Machine learning algorithm using publicly available echo database for simplified "visual estimation" of left...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Automated left ventricular ejection fraction (LV EF) calculation typically requires complex algorithms and precise endocardial border tracing.
• This complexity limits the usability of automated echocardiography at the bedside, particularly in point-of-care ultrasound (POCUS) scenarios.

Purpose of the Study:

• To develop a straightforward deep learning (DL) regression algorithm for visual estimation of LV EF using patient echocardiogram data.
• To compare the DL algorithm's EF estimations against standard echocardiography laboratory calculations.

Main Methods:

• A VGG16-based DL architecture integrated with a long short-term memory (LSTM) algorithm was used for analyzing echocardiogram videos.
• The algorithm was trained on visual echo video data correlated with calculated EF, excluding volumetric or coordinate data, to simulate POCUS conditions.
• Performance was evaluated using mean absolute error (MAE) and root mean square error (RMSE) against laboratory EF calculations on a separate test dataset.

Main Results:

• The DL algorithm achieved a mean absolute error (MAE) of 8.08% in visually estimating LV EF, indicating good performance relative to human experts (MAE typically 4-5%).
• The root mean square error (RMSE) was 11.98, with a correlation coefficient of 0.348.

Conclusions:

• A simplified DL algorithm demonstrates potential for accurate visual estimation of LV EF from short echocardiogram clips.
• This less complex approach may be more suitable for POCUS applications than intricate DL methods, warranting further research and development.