Personalized artificial intelligence based left ventricular ejection fraction and systolic dysfunction assessment.

Left-ventricular (LV) ejection fraction (LVEF) is a fundamental measure of cardiac function, typically assessed with resource-intensive imaging techniques, such as transthoracic echocardiography (TTE). We evaluated the electrocardiogram (ECG) as an alternative, easily accessible data to estimate LVEF in a large cohort of 191,941 patients, comprising 236,623 ECG/TTE pairs. Using either the ECG data alone or with structured features, we developed convolutional and probabilistic neural network models to estimate LVEF and quantify its uncertainty. The ECG-only model achieved a mean-absolute-error (MAE) of 7.71% and a root-mean-square-error (RMSE) of 10.36%, while the hybrid model achieved an MAE of 7.84% and an RMSE of 10.52%. Personalized models significantly improved performance, achieving MAEs of 5.98% (ECG-only) and 6.75% (hybrid). LV systolic dysfunction (LVEF ≤ 40%) was identified with an AUC of 0.88, sensitivity of 0.92 and negative predictive value of 0.98. The presented models demonstrated excellent performance in estimating LVEF and screening of LV systolic dysfunction.