The Taub Faculty of Computer Science Events and Talks

32% of all global deaths in the world are caused by cardiovascular diseases. The Electrocardiogram (ECG) is a non-invasive tool to measure the electrical activity of the heart, and it is the most common test performed by cardiologists to detect heart-diseases. Analyzing ECG signals manually is a hard task. Furthermore, abnormalities in the heart may occur at any time and not necessarily in the hospital. Many attempts were made to automate this task using machine learning algorithms. However, this task is challenging. There are many types of possible disorders in the heart and extreme variations exist between different patients. Furthermore, for a deep learning model to succeed in its task, a large amount of annotated data is needed. However, analyzing and labeling ECG signals manually is prone to errors and consumes expensive experts time. We study the use of generative adversarial networks (GANs) for the synthesis of ECG signals, which can then be used as additional training data to improve ECG classification. First, we present a GAN that learns to synthesize ECG heartbeats which can then be used as additional training data to improve classifier performance. Furthermore, to synthesize patient-specific ECG heartbeats, we propose a modified GAN optimized using a specialized loss function to mimic the morphology of the subject’s cardiac signal. In the second part of the thesis we take advantage of the nature of the ECG signal which describes the biological system of the heart. We study how to incorporate this knowledge into the generative process by leveraging a biological ECG simulator of the heart, defined by a system of ordinary differential equations (ODE) representing heart dynamics. Furthermore, We study how the ECG dynamics can be learned directly by a GAN that combines both physical and data considerations. We introduce an ECG-ODE-GAN framework, in which the generator learns the dynamics of a physical system in the form of an ODE. Finally, we study the dynamics of a full 12-lead ECG signal, the most commonly used ECG exam in medical facilities. ECG sensors were developed to allow for the recording of the full 12-lead ECG signal at home. However, if even a single lead is not correctly positioned on the body that lead becomes corrupted, making an automatic diagnosis on the basis of the full signal impossible. We present a methodology to reconstruct missing or noisy leads using the theory of Koopman Operators. We learn a dynamical system describing the evolution of the 12 individual signals together in time, which enables to impute missing leads by solving a novel least-squares system. We perform an empirical evaluation using 12-lead ECG signals from thousands of patients and show that we are able to reconstruct signals in such a way that enables accurate clinical diagnosis. This thesis is one of the first works to demonstrate the ability of deep learning to leverage knowledge about physiological systems to generate synthetic examples, specifically ECG signals. The algorithms presented enable full automation of ECG analysis, both in hospitals, and in home sensors.