would be required to detect one case of LVSD.

CONCLUSIONS In a real-world screening population with an extremely low prevalence of LVSD, the AiTiALVSD model demonstrated high diagnostic accuracy, supporting its potential role as a rule-out screening tool for HF prevention. Prospective validation is warranted.

Keywords

artificial intelligenceelectrocardiogramheart failureleft ventricular systolic dysfunctionprediction model

Figures

**Figure 1** **Study Flow** Between January 2011 and July 2023, a total of TTE and 12-lead ECG pairs were collected from the Seoul National University Hospital Healthcare System. Inclusion criteria were individuals aged 18 years or older who underwent a 12-lead ECG within 14 days of TTE. The final cohort consisted of 60,711 TTE-ECG pairs from 40,713 unique individuals. AI = artificial intelligence; ECG = electrocardiogram; LVSD = left ventricular systolic dysfunction; TTE = transthoracic echocardiography.

**Figure 2** **Architecture of AiTiALVSD Model** ECG = electrocardiogram; LVEF = left ventricular ejection fraction; ReLU = rectified linear unit.

**Figure 3** **AiTiALVSD Model Performance in Detecting LVSD** A and C show the ROC curves, indicating the discrimination performance of the AiTiALVSD model in identifying left ventricular systolic dysfunction with 2 thresholds: LVEF ≤40% (AUROC = 0.973) and LVEF ≤50% (AUROC = 0.937). B and D display the corresponding PR curves, with AUPRC values of 0.328 for LVEF ≤40% and 0.375 for LVEF ≤50%. The shaded areas around the curves represent the 95% CIs. AUROC = area under the receiver-operating characteristic curve; AUPRC = area under the precision–recall curve; other abbreviations as in Figures 1 and 2.

**Figure 4** **Comparison of AUROC Curves for AiTiALVSD Model and Conventional HF Prediction Models** AUROC curves illustrate the discrimination performance of the AiTiALVSD model (orange line, AUROC = 0.917) compared to models based on the MESA 5-year HF score (green line, AUROC = 0.696), and the PCP-HF score (red line, AUROC = 0.672). The shaded areas indicate the 95% CIs for each model. MESA = Multi-Ethnic Study of Atherosclerosis; PCP-HF = Pooled Cohort Equations to Prevent HF; other abbreviation as in Figure 3.

**Central Illustration** **AI-ECG Screening for Asymptomatic Left Ventricular Systolic Dysfunction** This illustration summarizes the study design, diagnostic performance, and screening efficiency of the AiTiALVSD model in a self-referred health checkup population. A total of 60,711 ECG-TTE pairs from 40,713 adults were analyzed to assess the ability of a 12-lead AI-enhanced ECG (AiTiALVSD) to detect asymptomatic LVSD, defined as LVEF ≤40% on echocardiography. The model demonstrated excellent discrimination (AUROC 0.973, AUPRC 0.328), high sensitivity (90.6%), specificity (99.4%) with a very high negative predictive value in this extremely low-prevalence screening setting. In a cohort with a 0.054% prevalence of asymptomatic LVSD, simulation suggests that 1,841 AI-ECG tests and 13 confirmatory echocardiograms are needed to identify one true positive case. The model’s high NPV allows for efficient rule-out in large populations, supporting its use. EF = ejection fraction; NPV = negative predictive value; PPV = positive predictive value; other abbreviations as in Figures 1 and 2.

그림 출처: PubMed Central 공개 논문 (PMC13015574)

Conflict of interest statement

Funding support and author disclosures This study was funded by Medical AI Co, Ltd. Authors affiliated with Medical AI contributed to the study design, data collection, analysis, and interpretation in their capacity as individual researchers. The funder did not impose restrictions on data access, data interpretation, or the decision to submit the manuscript for publication. Drs Kang, Min Sung Lee, Han, Yoo, Jang, Jo, Son, Kwon, and Hak Seung Lee are employees of Medical AI Co, Ltd and hold stocks in the company. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.

그 밖의 정보