Enhancing Chest Radiography Artificial Intelligence Model with Electrocardiogram for Improved Detection of Low...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Low ejection fraction (EF) detection traditionally relies on transthoracic echocardiograms (TTE).
• Artificial intelligence (AI) models show potential for EF detection using chest radiography (CXR).
• The efficacy of combining CXR and electrocardiogram (ECG) AI models for EF detection is not well-established.

Purpose of the Study:

• To evaluate the performance of individual AI models (CXR, ECG) and combined fusion strategies for detecting low EF.
• To compare early and late fusion methods for integrating CXR and ECG AI models.
• To determine if AI model fusion improves upon CXR-alone detection of low EF.

Main Methods:

• Utilized data from 7,246 patients who underwent CXR, ECG, and TTE.
• Employed two distinct AI models for ECG analysis: a convolutional neural network and a masked autoencoder.
• Implemented early fusion (simultaneous training) and late fusion (ensemble techniques) strategies to combine CXR and ECG models.

Main Results:

• The CXR-only AI model achieved an area under the curve (AUC) of 0.798.
• Both early and late fusion models significantly outperformed the CXR-only model.
• The early fusion model achieved an AUC of 0.937 (P=0.015), and the late fusion model achieved an AUC of 0.928 (P=0.010).

Conclusions:

• Combining AI models for CXR and ECG significantly enhances the detection performance for low EF.
• This integrated AI approach offers a more accurate method for identifying patients with reduced ejection fraction.
• The findings suggest a potential non-invasive screening tool for heart conditions, reducing reliance on TTE.