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

Abstract

Artificial intelligence models reportedly detect a low ejection fraction (EF) with chest radiography (CXR) examinations, which traditionally require a transthoracic echocardiogram (TTE) for evaluation. However, whether coupling a CXR model with an electrocardiogram (ECG) model improves the detection performance remains unclear. This study aimed to evaluate various models and fusion strategies for their detection performance.

This study included 7,246 patients who underwent CXR, ECG, and TTE at the University of Tokyo Hospital. Two ECG models were used-a convolutional neural network and masked autoencoder. Two methods for combining CXR and ECG models were tested, early and late fusion. In the early fusion method, the CXR and ECG models were trained simultaneously, whereas in the late fusion method, three ensemble techniques were implemented. The CXR single model achieved an area under the curve (AUC) of 0.798. Both fusion models significantly outperformed the CXR single model. The early fusion model achieved an AUC of 0.937 (P = 0.015), while the late fusion model had an AUC of 0.928 (P = 0.010).

Combining the CXR and ECG models significantly improved detection performance. This approach enables more accurate identification of patients with low EF, a condition typically requiring TTE for diagnosis.