Deep Learning-Based Automated Echocardiographic Quantification of Left Ventricular Ejection Fraction: A Point-of-Care...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Automated quantification of left ventricular (LV) ejection fraction (EF) typically relies on specific echocardiographic views.
• Apical 2-chamber views, often difficult to obtain in point-of-care (POC) settings, limit current automated approaches.
• Physicians commonly use apical 4-chamber and parasternal long-axis views for visual assessment in POC scenarios.

Purpose of the Study:

• To adapt and validate an automated machine-learning algorithm for LV EF quantification using commonly available POC echocardiographic views.
• To assess the accuracy of automated LV function classification compared to experienced physicians.
• To determine the algorithm's utility in a real-world POC setting with images acquired by nurses.

Main Methods:

• An automated machine-learning algorithm was adapted to utilize apical 4-chamber, parasternal long-axis, or combined views.
• Protocol 1 involved 166 clinical echocardiographic examinations comparing automated EF to reference biplane measurements and physician visual estimates.
• Protocol 2 tested the algorithm on images acquired by nurses using a portable system in a POC setting.

Main Results:

• Protocol 1 showed good agreement (intraclass correlation, 0.86-0.95) between automated and reference EF, with bias <2%.
• Automated classification accuracy for LV function was comparable to physician visual assessment.
• Protocol 2 demonstrated excellent agreement (intraclass correlation=0.84) with minimal bias (2.5±6.4%) in a POC setting.

Conclusions:

• The adapted machine-learning algorithm provides accurate automated LV function evaluation from standard POC echocardiographic views.
• This technology empowers POC personnel to reliably assess cardiac function, improving patient care.
• The algorithm enhances the utility of echocardiography in resource-limited or rapid assessment environments.