Cardiac parameter estimation from electrocardiograms based on inversion of physical simulator behavior.

Electrocardiograms (ECGs) are used in clinical medicine because their waveforms provide rich information reflecting heart states. It is believed that a specialist with extensive training can identify representative abnormalities and their predictive signs through template matching with typical cases. Furthermore, it is increasingly required to identify the cause of the abnormality, which is difficult even for specialists. To provide clues to internal heart states, we propose a deep neural network-based model composed of one encoder network and multiple decoder networks for converting a 12-lead ECG into multiple microscopic physical and chemical cardiac parameters that represent internal heart states and are potentially essential factors of ECG signals, such as multiple ionic parameters related to cardiac electrophysiology. Data pairs consisting of cardiac parameters and ECGs are needed for training the model, but these are unavailable in principle. Therefore, we propose to generate such data by using a white-box model that can simulate the physical behavior of the heart. The proposed method was experimentally evaluated in terms of the mean absolute error of cardiac parameters taking continuous values and in terms of the accuracy of those taking discrete values. The results showed that the method could accurately estimate these parameters. Moreover, we investigated the estimation errors in detail by visualizing the distributions of the estimated cardiac parameters around the ground-truth values. Our method can help doctors monitor heart conditions by automatically estimating cardiac parameters from ECGs. In addition, it can be used to discover relationships among cardiac parameters, ECG waveforms, and heart diseases. However, before our method can be used in actual clinical practice, it remains to be verified that it works correctly with real ECGs.