Deep Learning-Based Identification of Echocardiographic Abnormalities From Electrocardiograms.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Early diagnosis of heart failure is crucial.
• Current deep learning models for electrocardiogram (ECG) analysis predict limited echocardiographic findings.
• A comprehensive approach is needed for improved cardiac assessment.

Purpose of the Study:

• To develop a deep learning model for comprehensive prediction of echocardiographic findings from ECGs.
• To integrate various cardiac abnormality predictions into a single model.
• To enhance early detection of cardiac conditions.

Main Methods:

• Utilized a large dataset of 229,439 paired ECG and echocardiography sets from 8 centers.
• Developed convolutional neural networks (CNNs) to predict 12 key echocardiographic findings.
• Employed logistic regression for a composite label indicating any positive finding.

Main Results:

• The composite findings label achieved an area under the receiver-operating characteristic curve (AUC) of 0.80 in hold-out validation and 0.78 in external validation.
• Logistic regression analysis for the composite label yielded an AUC of 0.80, with 73.8% accuracy, 81.1% sensitivity, and 60.7% specificity.
• The CNN models successfully predicted a wide spectrum of cardiac abnormalities from ECGs.

Conclusions:

• Developed advanced CNN models capable of predicting diverse echocardiographic findings from ECGs.
• Created a logistic regression model for a composite label, facilitating a broader assessment of cardiac health.
• This AI-driven approach shows promise as an adjunct tool for early diagnosis and treatment of undiagnosed cardiac diseases.