ViViEchoformer: Deep Video Regressor Predicting Ejection Fraction.

Area of Science:

• Cardiology and Artificial Intelligence
• Medical Imaging Analysis
• Machine Learning in Healthcare

Background:

• Heart disease is the leading global cause of mortality.
• Accurate ejection fraction (EF) measurement is crucial for cardiac function assessment and patient outcomes.
• Current echocardiogram interpretation by humans has limitations due to reader variability.

Purpose of the Study:

• To introduce ViViEchoformer, a deep learning model for automated left ventricular ejection fraction (LVEF) prediction from echocardiogram videos.
• To evaluate the accuracy and reliability of ViViEchoformer compared to human interpretation.

Main Methods:

• Utilized a video vision transformer (ViViEchoformer) to directly regress LVEF from echocardiogram videos.
• Trained and validated the model on a dataset of 10,030 apical-4-chamber echocardiography videos from Stanford University Hospital.
• The model extracts spatiotemporal tokens to capture spatial information and inter-frame relationships.

Main Results:

• ViViEchoformer achieved a mean absolute error of 6.14% for EF prediction.
• The model demonstrated strong performance in predicting heart failure with reduced ejection fraction (HFrEF), with an AUC of 0.83 and 87% classification accuracy.
• The approach accurately quantifies left ventricular function, providing precise, automated EF predictions.

Conclusions:

• ViViEchoformer offers a precise and reliable video-based method for left ventricular function quantification.
• The model serves as a valuable tool to aid human assessment, reducing variability in echocardiogram interpretation.
• This deep learning approach establishes a fundamental basis for automated echocardiogram analysis.