Deep learning the inverse problem of electrocardiography

This report proposes a neural network-based solution for the inverse problem of electrocardiography. Given the electrical potential recordings of a mesh of electrodes from a patient’s chest, we predict the underlying heart surface potential (HSP) that propagated through the torso to the chest body surface potential (BSP)). Due to the lack of real-world data for training, a dataset is synthesized from a forward model based on experimental data. Both a dense fully connected architecture and deep convolutional neural network (CNN) architecture are tested and evaluated. The CNN outperforms the other architecture and is able to accurately model the relationship between the HSP-BSP pairs. Our final neural network achieves a 2.2% RMSE against the unseen test split of the synthetic dataset. These predicted HSPs can be used to noninvasively diagnose cardiac electrical dysfunctions, a significant step forward for clinical applications.