Respect The Apex - Deep Learning Enhanced Estimation of Left Ventricular Ejection Fraction: Application in Transthoracic and Point-of-Care Echocardiography

Transthoracic echocardiography (TTE) is increasingly utilized by point-of-care ultrasound (POCUS). This study aimed to evaluate whether artificial intelligence (AI) could estimate left ventricular ejection fraction (LVEF) using standard TTE without measuring chamber size or requiring image segmentation, focusing on the parasternal long-axis and apical four-chamber views. The Aisap.ai algorithm was applied in a deep learning network training with data from 76,342 patients for training (62%), 22,825 for validation (19%), and 20,960 for testing (17%), all of whom underwent comprehensive transesophageal echocardiography (TEE) studies between 2007 and 2021. An additional cohort of 556 patients from a different hospital provided external validation, while a prospective validation was conducted with a POCUS cohort of 237 patients. The final hold-out cohort included 20,960 patients with a median age of 66 years, of whom 58% were men. LVEF was estimated as a continuous variable, yielding a mean average error (MAE) of 3.4, a root mean square error (RMSE) of 4.83, and a Pearson correlation coefficient of 0.92. Results were consistent across the external validation and prospective POCUS cohort. During a median follow-up of five years, 4,682 patients (23.1%) died. A multivariate Cox regression model indicated that decreased LVEF estimation by the AI was linked to a 39% increased risk of death (95% CI 1.31-1.48, p<0.001). Consistent outcomes were found in a sub-group analysis of patients with preserved and reduced ejection fractions. Additionally, those in the discrepancy group faced a 58% increased risk of heart failure hospitalizations (95% CI 1.32-1.89, p<0.001). In conclusion, AI can effectively estimate LVEF and serves as an independent predictor of heart failure hospitalization and survival, suggesting its potential for risk stratification at the point of care.