Robust cardiac arrhythmia detection and classification from physiological signals

With the number of cardiovascular arrhythmia cases ever increasing, there is a dire need for accurate detection and classification methods. If the arrhythmia conditions in a patient are detected early, the chances of life threatening scenarios like that of heart failure would reduce through timely interventions and lifesaving therapies. Arrhythmia is an abnormal heart condition that can be studied through Electrocardiogram (ECG) analysis, which is a signal representation of cardiac activity. A detailed ECG analysis would allow performing a precise categorization of arrhythmia. In this research, a graphical computational tool has been developed, tested and analysed that can aid medical expertsto extract some of the clinically relevant medical information from the patient's cardiac system, analysing features from the ECG. The distorted beats which are the indicators of arrhythmia may sometimes randomly occur and could span over a short interval. The usage of an automated classification mechanism would minimize the complexity in the identification of these distorted intervals. The arrhythmia data sets used and analysed in this work are from the UCI Machine Learning Repository which contains data of 452 subjects. The relevant ECG features had been extracted in to a matrix and then six machine learning algorithms were trained individually and tested to determine their effectiveness. A comparison between these algorithms – fine tree, coarse tree, fine K-Nearest Neighbour (KNN), coarse KNN, medium KNN, and cosine KNN, was made. The observation was, among all the classifier learners for predicting normal vs abnormal heart beats from ECG, the fine tree algorithm performed the best. The fine tree had high model flexibility with a high prediction speed. It was determined to have an accuracy of 99.4% in predicting the heart beat as normal or abnormal from the ECG features and it had an average sensitivity of 0.9940