Leveraging ECG images for predicting ejection fraction using machine learning algorithms

INTRODUCTION: The capability to accurately predict the ejection fraction (EF) from an electrocardiogram (ECG) holds significant and valuable clinical implications. Various algorithms based on ECG images are currently being evaluated, with most methods requiring raw signal data from ECG devices. In this study, our objective was to train and validate a neural network on a readily available ECG trace image graph to determine the presence or absence of left ventricular dysfunction (LVD). METHODS: 12-lead ECG trace images paired with their echocardiogram reports performed on the same day were selected. A DenseNet121 model, using ECG images as input, was trained to identify EF <50 %. and then externally validated. RESULTS: 1,19,281 ECG-echocardiogram pairs were used for model development. The model demonstrated comparable performance in both the internal test data and external validation data. The area under receiver operating characteristic and precision-recall curves were 0.92 and 0.78, respectively, for the internal test data and 0.88 and 0.74, respectively, for the external validation data. The model accurately identified more than 85 % of cases with EF <50 % in both datasets. CONCLUSIONS: Actual images of ECGs with simple pre-processing and model architecture can be used as a reliable tool to screen for LVD. The use of images expands the reach of these algorithms to geographies with resource and technological limitations.