Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Mar 4;16:11996. doi: 10.1038/s41598-026-41824-5 Search in PMC Search in PubMed View in NLM Catalog Add to search Validation of AI-enhanced ECG image analysis for identifying extreme cardiac magnetic resonance metrics in a cross-ethnic UK biobank study Yerin Kim Yerin Kim 1 Division of Cardiology, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea Find articles by Yerin Kim 1, # , Haemin Lee Haemin Lee 2 Department of Emergency Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea 3 ARPI Inc., Seongnam, Republic of Korea Find articles by Haemin Lee 2, 3, # , Hong-Mi Choi Hong-Mi Choi 1 Division of Cardiology, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea Find articles by Hong-Mi Choi 1 , In-Chang Hwang In-Chang Hwang 1 Division of Cardiology, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea Find articles by In-Chang Hwang 1 , Joonghee Kim Joonghee Kim 2 Department of Emergency Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea 3 ARPI Inc., Seongnam, Republic of Korea Find articles by Joonghee Kim 2, 3 , Ji Hyun Lee Ji Hyun Lee 1 Division of Cardiology, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea Find articles by Ji Hyun Lee 1 , Il-Young Oh Il-Young Oh 1 Division of Cardiology, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea Find articles by Il-Young Oh 1 , Heesun Lee Heesun Lee 4 Department of Internal Medicine, Seoul National University College of Medicine, Seoul, Republic of Korea 5 Division of Cardiology, Department of Internal Medicine, Seoul National University Hospital Healthcare System Gangnam Center, 152 Teheran-ro, Gangnam-gu, Seoul, Republic of Korea Find articles by Heesun Lee 4, 5 , Su-Yeon Choi Su-Yeon Choi 4 Department of Internal Medicine, Seoul National University College of Medicine, Seoul, Republic of Korea 5 Division of Cardiology, Department of Internal Medicine, Seoul National University Hospital Healthcare System Gangnam Center, 152 Teheran-ro, Gangnam-gu, Seoul, Republic of Korea Find articles by Su-Yeon Choi 4, 5 , Tae-Min Rhee Tae-Min Rhee 4 Department of Internal Medicine, Seoul National University College of Medicine, Seoul, Republic of Korea 5 Division of Cardiology, Department of Internal Medicine, Seoul National University Hospital Healthcare System Gangnam Center, 152 Teheran-ro, Gangnam-gu, Seoul, Republic of Korea Find articles by Tae-Min Rhee 4, 5, ✉ , Youngjin Cho Youngjin Cho 1 Division of Cardiology, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea 3 ARPI Inc., Seongnam, Republic of Korea 6 Cardiovascular Center, Seoul National University Bundang Hospital, 82 Gumi-ro-173-gil, Bundang, Seongnam, Gyeonggi 13620 Republic of Korea Find articles by Youngjin Cho 1, 3, 6, ✉ Author information Article notes Copyright and License information 1 Division of Cardiology, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea 2 Department of Emergency Medicine, Seoul National University Bundang Hospital, Seongnam, Republic of Korea 3 ARPI Inc., Seongnam, Republic of Korea 4 Department of Internal Medicine, Seoul National University College of Medicine, Seoul, Republic of Korea 5 Division of Cardiology, Department of Internal Medicine, Seoul National University Hospital Healthcare System Gangnam Center, 152 Teheran-ro, Gangnam-gu, Seoul, Republic of Korea 6 Cardiovascular Center, Seoul National University Bundang Hospital, 82 Gumi-ro-173-gil, Bundang, Seongnam, Gyeonggi 13620 Republic of Korea ✉ Corresponding author. # Contributed equally. Received 2025 Oct 30; Accepted 2026 Feb 23; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13069017 PMID: 41776266 Abstract Artificial intelligence enhanced electrocardiography (AI-ECG) has shown promise in detecting cardiac abnormalities, but validation against cardiac magnetic resonance (CMR), the reference standard for cardiac structure and function, remains limited. We evaluated the diagnostic performance of six AI-ECG models in identifying CMR-defined cardiac abnormalities in a population-based cohort. This cohort study included 38,804 UK Biobank participants with paired 12-lead ECG and CMR data. AI-ECG models, originally developed using Korean hospital-based datasets, were externally validated in this sample. Cardiac abnormalities were defined as the top 1% of CMR-derived values. Outcomes were the performance of AI-ECG in detecting left and right ventricular dysfunction, based on ejection fraction (QCG-LVD and QCG-RVD) and strain (ECG-LVGLS and ECG-RVGLS), and structural abnormalities including left ventricular hypertrophy (LVH) and left atrial enlargement (LAE). Model performance was assessed using the area under the receiver operating characteristic curve (AUC), with subgroup analyses by age, sex, and comorbidities. Among 38,804 participants (mean age 64.2 years; 48.2% male), AI-ECG models showed high accuracy for detecting left ventricular dysfunction (AUC 0.887 for QCG-LVD and 0.896 for ECG-LVGLS) and right ventricular dysfunction (AUC 0.778 for QCG-RVD and 0.825 for ECG-RVGLS). In a sub-cohort of 21,267 individuals, AUCs were 0.824 for detecting LVH and 0.883 for LAE. Subgroup analyses showed consistent performance, with higher accuracy among older individuals, males, and those with hypertension or ischemic heart disease. In this large multiethnic cohort, AI-ECG models demonstrated strong performance in detecting CMR-defined abnormalities, supporting their potential as a scalable, noninvasive screening tool for cardiovascular disease. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-026-41824-5. Keywords: Artificial intelligence, Electrocardiogram, Cardiac magnetic resonance imaging Subject terms: Cardiology, Diseases, Medical research Introduction Electrocardiography (ECG) is a widely accessible, non-invasive, and cost-effective tool for evaluating cardiovascular health. Recent advances in artificial intelligence (AI) have significantly expanded its clinical utility 1 – 3 . AI-enhanced ECG (AI-ECG) enables the prediction of structural and functional cardiac abnormalities that have typically required advanced imaging modalities such as echocardiography or cardiac magnetic resonance imaging (CMR). By leveraging the ubiquity of ECG, AI-ECG offers a scalable strategy for early detection and screening of cardiovascular disease, potentially reducing dependence on resource-intensive imaging tools. Our group has previously developed and validated deep learning based AI-ECG models that analyze ECG images to generate digital biomarkers for various cardiac conditions 4 – 7 . These models rely on image-based ECG input rather than raw signal data, offering flexibility in clinical implementation and demonstrating high diagnostic performance. However, despite their promising performance, further validation is necessary to assess their generalizability in broader populations, particularly across different ethnicities and outside of hospital-based cohorts. This study aims to validate previously developed AI-ECG models in a general population from the UK Biobank, comprising over 38,000 participants with paired ECG and CMR data. Results Baseline characteristics Baseline characteristics of the study population are summarized in Table 1 . The mean age of participants was 64.2 ± 7.8 years, with 48.2% being male. Comorbidities included hypertension in 32.9%, diabetes mellitus in 5.9%, and dyslipidemia in 24.1%. The mean left ventricular ejection fraction (LVEF) measured by CMR was 59.5 ± 6.1%, and the mean right ventricular ejection fraction (RVEF) was 57.3 ± 6.2%. Table 1. Baseline characteristics of the overall study population. Total ( n = 38,804) Age, years 64.2 ± 7.8 Sex, male (%) 18,709 (48.2%) Ethnicity (%) White 37,480 (96.6%) Mixed 193 (0.5%) Asian 540 (1.4%) Black 272 (0.7%) Others 211 (0.5%) Unknown 108 (0.3%) BMI, kg/m 2 26.5 ± 4.4 Hypertension (%) 12,778 (32.9%) Diabetes mellitus (%) 2,275 (5.9%) Dyslipidemia (%) 9,365 (24.1%) Heart Failure (%) 249 (0.6%) Ischemic heart disease (%) 2,401 (6.2%) Cerebrovascular disease (%) 892 (2.3%) CMR parameters LVEF (%) 59.5 ± 6.1 RVEF (%) 57.3 ± 6.2 LV mass (g) 85.9 ± 22.2 LVMI (g/m 2 )* 45.9 ± 8.5 LA volume (mm 3 ) 72.7 ± 23.6 LAVI (mm 3 /m 2 )* 39.0 ± 11.1 AI-ECG scores QCG-LVD 0.03 ± 0.07 ECG-LVGLS 18.01 ± 1.14 QCG-RVD 0.04 ± 0.06 ECG-RVGLS 23.19 ± 1.59 AI-ECG-LVH 0.22 ± 0.15 AI-ECG-LAE 0.20 ± 0.15 Open in a new tab * The subgroup analysis includes data from a separate population ( n = 21,267). This subgroup was specifically adjusted for body surface area (BSA) to account for differences in LVMI and LAVI. LVEF left ventricular ejection fraction, RVEF right ventricular ejection fraction, LV mass left ventricle mass, LVMI left ventricle mass index, LA volume left atrium volume, LAVI left atrium volume index, QCG quantitative electrocardiogram, LVD left ventricle dysfunction, LVGLS left ventricular global longitudinal strain, RVD right ventricle dysfunction, RVGLS right ventricular global longitudinal strain, LVH left ventricular hypertrophy, LAE left atrial enlargement. AI-ECG performance for detecting left and right ventricular dysfunction Table 2 presents baseline characteristics stratified by the presence of left ventricular dysfunction (LVD) and right ventricular dysfunction (RVD). Participants with LVD or RVD were generally older and had a higher prevalence of comorbid conditions compared to those without dysfunction. In particular, hypertension was more prevalent in both LVD and RVD groups. Individuals with LVD had higher rates of diabetes mellitus and ischemic heart disease, while those with RVD more frequently had heart failure. Table 2. Baseline characteristics according to the presence of left and right ventricular dysfunction. LVD ( n = 389) No LVD ( n = 38,415) p RVD ( n = 389) No RVD ( n = 38,415) p Age, years 69.2 ± 7.0 64.2 ± 7.8 < 0.001 68.5 ± 7.5 64.2 ± 7.8 < 0.001 Male, sex 308 (79.2%) 18,401 (47.9%) < 0.001 18,414 (47.9%) < 0.001 Ethnicity 0.501 0.904 White 379 (97.4%) 37,101 (96.6%) 377 (96.9%) 37,103 (96.6%) Mixed 1 (0.3%) 192 (0.5%) 2 (0.5%) 191 (0.5%) Asian 5 (1.3%) 535 (1.4%) 4 (1.0%) 536 (1.4%) Black 4 (1.0%) 268 (0.7%) 4 (1.0%) 268 (0.7%) Others 0 (0.0%) 211 (0.5%) 1 (0.3%) 210 (0.5%) Unknown 0 (0.0%) 108 (0.3%) 1 (0.3%) 107 (0.3%) BMI, kg/m 2 27.7 ± 4.8 26.5 ± 4.3 < 0.001 27.5 ± 5.0 26.5 ± 4.3 < 0.001 Hypertension 213 (54.8%) 12,565 (32.7%) < 0.001 194 (49.9%) 12,584 (32.8%) < 0.001 Diabetes mellitus 50 (12.9%) 2,225 (5.8%) < 0.001 44 (11.3%) 2231 (5.8%) < 0.001 Dyslipidemia 149 (38.3%) 9,216 24.0%) < 0.001 139 (35.7%) 9,226 (24.0%) < 0.001 Heart Failure 41 (10.5%) 208 (0.5%) < 0.001 27 (6.9%) 222 (0.6%) < 0.001 Ischemic heart disease 96 (24.7%) 2305 (6.0%) < 0.001 62 (15.9%) 2,339 (6.1%) < 0.001 Cerebrovascular disease 26 (6.7%) 866 (2.3%) < 0.001 23 (5.9%) 869 (2.3%) < 0.001 CMR parameters LVEF (%) 36.9 ± 5.1 59.8 ± 5.7 < 0.001 44.4 ± 9.5 59.7 ± 5.9 < 0.001 RVEF (%) 42.6 ± 10.2 57.4 ± 5.9 < 0.001 35.2 ± 5.7 57.5 ± 5.8 < 0.001 AI-ECG scores QCG-LVD 0.28 ± 0.26 0.03 ± 0.06 < 0.001 0.16 ± 0.22 0.03 ± 0.06 < 0.001 ECG-LVGLS 14.44 ± 2.71 18.05 ± 1.05 < 0.001 15.60 ± 2.69 18.04 ± 1.09 < 0.001 QCG-RVD 0.14 ± 0.13 0.04 ± 0.05 < 0.001 0.14 ± 0.15 0.04 ± 0.05 < 0.001 ECG-RVGLS 18.56 ± 3.53 23.24 ± 1.49 < 0.001 19.18 ± 3.88 23.23 ± 1.50 < 0.001 Open in a new tab LVEF left ventricular ejection fraction, RVEF right ventricular ejection fraction, QCG quantitative electrocardiogram, LVD left ventricle dysfunction, LVGLS left ventricular global longitudinal strain, RVD right ventricle dysfunction, RVGLS right ventricular global longitudinal strain, LVH left ventricular hypertrophy, LAE left atrial enlargement. As expected, CMR-derived LVEF was significantly lower in the LVD group compared to those without LVD (36.9 ± 5.1% vs. 59.5 ± 6.1%; p < 0.001), and RVEF was markedly reduced in the RVD group (35.2 ± 5.7% vs. 57.4 ± 5.9%; p < 0.001). Corresponding AI-ECG scores were significantly higher among participants with ventricular dysfunction (all p < 0.001), highlighting the clear separation in AI-ECG values between normal and abnormal groups (Fig. 1 ). Fig. 1. Open in a new tab Flowchart illustrating AI-ECG algorithm development and validation dataset. The six AI-ECG models used in this study s– four targeting functional abnormalities and two targeting structural abnormalities – were previously developed using ECG datasets from Seoul National University Bundang Hospital (SNUBH). These models were validated against CMR-derived values in the UK Biobank population. CMR = cardiac magnetic resonance imaging, ECG electrocardiography, QCG quantitative ECG, LVD left ventricular dysfunction, RVD right ventricular dysfunction, GLS global longitudinal strain, LVH left ventricular hypertrophy, LAE left atrial enlargement. AI-ECG models demonstrated excellent performance in detecting LVD, with an area under the curve (AUC) of 0.887 (95% CI 0.866–0.907) for Quantitative ECG (QCG™)-LVD and 0.896 (95% CI 0.876–0.916) for ECG-LV global longitudinal strain (LVGLS). For RVD, AUCs were 0.778 (95% CI 0.751–0.804) for QCG-RVD and 0.825 (95% CI 0.800–0.850) for ECG-RVGLS (Fig. 2 ). Although performance for RVD detection was slightly lower than for LVD, the AI-ECG models still demonstrated substantial diagnostic value, particularly given the challenges of assessing RVD using ECG alone. In subgroup analyses, AI-ECG performance for LVD detection was particularly high in older individuals and males, suggesting potential demographic variations in model accuracy. Subgroup analyses revealed higher diagnostic accuracy for LVD in older individuals and males. For RVD, model performance remained consistent across subgroups stratified by body mass index (BMI) groups and individuals with and without comorbidities. Notably, the model exhibited slightly higher accuracy in subgroups with a higher prevalence of ventricular dysfunction, such as older individuals and those with ischemic heart disease. (Fig. 3 ) Using bootstrap resampling (200 iterations), bias-corrected calibration and discrimination metrics were estimated (Supplementary Fig. 1 , 2 , and Supplementary Table 2 ), which demonstrated reasonable calibration. Fig. 2. Open in a new tab Diagnostic performance of AI-ECG for left and right ventricular dysfunction. Receiver operating characteristic (ROC) curves for AI-ECG models predicting ( A ) left and ( B ) right ventricular dysfunction, as defined by CMR-derived eft ventricular ejection fraction (LVEF) and right ventricular ejection fraction (RVEF), respectively. Additionally, in the sub-cohort with available CMR-derived LVGLS data, the ECG-LVGLS model demonstrated robust performance in identifying abnormal LVGLS, defined as the lowest 1% of CMR-LVGLS distribution, with an AUC of 0.829 (95% CI 0.794–0.865). The ECG-LVGLS was significantly lower in participants with abnormal CMR-LVGLS. (Fig. 4 ) Fig. 3. Open in a new tab Subgroup analysis for ventricular dysfunction detection performances. ( A ) Forest plots showing AUC of QCG-LVD and ECG-LVGLS for LVD, and ( B ) QCG-RVD and ECG-RVGLS for RVD, across clinical subgroups stratified by age, sex, BMI, and comorbidities. AUC area under the curve. Other abbreviations are as Fig. 1 . AI-ECG performance for detecting structural abnormalities Table 3 presents baseline characteristics stratified by the presence of left ventricular hypertrophy (LVH) and left atrial enlargement (LAE), as defined by CMR. Participants with LVH had a higher BMI (27.7 ± 5.3 vs. 26.4 ± 4.3 kg/m²; p < 0.001) and were more likely to have hypertension (55.4% vs. 31.0%; p < 0.001), heart failure (2.3% vs. 0.5%; p < 0.001), and ischemic heart disease (12.7% vs. 5.7%; p < 0.001). Similarly, individuals with LAE were older (67.6 ± 7.4 vs. 62.8 ± 7.5 years; p < 0.001), predominantly male (66.2% vs. 47.7%; p < 0.001), and had higher rates of hypertension (61.0% vs. 30.9%; p < 0.001) and ischemic heart disease (18.8% vs. 5.7%; p < 0.001). Table 3. Baseline characteristics according to the presence of left ventricular hypertrophy and left atrial enlargement. LVH ( n = 213) No LVH ( n = 21,054) p LAE ( n = 213) No LAE ( n = 21,054) p Age, years 62.9 ± 8.0 62.8 ± 7.5 0.794 67.6 ± 7.4 62.8 ± 7.5 < 0.001 Male, sex 102 (47.9%) 10,080 (47.9%) 1.000 141 (66.2%) 10,041 (47.7%) < 0.001 Ethnicity 0.249 0.388 White 201 (94.4%) 20,443 (97.1%) 208 (97.7%) 20,436 (97.1%) Mixed 2 (0.9%) 87 (0.4%) 1 (0.5%) 88 (0.4%) Asian 5 (2.3%) 266 (1.3%) 2 (0.9%) 269 (1.3%) Black 3 (1.4%) 114 (0.5%) 0 (0.0%) 117 (0.6%) Others 1 (0.5%) 82 (0.4%) 0 (0.0%) 83 (0.4%) Unknown 1 (0.5%) 62 (0.3%) 2 (0.9%) 61 (0.3%) BMI, kg/m 2 27.7 ± 5.3 26.4 ± 4.3 0.001 27.1 ± 4.5 26.4 ± 4.3 0.019 Hypertension 118 (55.4%) 6,520 (31.0%) < 0.001 130 (61.0%) 6,508 (30.9%) < 0.001 Diabetes mellitus 15 (7.0%) 1122 (5.3%) 0.341 15 (7.0%) 1,122 (5.3%) 0.341 Dyslipidemia 56 (26.3%) 4,825 (22.9%) 0.279 77 (36.2%) 4,804 (22.8%) < 0.001 Heart Failure 5 (2.3%) 105 (0.5%) 0.001 12 (5.6%) 98 (0.5%) < 0.001 Ischemic heart disease 27 (12.7%) 1,208 (5.7%) < 0.001 40 (18.8%) 1,195 (5.7%) < 0.001 Cerebrovascular disease 6 (2.8%) 459 (2.2%) 0.691 9 (4.2%) 456 (2.2%) 0.070 CMR parameters LV mass (g) 135.1 ± 32.6 85.7 ± 21.4 < 0.001 111.7 ± 29.7 85.9 ± 21.8 < 0.001 LVMI (g/m 2 ) 71.3 ± 11.8 45.7 ± 8.0 < 0.001 57.9 ± 12.6 45.8 ± 8.3 < 0.001 LA volume (mm 3 ) 100.8 ± 34.2 72.4 ± 22.3 < 0.001 154.5 ± 29.8 71.8 ± 21.0 < 0.001 LAVI (mm 3 /m 2 ) 53.8 ± 17.8 38.9 ± 10.9 < 0.001 80.6 ± 13.9 38.6 ± 10.3 < 0.001 AI-ECG scores AI-ECG-LVH 0.51 ± 0.21 0.22 ± 0.14 < 0.001 0.45 ± 0.20 0.22 ± 0.14 < 0.001 AI-ECG-LAE 0.41 ± 0.20 0.19 ± 0.14 < 0.001 0.53 ± 0.25 0.19 ± 0.13 < 0.001 Open in a new tab LV mass left ventricle mass, LVMI left ventricle mass index, LA volume left atrium volume, LAVI left atrium volume index, LVH left ventricular hypertrophy, LAE left atrial enlargement. CMR-derived left ventricular mass index (LVMI) (76.6 ± 7.2 vs. 45.6 ± 7.9 g/m²; p < 0.001) and LAVI (80.6 ± 13.9 vs. 38.6 ± 10.3 mm³/m²; p < 0.001) were significantly elevated in LVH and LAE groups, respectively. AI-ECG scores were significantly elevated in participants with structural abnormalities. AI-ECG-LVH scores were markedly increased in participants with LVH (0.51 ± 0.21 vs. 0.22 ± 0.14; p < 0.001), and AI-ECG-LAE scores were notably higher among those with LAE (0.53 ± 0.25 vs. 0.19 ± 0.13; p < 0.001), demonstrating clear separation between groups. AI-ECG demonstrated strong diagnostic accuracy for identifying structural abnormalities. AI-ECG-LVH achieved an AUC of 0.877 (95% CI 0.851–0.902) for detecting LVH, while AI-ECG-LAE exhibited AUC of 0.883 (95% CI 0.858–0.908) for detecting LAE (Fig. 5 ). In subgroup analyses, predictive performance for LVH and LAE detection was robust across demographic categories, including age, sex, and BMI subgroups. The accuracy of AI-ECG was particularly strong in older individuals and males, reflecting trends similar to those observed in functional abnormality (ventricular dysfunction) analyses. Model performance was also enhanced in subgroups characterized by higher prevalence rates of cardiovascular comorbidities, notably hypertension and ischemic heart disease, highlighting the clinical relevance of AI-ECG as a potential non-invasive screening tool for structural cardiac abnormalities. (Fig. 6 ) The AI-ECG-LVH and AI-ECG-LAE scores demonstrated superior diagnostic performance compared with conventional ECG criteria. (Supplementary Fig. 3 ) Detailed diagnostic performance metrics—including sensitivity, specificity, positive predictive value, and negative predictive value—are summarized in Table 4 for each AI-ECG model. Additional sensitivity analyses using broader percentile-based definitions (top/bottom 1%, 2%, and 5%) demonstrated that model performance remained acceptable across the various thresholds tested (Supplementary Table 1 ). Fig. 4. Open in a new tab Diagnostic performance of AI-ECG for LVGLS. ( A ) ROC curve demonstrating the predictive performance of AI-ECG model for detecting abnormal LVGLS and ( B ) Boxplot of AI-ECG-LVGLS according to abnormal LVGLS. Abbreviations are as Figs. 1 and 2 . Fig. 5. Open in a new tab Diagnostic performance of AI-ECG for LVH and LAE. ROC curves demonstrating the predictive performance of AI-ECG models for detecting structural abnormalities: ( A ) LVH and ( B ) LAE defined by CMR. Abbreviations are as Figs. 1 and 2 . Table 4. Diagnostic performances of AI-ECG models. AI-ECG models AUROC Threshold* Sensitivity Specificity PPV NPV AUPRC F1 score QCG-LVD 0.887 (0.866–0.907) 0.048 77.1 (73.0-81.2) 88.1 (87.8–88.4) 6.2 (5.8–6.5) 99.7 (99.7–99.8) 0.264 (0.216–0.314) 0.114 ECG-LVGLS 0.896 (0.876–0.916) 17.08 80.5 (76.6–84.3) 86.5 (86.2–86.9) 5.7 (5.4-6.0) 99.8 (99.7–99.8) 0.300 (0.252–0.352) 0.106 QCG-RVD 0.778 (0.751–0.804) 0.060 64.5 (59.9–69.2) 80.5 (80.1–80.9) 3.2 (3.0-3.5) 99.6 (99.5–99.6) 0.083 (0.061–0.112) 0.062 ECG-RVGLS 0.825 (0.800–0.850) 21.37 62.2 (57.6–67.1) 91.1 (90.8–91.3) 6.6 (6.1–7.1) 99.6 (99.5–99.6) 0.206 (0.163–0.254) 0.119 AI-ECG-LVH 0.877 (0.851–0.902) 0.338 77.9 (72.3–83.1) 81.6 (81.1–82.2) 4.1 (3.8–4.4) 99.7 (99.7–99.8) 0.212 (0.156–0.267) 0.078 AI-ECG-LAE 0.862 (0.835–0.889) 0.303 76.1 (70.0-81.7) 82.8 (82.3–83.3) 4.3 (4.0-4.6) 99.7 (99.6–99.8) 0.206 (0.150–0.267) 0.081 Open in a new tab AUROC area under the receiver operating characteristic curve, PPV positive predictive value, NPV negative predictive value, AUPRC area under the precision-recall curve, *optimal threshold maximizing Youden’s index. Discussion In this study of 38,804 UK Biobank participants, we demonstrated that AI-ECG can accurately detect structural and functional cardiac abnormalities, as defined by CMR. Using the top 1% of extreme CMR-derived values to define abnormalities, AI-ECG models achieved high diagnostic performance across a range of conditions: left ventricular dysfunction (AUC 0.887–0.896), right ventricular dysfunction (AUC 0.778–0.825), left ventricular hypertrophy (AUC 0.824), and left atrial enlargement (AUC 0.883). These findings support the potential utility of AI-ECG as a scalable, non-invasive tool for population-level cardiovascular screening, extending its application beyond the hospital-based Korean cohorts in which the models were originally developed. Previous research has shown that AI-ECG can effectively identify various cardiac conditions. Notably, a Mayo Clinic study demonstrated reliable detection of left ventricular dysfunction using AI applied to ECGs, validated against echocardiography-defined ejection fraction 8 . Our group has similarly developed ECG image–based models that perform well in detecting structural and functional abnormalities 4 , 5 . However, prior studies have primarily focused on hospital-based or ethnically homogeneous populations. The present study uniquely validates these models in a multiethnic, community-based cohort from the UK Biobank, enhancing their generalizability and external validity. Despite differences in ethnicity and healthcare settings, our AI-ECG algorithms maintained high accuracy, suggesting the robust nature of ECG image-based analysis and generalizable convolutional neural network-based AI methodologies. This multiethnic validation significantly strengthens the external validity and potential clinical utility of AI-ECG for broad population-level screening. CMR is widely regarded as the gold standard for evaluating cardiac structure and function due to its reproducibility and accuracy, particularly for parameters such as ventricular volumes and myocardial mass. Compared with echocardiography, CMR avoids geometric assumptions and operator dependency, resulting in more reliable assessments 9 . Consequently, significant discrepancies between echocardiography and CMR-derived parameters have been consistently reported, with one study noting up to 60% discordance in patients with reduced LV ejection fraction (≤ 45%) 10 – 12 . This is especially important for right ventricular function, which is often difficult to evaluate via echocardiography due to its complex geometry and limited acoustic windows, making CMR a more reliable reference standard 13 . Given these inherent methodological differences and the population-based nature of the UK Biobank cohort, we defined cardiac structural and functional abnormalities using the top 1% CMR-measured values rather than applying traditional echocardiographic thresholds. This approach enabled identification of individuals with marked deviations from normal cardiac function and provided a rigorous reference standard for AI-ECG validation in a general population setting. Beyond confirming strong overall performance, our subgroup analyses offered additional insights into its performance across various demographic and clinical subgroups. Notably, diagnostic accuracy for both LVD and RVD was significantly higher among older participants and males compared to younger individuals and females. The presence of hypertension significantly enhanced the diagnostic accuracy of AI-ECG for both LVD and RVD, while diabetes and dyslipidemia tended to improve RVD detection. Regarding structural abnormalities, the diagnostic accuracy of AI-ECG for LVH was greater in participants with diabetes, whereas the predictive performance for LAE was enhanced among older individuals and those with higher BMI. These trends likely reflect the greater degree of cardiac remodeling in individuals with established cardiovascular risk factors, resulting in more pronounced ECG changes and improved model discrimination. Although performance was broadly consistent across subgroups, these findings suggest that AI-ECG may have particularly high clinical value in high-risk populations. Results of sensitivity analyses supported the interpretation that the models detect extreme CMR measurements across varying degrees of stringency. Notably, while the strict 1% definition yielded very high NPV, expanding the definition to 2% and 5% resulted in progressive improvements in PPV. This pattern suggests that the models capture a spectrum of extreme CMR measurements and may have potential utility as a rule-out–oriented screening tool for identifying individuals unlikely to harbor marked structural or functional deviations, rather than as a definitive diagnostic test. The robust diagnostic performance and high accessibility of AI-ECG demonstrated in this study suggest substantial implications for clinical practice. In real-world healthcare settings, AI-ECG could serve as a scalable, non-invasive initial screening tool, efficiently identifying individuals at higher cardiovascular risk who may benefit from more detailed diagnostic assessments or preventive interventions. For instance, AI-ECG could be integrated into routine primary care visits or community-based screening programs, promptly flagging patients with subtle but clinically meaningful abnormalities such as early-stage heart failure, asymptomatic left ventricular hypertrophy, or atrial remodeling indicative of increased atrial fibrillation risk. This early identification could facilitate timely intervention, potentially reducing morbidity and mortality associated with undetected cardiac conditions. Furthermore, because our AI-ECG algorithms operate on standard printed ECG images, their implementation requires minimal infrastructural changes, allowing rapid integration into diverse healthcare environments, including resource-limited settings where access to advanced cardiac imaging modalities remains challenging. Ultimately, our results highlight the practical and clinical value of incorporating AI-ECG into broader cardiovascular disease prevention and management strategies, emphasizing its potential to improve patient outcomes on a population scale. This study has several limitations. While the UK Biobank dataset expands research from Korean populations to a predominantly European cohort, this represents a cross-ethnic validation between the development and validation cohorts. However, its primarily White demographic composition may still limit the generalizability within the validation population, highlighting the need for validation in more ethnically diverse cohorts. The top 1% CMR values offers reasonable and practical cut-off value for defining the functional and structural abnormalities in general population, but may not align with previously established thresholds, potentially limiting its application in routine clinical workflows where fixed cut-offs are standard. Moreover, good calibration in a retrospective cohort does not automatically translate to prospective performance. The healthy volunteer effect in the UK Biobank may introduce spectrum bias, potentially impacting model performance compared to more diverse or high-acuity clinical populations. Furthermore, identifying extreme percentile metrics focuses on population-based structural deviations rather than clinical disease, a distinction that is essential for interpreting the model’s diagnostic accuracy. The black-box nature of AI-ECG models may pose challenges for clinical adoption, underscoring the need for further efforts to improve interpretability. Lastly, while our AI-ECG algorithms demonstrated robust diagnostic performance, this study was limited by the lack of longitudinal follow-up data. Future prospective studies are required to evaluate the prognostic utility of these AI-ECG scores and to confirm whether their application can meaningfully enhance cardiovascular outcomes. In conclusion, the newly developed AI-ECG scores are highly effective in detecting functional and structural cardiac abnormalities within a general population. These findings highlight AI-ECG’s potential as a cost-effective and accessible screening tool, addressing limitations in availability and diagnostic accuracy associated with current imaging modalities, such as echocardiography and CMR. Methods Study population This study utilized data from the UK Biobank, a large-scale, prospective cohort comprising over 500,000 participants aged 40 to 69 years, recruited between 2006 and 2010. The cohort includes data collected from hospital records and health check-ups in the general population. Detailed protocols for recruitment and imaging have been previously described 14 . Among the 502,386 participants, 61,292 underwent imaging visits that included both ECG and CMR. According to the UK Biobank imaging protocol, the resting 12-lead ECG and cardiac MRI were performed during the same imaging visit. We excluded 10,808 participants without ECG data and 11,680 without available CMR parameters, yielding a final study population of 38,804 individuals (Fig. 1 ). A sub-cohort of 21,267 participants with complete structural CMR measurements, specifically left ventricular mass and left atrial volume, was used for the evaluation of structural abnormalities. This sub-cohort also included CMR-derived left ventricular global longitudinal strain (LVGLS) data, allowing additional validation of AI-ECG models targeting myocardial strain. This study was conducted under the ethical approval granted to the UK Biobank by the National Health Service National Research Ethics Service (Ref 11/NW/0382, extended under Ref 16/NW/0274). This study adhered to the principles of the Declaration of Helsinki. Access to anonymized participant data was granted through an approved UK Biobank application. Fig. 6. Open in a new tab Subgroup analysis for structural abnormality detection. Forest plots showing ( A ) AUCs of AI-ECG-LVH and ( B ) AI-ECG-LAE across demographic and clinical subgroups. Abbreviations are as Figs. 1 and 3 . AI algorithms We evaluated six previously developed AI-ECG scores: four targeting functional abnormalities (QCG-LVD for left ventricular dysfunction, QCG-RVD for right ventricular dysfunction, ECG-LVGLS for LV global longitudinal strain, and ECG-RVGLS for RV global longitudinal strain) and two targeting structural abnormalities (AI-ECG-LVH for LV hypertrophy and AI-ECG-LAE for LA enlargement). These models are based on modified convolutional neural networks (CNNs) that utilize a shared encoder and task-specific output layers to analyze ECG signals or images. In all six models, the shared encoder is a 2D CNN feature extractor that maps a single ECG image input (signal-derived data are transformed into an image input) into a fixed-length latent representation, and each task-specific head is a multilayer network that transforms this shared representation into a task-specific output. The image input is a standardized 12-lead ECG report image, and each task-specific head produces one corresponding AI-ECG score. The encoder is a ResNet-based neural network with squeeze-excitation layers and a non-local network block. For QCG-LVD, QCG-RVD, AI-ECG-LVH, and AI-ECG-LAE, the task head outputs a continuous probability-like score (sigmoid output), whereas for ECG-LVGLS and ECG-RVGLS the task head outputs a continuous regression estimate (linear output). The development of QCG-LVD, QCG-RVD, and ECG-LVGLS has been described in previous papers, and ECG-RVGLS was developed in an identical manner to ECG-LVGLS 4 , 5 , 15 , 16 . Briefly, the QCG-LVD and QCG-RVD models were developed as part of the QCG™ framework and are embedded in the ‘ECG Buddy’ software, which received regulatory approval from the Korean Ministry of Food and Drug Safety (MFDS) in January 2024. The encoder component was pretrained on 49,731 public ECG datasets via self-supervised learning and fine-tuned using 47,194 annotated ECG images from 32,968 patients at Seoul National University Bundang Hospital between 2017 and 2019. This multi-task learning framework enabled the models to classify rhythms and predict critical cardiovascular conditions such as shock, cardiac arrest, acute coronary syndrome, and heart failure 5 , 15 , 16 . For their labels, QCG-LVD and QCG-RVD were trained using echocardiography-derived reference standards linked to the corresponding clinical ECGs (binary labels reflecting the presence/absence of LV systolic dysfunction [LVEF < 40%] and the presence of RV dysfunction based on qualitative statements in echocardiography reports). The ECG-LVGLS and ECG-RVGLS models were developed using a transfer learning approach. The CNN encoder from the QCG system was retained, while new task-specific networks were trained to output ECG-based global longitudinal strain estimates. Training utilized data from 2882 patients across four Korean hospital cohorts, with external validation in an independent sample 4 . For these regression models, the training target was the echocardiography-derived (speckle-tracking) global longitudinal strain value. The AI-ECG-LVH and AI-ECG-LAE models share a similar CNN architecture and were trained on 254,356 standard 12-lead ECGs matched with echocardiography results from 139,877 patients at Seoul National University Bundang Hospital between 2003 and 2020. Their labels were also derived from echocardiography reports (binary targets reflecting LVH defined as LVMI > 115 g/m² for males and > 95 g/m² for females, and LAE defined as LAVI > 34mL/m²) according to the original development definitions. Only ECG-echocardiography pairs acquired within a 7-day interval were included to ensure temporal consistency. This study adhered to the principles of the Declaration of Helsinki and was approved by the institutional review board (IRB No. B-2306-832-101). AI analysis of ECG Resting ECGs (RestECG, Field 20205) were obtained from participants during initial imaging (instance 2), or repeat imaging (instance 3) visits and were provided by the UK Biobank in XML format ( https://biobank.ndph.ox.ac.uk/showcase/field.cgi?id=20205 ). ECGs were acquired at UK Biobank imaging assessment centres using GE CardioSoft software, with automated measurements generated by the Marquette 12SL algorithm. Voltage signals for the 12 standard leads (I, II, III, aVR, aVL, aVF, V1–V6) were extracted by parsing each < WaveformData> element. The < Resolution units="uVperLsb”>5</Resolution > tag indicates that each integer tick represents 5 µV. Therefore, the raw values were multiplied by 5 to convert them to µV and then divided by 1,000 to express them in mV. The resulting 10-second recordings sampled at 500 Hz (5000 samples per lead) were split into four consecutive 2.5-second segments and arranged in the first three rows in a 3 × 4 format (columns: 1st— I, II, III; 2nd— aVR, aVL, aVF; 3rd— V1, V2, V3; 4th— V4, V5, V6) with 10-second Lead II tracing as a 4th row. The choice of this standard ECG format reflects its widespread use in clinical practice and matches the input format covered by the evaluated models. All six models received the same rendered report image in PNG format as input; the shared encoder produced a latent feature representation, and each task-specific head generated its corresponding score. CMR image acquisition and parameter extraction CMR imaging in the UK Biobank was conducted using 1.5-Tesla MAGNETOM Aera scanners (Siemens Healthcare, Syngo Platform VD13A), following a previously established protocol 17 . Quantitative cardiac and aortic parameters were extracted using a validated automated analysis pipeline powered by machine learning 18 . This pipeline, based on convolutional neural networks, performed segmentation of cardiac and aortic structures from short-axis, long-axis, and cine images. Specifically, short-axis cine images were used to segment the ventricles and myocardium through a fully convolutional neural network trained on expert-labeled data from 3,975 individuals. This enabled the computation of ventricular volumes, myocardial mass, and ejection fractions. Comprehensive morpho-functional measurements were derived for all four cardiac chambers and two aortic segments. To account for body size, all volume-based measurements were indexed to body surface area. Definitions In this study, abnormalities in CMR parameters were operationally defined to assess various aspects of cardiac function and structures. We utilized preprocessed CMR measurements provided by the UK Biobank imaging project, which were generated through standardized imaging protocols and automated post-processing pipelines developed and validated in prior studies 17 , 18 . Specifically, we focused on LVD, RVD, LVH, and LAE. LVD and RVD were defined using CMR-derived LVEF, RVEF, or global longitudinal strain (LVGLS and RVGLS), respectively. LVH and LAE were defined based on left ventricular mass and left atrial maximum volume, respectively, both indexed to body surface area. To account for population-based variability, abnormalities were operationally defined as the top 1% values of the cohort distribution for each CMR-derived parameter, as surrogate indicators of abnormal findings. For LVH, the cutoff was determined separately for males and females, reflecting sex-specific differences in normative values of LV mass index, as commonly applied in echocardiographic criteria 19 . The specific cutoff values corresponding to the top 1% for each parameter were 42.2% for LVEF, 40.7% for RVEF, 74.3 g/m 2 (men)/56.9 g/m 2 (women) for LVMI, 68.1 mL/m 2 (men)/73.4 mL/m 2 (women) for LAVI, and 11.9% for LVGLS. The performance of the AI-ECG algorithm in detecting these defined abnormalities were evaluated, aiming to determine its effectiveness in identifying individuals with marked deviations from the population distribution and its potential applicability in a clinical context. Statistical analysis Continuous variables were reported as mean ± standard deviation, and categorical variables as counts and percentages. The predictive performance of AI-ECG models was assessed using the area under the receiver operating characteristic curve (AUROC) with 95% confidence interval (CI) calculated using DeLong’s method. The area under the precision-recall curve (AUPRC) and its bootstrapped 95% CI (1,000 replications) were additionally calculation. The optimal thresholds were determined using the Youden index. At these optimal thresholds, we calculated sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and the F1-score. Subgroup analyses were performed across various clinical characteristics, calculating the AUROC and 95% CI for each subgroup, and P for difference was calculated using DeLong’s test for unpaired ROC curves. Model calibration was visualized using calibration plots with non-parametric locally weighted scatterplot smoothing (LOESS) curves and risk deciles. Quantitative performance was evaluated using bias-corrected calibration slope, intercept, and C-statistic derived from 200 bootstrap resamples. Additionally, overall accuracy and error were assessed using the Brier score, average and maximum calibration errors (E avg and Emax). Decision curve analysis was additionally performed by calculating the net benefit across a range of threshold probabilities, comparing the model against default strategies of treating all or no patients. To further address concerns regarding clinical utility and the effect of fixed low prevalence, we conducted additional sensitivity analyses using broader percentile-based definitions (top/bottom 1%, 2%, and 5%). A p -value < 0.05 was considered statistically significant. All analyses were performed using R (version 4.4.1; https://www.R-project.org ). Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (25.2KB, docx) Supplementary Material 2 (2.1MB, tif) Supplementary Material 3 (1.7MB, tif) Supplementary Material 4 (714.8KB, tif) Acknowledgements None. Author contributions Conceptualization: YK, HL, JK, TMR, and YCData Curation: YK, HL, JK, TMR, and YCFormal Analysis: YK and HLMethodology: JK, TMR, and YCProject Administration: YCWriting—Original Draft: YK, HLWriting—Review & Editing: YK, HL, TMR, and YC. Funding 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 (grant number: RS-2023-00265933). Data availability Data used in this study will be made available on reasonable request to the corresponding authors. Declarations Competing interests Joonghee Kim, MD, PhD developed the algorithm. He also founded a start-up company, ARPI Inc., where he serves as the CEO. Youngjin Cho, MD, PhD works for the company as a research director. Haemin Lee works for the company as data scientist. Otherwise, there is no conflict of interest for the other authors. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Yerin Kim and Haemin Lee are contributed equally to this work. Contributor Information Tae-Min Rhee, Email: [email protected]. Youngjin Cho, Email: [email protected]. References 1. Attia, Z. I. et al. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nat. Med. 25 , 70–74. 10.1038/s41591-018-0240-2 (2019). [ DOI ] [ PubMed ] [ Google Scholar ] 2. Yao, X. et al. Artificial intelligence-enabled electrocardiograms for identification of patients with low ejection fraction: a pragmatic, randomized clinical trial. Nat. Med. 27 , 815–819. 10.1038/s41591-021-01335-4 (2021). [ DOI ] [ PubMed ] [ Google Scholar ] 3. Sau, A. et al. Artificial intelligence-enabled electrocardiogram for mortality and cardiovascular risk estimation: a model development and validation study. Lancet Digit. Health . 6 , e791–e802. 10.1016/S2589-7500(24)00172-9 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 4. Choi, H. M. et al. AI derived ECG global longitudinal strain compared to echocardiographic measurements. Sci. Rep. 14 , 26458. 10.1038/s41598-024-78268-8 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Cho, Y. et al. Artificial intelligence-based electrocardiographic biomarker for outcome prediction in patients with acute heart failure: Prospective cohort study. J. Med. Internet Res. 26 , e52139. 10.2196/52139 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Park, J. et al. Artificial intelligence-enhanced electrocardiography analysis as a promising tool for predicting obstructive coronary artery disease in patients with stable angina. Eur. Heart J. Digit. Health . 5 , 444–453. 10.1093/ehjdh/ztae038 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Choi, J. et al. Artificial intelligence predicts undiagnosed atrial fibrillation in patients with embolic stroke of undetermined source using sinus rhythm electrocardiograms. Heart Rhythm . 21 , 1647–1655. 10.1016/j.hrthm.2024.03.029 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 8. Kashou, A. H. et al. Artificial intelligence-augmented electrocardiogram detection of left ventricular systolic dysfunction in the general population. Mayo Clin. Proc. 96 , 2576–2586. 10.1016/j.mayocp.2021.02.029 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Alfakih, K., Reid, S., Jones, T. & Sivananthan, M. Assessment of ventricular function and mass by cardiac magnetic resonance imaging. Eur. Radiol. 14 , 1813–1822. 10.1007/s00330-004-2387-0 (2004). [ DOI ] [ PubMed ] [ Google Scholar ] 10. Clark, J. et al. Interchangeability in left ventricular ejection fraction measured by echocardiography and cardiovascular magnetic resonance: Not a perfect match in the real world. Curr. Probl. Cardiol. 48 , 101721. 10.1016/j.cpcardiol.2023.101721 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 11. Marcos-Garces, V. et al. Ejection fraction by echocardiography for a selective use of magnetic resonance after infarction. Circ. Cardiovasc. Imaging . 13 , e011491. 10.1161/CIRCIMAGING.120.011491 (2020). [ DOI ] [ PubMed ] [ Google Scholar ] 12. Marwick, T. H., Neubauer, S. & Petersen, S. E. Use of cardiac magnetic resonance and echocardiography in population-based studies: Why, where, and when? Circ. Cardiovasc. Imaging . 6 , 590–596. 10.1161/CIRCIMAGING.113.000498 (2013). [ DOI ] [ PubMed ] [ Google Scholar ] 13. Bowen, D. J. et al. Comparison of advanced echocardiographic right ventricular functional parameters with cardiovascular magnetic resonance in adult congenital heart disease. Eur. Heart J. Imaging Methods Pract. 1 , qyad033. 10.1093/ehjimp/qyad033 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Sudlow, C. et al. UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med. 12 , e1001779. 10.1371/journal.pmed.1001779 (2015). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Choi, Y. J. et al. Screening for RV dysfunction using smartphone ECG analysis app: Validation study with acute pulmonary embolism patients. J. Clin. Med. 13 10.3390/jcm13164792 (2024). [ DOI ] [ PMC free article ] [ PubMed ] 16. Kim, J. H. et al. Non-Inferiority analysis of electrocardiography analysis application vs. point-of-care ultrasound for screening left ventricular dysfunction. Yonsei Med. J. 66 , 172–178. 10.3349/ymj.2024.0148 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Pennell, D. J. et al. Review of Journal of Cardiovascular Magnetic Resonance 2015. J. Cardiovasc. Magn. Reson. 18 10.1186/s12968-016-0305-7 (2016). [ DOI ] [ PMC free article ] [ PubMed ] 18. Bai, W. et al. A population-based phenome-wide association study of cardiac and aortic structure and function. Nat. Med. 26 , 1654–1662. 10.1038/s41591-020-1009-y (2020). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Oikonomou, E. et al. Diagnostic performance of electrocardiographic criteria in echocardiographic diagnosis of different patterns of left ventricular hypertrophy. Ann. Noninvasive Electrocardiol. 25 , e12728. 10.1111/anec.12728 (2020). [ 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. Supplementary Materials Supplementary Material 1 (25.2KB, docx) Supplementary Material 2 (2.1MB, tif) Supplementary Material 3 (1.7MB, tif) Supplementary Material 4 (714.8KB, tif) Data Availability Statement Data used in this study will be made available on reasonable request to the corresponding authors. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (2.4 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top