Why the study?
Diagnosing left ventricular diastolic dysfunction relies on Doppler echocardiography, which limits large-scale screening, and AI-enabled ECG detection of diastolic dysfunction remains less explored than systolic dysfunction.
Does a Vision Transformer deep learning model applied to 12-lead ECGs accurately identify guideline-defined left ventricular diastolic dysfunction compared to echocardiography in patients at a tertiary academic medical center?
Observational (n=21,385)
No
Does a Vision Transformer deep learning model applied to 12-lead ECGs accurately identify guideline-defined left ventricular diastolic dysfunction compared to echocardiography in patients at a tertiary academic medical center?
A Vision Transformer deep learning model applied to 12-lead ECGs achieved high discrimination (AUROC 0.913) and negative predictive value (96.6%) for detecting left ventricular diastolic dysfunction, supporting its potential utility as a scalable, non-invasive rule-out screening tool for early HFpEF.
High NPV supports rule-out potential in retrospective data; extends AI-ECG to LVDD but leaves open prospective validation.
Background Left ventricular diastolic dysfunction (LVDD) is highly prevalent in the general population and represents an early stage of heart failure with preserved ejection fraction (HFpEF). Diagnosis relies on Doppler-based echocardiographic parameters, limiting its feasibility for large-scale screening. While artificial intelligence (AI)–enabled electrocardiography has shown strong performance for detecting systolic dysfunction, ECG-based detection of LVDD remains less explored. We evaluated whether a deep learning model could identify guideline-defined LVDD using 12-lead ECGs. Methods We retrospectively analyzed paired ECG and echocardiographic examinations obtained between 2020 and 2023 at a tertiary academic medical center. LVDD was defined according to the 2016 ASE/EACVI recommendations using four parameters: mean E/e′ ratio (>14), septal e′ velocity (<0.07 m/s) or lateral e′ velocity (<0.10 m/s), tricuspid regurgitation velocity (>2.8 m/s), and left atrial volume index (>34 mL/m²). Patients meeting more than half of the evaluable criteria were classified as LVDD. ECG–echocardiography pairs were matched within ±14 days. A Vision Transformer pretrained on unlabeled ECGs was fine-tuned to classify LVDD. Data were split at the patient level into training, validation, and test sets (7:1:2). After exclusions, 46,564 ECG–echocardiography pairs from 21,385 patients were included. Results LVDD prevalence in the test cohort was 15.3%. The model achieved an AUROC of 0.908 (95% CI 0.893–0.922) in the validation set and 0.913 (95% CI 0.903–0.924) in the test set. Using a threshold defined in the validation cohort where sensitivity equaled specificity, the model demonstrated sensitivity 0.838, specificity 0.825, and overall accuracy 0.827. The positive predictive value was 0.463, whereas the negative predictive value reached 0.966, supporting potential utility as a rule-out screening tool. Conclusion A deep learning model using 12-lead ECGs demonstrated strong discrimination for detecting LVDD with an AUROC exceeding 0.90. ECG-based AI screening may provide a scalable and non-invasive strategy for early identification of diastolic dysfunction and facilitate earlier intervention before progression to symptomatic HFpEF.
No takes yet. Share an insight, caveat, or question.
Soh et al. (2026) conducted an observational in Left ventricular diastolic dysfunction (LVDD) (n=21,385). Vision Transformer deep learning model on 12-lead ECG vs. 2016 ASE/EACVI guideline-defined LVDD on echocardiography was evaluated on Detection of guideline-defined left ventricular diastolic dysfunction (LVDD) in the test set (AUROC 0.913, 95% CI 0.903-0.924).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: