An Artificial Intelligence Algorithm for Early Detection of Left Ventricular Systolic Dysfunction in Patients with...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Previous AI models for detecting left ventricular systolic dysfunction (LVSD) often used data near echocardiography or included patients with known heart disease.
• This limitation reduced the specificity of AI for screening purposes.

Purpose of the Study:

• To evaluate AI models' ability to predict future LVSD using electrocardiograms (ECGs) initially interpreted as normal.
• To assess if ECGs recorded one to two years before echocardiography could predict LVSD.

Main Methods:

• Retrospective analysis of 24,203 sinus rhythm ECGs from 11,131 patients.
• Training and testing of two convolutional neural network models (DenseNet-121 and ResNet-101) to predict LVSD (ejection fraction ≤50%).
• Survival analysis using Kaplan-Meier curves and log-rank tests.

Main Results:

• Both AI models demonstrated high accuracy in predicting LVSD (AUROCs 0.930 and 0.925).
• Patients predicted to have LVSD by AI showed a significantly higher risk of developing echocardiographic LVSD (HR 9.89).
• Predicted LVSD was associated with significantly lower 24-month survival rates.

Conclusions:

• AI-enabled ECG models can predict future LVSD from normal ECGs obtained up to two years prior.
• These findings highlight the potential of AI-ECG for early detection of subclinical LVSD.
• AI-ECG may improve risk stratification in asymptomatic individuals.