A method to screen left ventricular dysfunction through ECG based on convolutional neural network

Objective: This study aims to develop an artificial intelligence-based method to screen patients with left ventricular ejection fraction (LVEF) of 50% or lesser using electrocardiogram (ECG) data alone.

Methods: Convolutional neural network (CNN) is a class of deep neural networks, which has been widely used in medical image recognition. We collected standard 12-lead ECG and transthoracic echocardiogram (TTE) data including the LVEF value. Then, we paired the ECG and TTE data from the same individual. For multiple ECG-TTE pairs from a single individual, only the earliest data pair was included. All the ECG-TTE pairs were randomly divided into the training, validation, or testing data set in a ratio of 9:1:1 to create or evaluate the CNN model. Finally, we assessed the screening performance by overall accuracy, sensitivity, specificity, positive predictive value, and negative predictive value.

Results: We retrospectively enrolled a total of 26 786 ECG-TTE pairs and randomly divided them into training (n = 21 732), validation (n = 2 530), and testing data set (n = 2 530). In the testing set, the CNN algorithm showed an overall accuracy of 73.9%, sensitivity of 69.2%, specificity of 70.5%, positive predictive value of 70.1%, and negative predictive value of 69.9%.

Conclusion: Our results demonstrate that a well-trained CNN algorithm may be used as a low-cost and noninvasive method to identify patients with left ventricular dysfunction.