Title:  Morphological Variability Analysis of Physiologic Waveform for Prediction and Detection of Diseases

Committee:

Dr. Gari Clifford, BME, Chair, Advisor

Dr. Omer Inan, ECE, Co-Advisor

Dr. David Anderson, ECE

Dr. Vince Calhoun, ECE

Dr. Fatih Sarioglu, ECE

Dr. Aris Georgakakos, CEE

Abstract: The morphology of physiological signals offers important health information that is often difficult to discern through visual inspection. Microvolt T-Wave Alternans and morphological variability (MV) examine the shape of the electrocardiogram (ECG) and help to understand cardiac health on a cellular level, thus providing additional insight over more traditional health measures such as heart rate (HR), breathing rate (BR) and blood pressure (BP). However, the field has remained rather ad hoc, based on hand-crafted features. Using a model-based approach we explore the nature of these features and their sensitivity to variabilities introduced by changes in both the sampling period (HR) and observational reference frame (through breathing). HR and BR are determined as having a statistically significant confounding effect on the MV evaluated in high-resolution physiological time series data, thus an important gap is identified in previous studies that ignored the effects of HR and BR when measuring MV. We build a best-in-class open-source toolbox for exploring MV that accounts for the confounding factors of HR and BR. We demonstrate the toolbox's utility in three domains on three different signals: arterial BP in sepsis; photoplethysmogram in coarctation of the aorta; and ECG in post-traumatic stress disorder (PTSD). In each of the three case studies, incorporating features that capture MV while controlling for BR and/or HR improved disease classification performance compared to previously established methods that used features from lower resolution time series data. Using the PTSD example, we then introduce a deep learning approach that significantly improves our ability to identify the effects of PTSD on ECG morphology. In particular, we show that pre-training the algorithm on a database of over 70,000 ECGs containing a set of 25 rhythms, allowed us to boost performance from an AUC of 0.61 to 0.85. This novel approach to identifying morphology indicates that variation of additional elements of ECG morphology are important in risk stratification beyond amplitude oscillations in the repolarization period. This research indicates that future work should focus on identifying the nature and etiology of the dynamic features in the ECG that provided such a large boost in performance, since this may reveal novel underlying mechanisms of the influence of PTSD on the myocardium.