Title:  Processing and Learning of Cardiac Signals: Application to Heart Failure

Committee:

Dr. James Rehg, CoC, Chair, Advisor

Dr. Omer Inan, ECE, Co-Advisor

Dr. Mark Davenport, ECE

Dr. David Anderson, ECE

Dr. Rishikesan Kamaleswaran, BME

Dr. Santosh Kumar, U of Memphis

Abstract: Heart Failure (HF) is a debilitating disorder contributing to nearly 300,000 deaths and more than 800,000 hospitalizations each year in the US. The cost associated with HF exceeds $30 billion per year and expected to reach $70 billion by 2030. One of the driving factors of the mortality and cost of HF is the high rate of readmission after patients’ initial hospitalization. Rehospitalizations can be reduced by remote monitoring of HF patients. Large-scale remote monitoring is possible via non-invasive and unobtrusive devices that can measure cardiac function. Two cardiac signals acquired from such devices are called Ballistocardiography (BCG) and Seismocardiography (SCG). BCG and SCG measure the whole-body reaction forces and local chest vibrations in response to cardiac ejection of the blood from the heart, respectively. The goal of this thesis is to develop effective processing and learning methods for cardiac signals, such as BCG and SCG, with an application to HF care. This work describes two processing and modeling approach for BCG and SCG signals in classification of the HF patients’ clinical status. In the first approach, processing pipeline for BCG signals is developed to classify clinical decompensation. Area under the Receiver Operating Characteristic Curve (AUC) of 0.78 is achieved in BCG data collected from HF patients from home and hospital settings. In the second effort, processing pipeline is developed for SCG signals to classify hemodynamic decompensation. AUC of 0.84 is achieved from SCG data collected in the hospital. The final piece of work aims to handle the scarce labeled data in cardiac signal domain. Recently, Self-supervised Learning (SSL) methods in computer vision and natural language processing saw great successes with limited amount of labeled data. Motivated by these successes, we propose a novel SSL task where a model is learned through matching the representations of simultaneously recorded cardiac signals. We show that the novel SSL approach shows an improvement of 5% against the state-of-the-art SSL method in scarce data settings. Furthermore, the approach has comparable performance in a variety of settings with much lower data requirement (19 vs. 8800 subjects).