Deep Learning-Driven Automated Assessment of Left Ventricular Diastolic Function in Echocardiography

This study sought to evaluate the feasibility, accuracy, and diagnostic performance of a fully automated deep learning model for assessing left ventricular diastolic function (LVDF) using 2D transthoracic echocardiography (TTE).

In this prospective observational study, 302 patients underwent 2D TTE for suspected diastolic dysfunction. Diastolic parameters, such as mitral inflow velocities, tissue Doppler indices, left atrial volumes, and tricuspid regurgitation velocity, were automatically analyzed using AI-based software (Ligence Heart) and compared with expert manual measurements. The AI achieved a maximal measurement success rate for spectral and tissue Doppler parameters. Strong correlation was observed for key variables such as E velocity (r = 0.93), A velocity (r = 0.88), E/A ratio (r = 0.94), and LAVi (r = 0.92), while lower agreement was noted for TR velocity. Classification of diastolic dysfunction showed high accuracy in identifying normal and severe grades, with more variability in intermediate categories. No consistent bias or directional error was observed across measured parameters.

Automated AI-based evaluation of diastolic function using 2D TTE is feasible and provides accurate, guideline-based measurements with strong correlation to expert interpretation. The software demonstrated promising results in classifying LVDF grades, particularly in distinguishing normal and severe dysfunction. This approach shows potential to enhance diagnostic consistency and efficiency in clinical practice, although further validation is needed.