Abnormal ECG search in long-term electrocardiographic recordings from an animal model of heart failure

Abstract

Heart failure is one of the leading causes of death in the United States. Five million Americans suffer from heart failure. Advances in portable electrocardiogram (ECG) monitoring systems and large data storage space allow the ECG to be recorded continuously for long periods. Long-term monitoring could potentially lead to better diagnosis and treatment if the progression of heart failure could be followed. The challenge is to analyze the sheer mass of data. Manual analysis using the classical methods is impossible. In this dissertation, a framework for analysis of long-term ECG recording and methods for searching an abnormal ECG are presented. The data used in this research were collected from an animal model of heart failure. Chronic heart failure was gradually induced in rats by aldosterone infusion and a high Na and low Mg diet. The ECG was continuously recorded during the experimental period of 11-12 weeks through radiotelemetry. The ECG leads were placed subcutaneously in lead-II configuration. In the end, there were 80 GB of data from five animals. Besides the massive amount of data, noise and artifacts also caused problems in the analysis. The framework includes data preparation, ECG beat detection, EMG noise detection, baseline fluctuation removal, ECG template generation, feature extraction, and abnormal ECG search. The raw data was converted from its original format and stored in a database for data retrieval. The beat detection technique was improved from the original algorithm so that it was less sensitive to signal baseline jump and more sensitive to beat size variation. A method for estimating a parameter required for baseline fluctuation removal is proposed. It provides a good result on test signals. A new algorithm for EMG noise detection was developed using morphological filters and moving variance. The resulting sensitivity and specificity are 94% and 100%, respectively. A procedure for ECG template generation was proposed to capture gradual change in ECG morphology and manage the matching process if numerous ECG templates are created. RR intervals and heart rate variability parameters are extracted and plotted to display progressive changes as heart failure develops. In the abnormal ECG search, premature ventricular complexes, elevated ST segment, and split-R-wave ECG are considered. New features are extracted from ECG morphology. The Fisher linear discriminant analysis is used to classify the normal and abnormal ECG. The results provide classification rate, sensitivity, and specificity of 97.35%, 96.02%, and 98.91%, respectively.