Identifying Best Feature Subset for Classification of Cardiac Arrhythmia

Abstract

Functioning of human heart can be checked by obtaining an ECG (electrocardiogram) Signal. Correct identification of presence or absence of cardiac arrhythmia is not possible unless the examiner is a cardiologist. As a result, in those times, when a cardiologist is not available, it becomes intricate to provide all necessary treatment to the suffering patient; it happens because the ECG technicians don’t have enough competence to identify the abnormalities in heart rhythm. This necessitates that a system be devised, which takes data from ECG and identifies the type of Heart Rhythm as well as suggest possible treatment for the patient. This document proposes an Arrhythmia Classification model based on statistical data obtained from the ECG signal. The proposed model takes in the ECG data and identifies the type of heart rhythm for the subject patient; it utilizes a combination of feature selection and classification algorithms to identify the best feature subset for arrhythmia classification. The classification is performed considering the training data which is used to train the classification algorithm. Training data is the annotated data containing instances for different types of arrhythmia. The proposed model evolved through total of five phases; at every phase rigorous experimentations were performed to validate the results. In phase 1 the data from the ECG signal is taken in as an input for the two classification algorithms (i.e. KNN and SVM). In phase 2 data is input to the feature selection part of the model; here feature selection is performed using PCA, and then tenfold and twentyfold cross validation is performed to identify the maximum classification accuracy attained by the model. In phase 3 the ECG data is input to the feature selection part of the model; here feature selection is performed using Improved F-Score, and then tenfold and twentyfold cross validation is performed to identify the maximum classification accuracy achieved by the model. Next, in phase 4 the filter and wrapper based approaches for feature selection are combined and introduced to achieve more accurate results for diagnosis of cardiac arrhythmia. Improved F-Score is used as a criterion in the Filter part and SFS (Sequential Forward Search) is used as criterion in the wrapper part. Results for the model are validated by performing twentyfold cross validation by using SVM and KNN as the classification models for the purpose. Finally, in phase 5, the concept proposed in phase 4 is extended to the next level by embedding averaged improved F-Score as the criterion in the filter part of the model. Eventually the accuracy of model is tested by performing twentyfold and leave one out cross validation. The performance of the model comes out to be more stable and accurate than all previous phases and also it out performs the accuracies achieved by the researchers. Conclusively, it can be inferred that progressive research as presented in this document can reduce dependence on cardiologists in the futuristic perspective.