An Artificial Intelligence Algorithm for Early Detection of Left Ventricular Systolic Dysfunction in Patients with Normal Sinus Rhythm.

Most previous studies using artificial intelligence (AI) to detect left ventricular systolic dysfunction (LVSD) from electrocardiograms (ECGs) relied on data obtained near the time of echocardiography or included patients with known cardiac disease, limiting their specificity for screening. We aimed to evaluate whether AI models could predict future LVSD from ECGs interpreted as normal and recorded one to two years before echocardiography. We retrospectively analyzed 24,203 sinus rhythm ECGs from 11,131 patients. Two convolutional neural network models (DenseNet-121 and ResNet-101) were trained (70%), validated (10%), and tested (20%) to predict LVSD (defined as ejection fraction ≤50%). Survival analysis was performed using Kaplan-Meier curves and the log-rank test. Of the total population, 2734 patients had LVSD and 8397 had preserved EF. DenseNet-121 and ResNet-101 demonstrated excellent discrimination for LVSD with AUROCs of 0.930 and 0.925, accuracies of 0.887 and 0.860, sensitivities of 0.821 and 0.856, and specificities of 0.908 and 0.861, respectively. In the test set, patients predicted to have LVSD showed a significantly higher risk of echocardiographic LVSD (hazard ratio 9.89, 95% CI 8.20-11.92, = 0.005) and lower 24-month survival (log-rank < 0.001). AI-enabled ECG models predicted future LVSD from clinically normal ECGs recorded up to two years prior to imaging. These findings suggest a potential role for AI-ECG in the early detection of subclinical LVSD and improved risk stratification in asymptomatic individuals.