PO-04-175 RECONSTRUCTING TWELVE-LEAD ELECTROCARDIOGRAM SIGNALS FROM A SINGLE-LEAD SIGNAL USING GENERATIVE ADVERSARIAL NETWORKS

Abstract

Recently, many highly accurate artificial intelligence (AI)-based prediction models have been developed for 12-lead electrocardiogram (ECG) signals. Therefore, if one can faithfully reconstruct 12-lead signals from a single-lead signal, such prediction models can seamlessly be integrated into various wearable devices that measure only a single-lead ECG signal. We aim to verify the performance of AI model that constructs estimated 12-lead ECG signals from a single-lead ECG signal. To train and validate AI models, we used 36,610 12-lead standard ECG (10 seconds) signals recorded from Ewha Womans University Medical Center between 23 May 2017 and 2 September 2020. In this study, we examined whether it is possible to generate ECGs that reflect the characteristics of various cardiovascular diseases such as LBBB, RBBB, and AF. All input signals were normalized by using a bandpass filter and into [-1, 1]. To reconstruct 12-lead signals from a single-lead signal, we developed a novel conditional generative adversarial network, called EKGAN, which is based on Pix2Pix [1] and consists of two generators and one discriminator. Then, the performance of EKGAN was evaluated with respect to root mean square error (RMSE), mean absolute error (MAE) and percentage root mean squared difference (PRD). We used 36,610 and 9,154 ECGs to train and test AI models that generate other leads from a Lead I signal. Table 1 shows the performance of EKGAN and Pix2Pix for generating all lead signals except Lead I, where EKGAN outperformed Pix2Pix in terms of all measures. Figure 1 shows an example of original and generated 12-lead ECG signals by EKGAN and Pix2Pix using only the Lead I signal. Twelve-lead ECG signals can faithfully be reconstructed from a single-lead ECG signal by using our GAN-based reconstruction model, which opens up a wide range of medical applications of wearable ECG devices.