Cardiac Arrhythmia Monitoring and Severe Event Prediction System

Abstract

Abnormalities of cardiac rhythms are correlated with significant morbidity. For example, atrial fibrillation, affecting at least 2.3 million people in the United States, is associated with an increased risk of both stroke and mortality; supra-ventricular tachycardia, detected in approximately 90,000 cases annually in the United States, ventricular arrhythmias cause 75% to 80% of the cases of sudden cardiac death; bradyarrhythmias may cause syncope, fatigue from chronotropic incompetence, or sudden death from asystole or ventricular tachycardia. Due to the time-sensitive nature of cardiac events, it is of utmost importance to ensure that medical intervention is provided in a timely manner, which could benefit greatly from a cardiac arrhythmia monitoring system that can detect and preferably also predict abnormal cardiac events. In recent years, with the development of medical monitoring devices, vast amounts of physiological signal data have been collected and become available for analysis. However, the extraction of the relevant information from physiological signals is hindered by the complexity within signal morphology, which leads to vague definitions and ambiguous guidelines, causing difficulties even for medical expert. To address the variability-related issues, most traditional methods depend heavily on pre-processing to identify specific morphology types and extract the related features. Despite many successes, one of the drawbacks of these methods is that they require signal data of high quality and tend to be less effective in the presence of noise. Although not an issue in almost noiseless situations, such pre-processing--based methods have become insufficient with the advent of portable arrhythmia monitoring devices in recent years capable of collecting physiological signals in real time, albeit with more noise. Therefore, to enable automated clinical decision, it is desirable to introduce new methods that require minimal pre-processing prior to analysis and are robust to noise. This thesis aims to develop a cardiac arrhythmia monitoring and prediction system applicable to portable arrhythmia monitoring devices. The analysis is based on a novel algorithm which does not rely on the detailed morphological information contained within each heartbeat, thus minimizing the impact of noise. Instead, the method works by analyzing the similarity and variability within strings of consecutive heartbeats, relying only on the broad morphology type of each heartbeat and utilizing the computer's ability to store and process a large number of heartbeats beyond humanly possible. The novel algorithm is based on deterministic probabilistic finite-state automata which have found great success in the field of natural language processing by studying the relationships among different words in a sentence rather than the detailed structure of the individual words. The proposed algorithm has been employed in experiments on both detection and prediction of various cardiac arrhythmia types and has achieved an AUC in the range of 0.70 to 0.95 for detection and prediction of different types of cardiac arrhythmias and cardiac events with data collected from publicly available databases, hospital bedside database and data collected from portable devices. Comparing with other well-established methods, the proposed algorithm has achieved equal or better classification results. In addition, the performance of the proposed algorithm is almost identical with or without any pre-processing on the data. The work in the thesis could be deployed as a cardiac arrhythmia monitoring and severe event prediction system which could alert patients and clinicians of an impending event, thereby enabling timely medical interventions.