Deep learning model to detect significant aortic regurgitation using electrocardiography.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Aortic regurgitation (AR) is a prevalent heart disease, affecting 4.9% in the Framingham Heart Study.
• AR prevalence increases with age, potentially increasing the future burden of the disease.
• Effective screening methods are needed to manage the growing prevalence of AR.

Purpose of the Study:

• To develop and evaluate a deep learning-based artificial intelligence algorithm for diagnosing significant aortic regurgitation (AR) using electrocardiography (ECG).
• To assess the performance of a novel multi-input neural network model compared to existing machine learning approaches.

Main Methods:

• A dataset of 29,859 paired ECG and echocardiography records (including 412 AR cases) from 2015-2019 was utilized.
• A multi-input neural network combining a 2D-CNN for raw ECG data and an FC-DNN for ECG features was developed.
• Gradient-weighted class activation mapping (Grad-CAM) was employed to interpret the model's decision-making process.

Main Results:

• The multi-input model achieved an area under the receiver operating characteristic curve (AUC) of 0.802, significantly outperforming a 2D-CNN alone (AUC=0.734) and other machine learning models.
• Grad-CAM analysis indicated the model primarily focused on the QRS complex in ECG leads I and aVL for AR detection.
• The model demonstrated modest predictive value in detecting significant AR.

Conclusions:

• A multi-input deep learning model utilizing 12-lead ECG data can effectively screen for significant aortic regurgitation.
• The AI algorithm shows potential as a non-invasive tool for AR detection.
• Further validation and clinical implementation of this ECG-based AI model are warranted.