Deep Learning-Based Prediction of Right Ventricular Ejection Fraction Using 2D Echocardiograms

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Right ventricular (RV) function is a crucial independent prognostic indicator, even in patients with left-sided heart disease.
• Conventional 2D echocardiography has limitations in assessing RV function compared to 3D echocardiography-derived RVEF.
• Accurate RV function assessment is vital for patient management and risk stratification.

Purpose of the Study:

• To develop and implement a deep learning (DL) tool for estimating RVEF from 2D echocardiographic videos.
• To benchmark the DL tool's performance against human expert interpretation.
• To evaluate the prognostic significance of DL-predicted RVEF values.

Main Methods:

• Retrospective analysis of 831 patients with 3D RVEF measurements.
• Training spatiotemporal convolutional neural networks on 3,583 2D echocardiographic videos (80:20 train-validation split).
• External validation of an ensemble DL model on 1,493 videos from 365 patients with a median follow-up of 1.9 years.

Main Results:

• The DL ensemble model achieved a mean absolute error of 4.57% (internal) and 5.54% (external) for RVEF prediction.
• The model identified RV dysfunction with 78.4% accuracy, comparable to expert readers (77.0%).
• DL-predicted RVEF was independently associated with major adverse cardiac events (HR: 0.924, P=0.025).

Conclusions:

• A DL-based tool can accurately assess RV function using only 2D echocardiographic videos.
• The tool demonstrates comparable diagnostic and prognostic capabilities to 3D imaging.
• This AI approach offers a promising method for enhancing RV function assessment in clinical practice.