Abstract 13971: ECG Identifies Diastolic Dysfunction and Predicts Future Sudden Cardiac Arrest in Patients with Preserved Ejection Fraction
Abstract
Background: Diastolic dysfunction (DD) is a common precursor of heart failure and an important marker of adverse cardiovascular outcomes, yet its identification relies largely on echocardiography. We hypothesized that a convolutional neural network (CNN) applied to the standard 12-lead electrocardiogram (ECG) could identify the presence and severity of DD in patients with preserved left ventricular ejection fraction (LVEF) and that ECG-derived estimates of DD would be associated with future risk of sudden cardiac arrest (SCA). Methods: We retrospectively identified patients with preserved LVEF (≥ 50%) who underwent clinically indicated echocardiography and had a contemporaneous digital 12-lead ECG. ECG waveforms, along with age and sex were used to train, validate and test a CNN model for detection of DD on 64%, 16%, and 20% of the total dataset, respectively. Models were trained to discriminate (1) normal diastolic function versus any DD and (2) increasing grades of DD (Grade I, Grade II, and Grade III). Probabilities generated by the model were subsequently used to predict incident SCA. Results: The study included 19,514 patients (mean age 61±17years; 54% women) with preserved LVEF and paired ECG-echocardiographic data. In the holdout test set, the CNN identified the presence of any DD with an area under the ROC curve (AUC) of 0.78 and an F1 score of 0.45. Performance for discrimination of higher grades of DD yielded AUCs of 0.82 for grade II-III and 0.87 for Grade III. Patients with DD that could not be determined by echocardiography had ROC of 0.83. In the detection of any DD, the sensitivity, specificity, positive predictive value, and negative predictive value were 70%, 73%, 33%, and 93%, respectively. At 12 months follow-up, 0.5% of patients in the holdout set experienced SCA. Higher predicted risk of any DD by ECG was associated with increased risk of SCA (HR=17.8, p = 0.005, Figure). Conclusions: A deep learning model applied to ECGs can identify the presence of DD with improved discrimination with higher grades of DD, in patients with preserved LVEF. ECG-derived estimates of DD were strongly associated with future risk of SCA, suggesting that CNN-enabled ECG analysis may provide a scalable, noninvasive approach for identifying patients with occult myocardial dysfunction and increased arrhythmic risk.