Machine learning classifiers for critical cardiac conditions

Abstract

Cardiac diseases are one of the leading causes of death in Canada. Current methods of diagnosing cardiac conditions require a manual and visual analysis of ECG and heart rate (RR interval) data. In this thesis, novel features and machine learning classifiers are developed for automating the detection of Congestive Heart Failure (CHF) and Atrial Fibrillation (AFIB). These classifiers can potentially trigger alarms when implemented in wearable devices. In the first experiment, quantitative analysis of easily measurable RR interval data is employed to detect the change in CHF severity. This experiment demonstrates that for a progressive disease such as CHF, 6 hour RR interval data can be used to classify CHF severity. It shows that the separating patients in the least severe class from more severe classes performs better than the separating all three severity classes. AFIB is one of the most frequently occurring critical event that occurs in patients with CHF. The second experiment uses novel features extracted from RR interval data, to detect AFIB within the first minute of its occurrence. This experiment evaluates feature sets engineered through different feature selection techniques. The sensitivity and specificity of the proposed classifier is 98% and 95% respectively. The third experiment aims to identify subjects at a high risk of experiencing AFIB in the future. This experiment develops novel features using RR intervals and ECG signal analysis. The results show that feature sets obtained from ECG signals can improve the classifier performance (92.4% sensitivity and 82.8% specificity). However, classifiers based on RR interval alone also show high sensitivity and specificity (81.4% and 76.9% respectively).