Electrocardiogram signal based sudden cardiac arrest prediction using machine learning approaches

This thesis focuses on predicting occurrence of imminent sudden cardiac arrest (SCA) using heart rate variability (HRV) and electrocardiogram (ECG) signals. Sudden cardiac death (SCD) is a devastating cardiovascular disease that responsible for millions of deaths per year. SCD occurs when SCA went...

Full description

Saved in:
Bibliographic Details
Main Author: L Murukesan, Loganathan
Format: Thesis
Language:English
Subjects:
Online Access:http://dspace.unimap.edu.my:80/xmlui/bitstream/123456789/61540/1/Page%201-24.pdf
http://dspace.unimap.edu.my:80/xmlui/bitstream/123456789/61540/2/Full%20text.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:This thesis focuses on predicting occurrence of imminent sudden cardiac arrest (SCA) using heart rate variability (HRV) and electrocardiogram (ECG) signals. Sudden cardiac death (SCD) is a devastating cardiovascular disease that responsible for millions of deaths per year. SCD occurs when SCA went untreated for more than 10 minutes. Hence, predicting imminent SCA before its occurrence or identification of high-risk patients for SCD can save millions of lives. Two international databases, namely MIT/BIH Sudden Cardiac Death database (20 subjects) and MIT/BIH Normal Sinus Rhythm database (18 subjects) were used in this work. Both databases have two leads ECG recording of patients in supine condition. In addition, HRV signals are provided in these databases. Two segments of HRV signals were used in this work. First segment is five minutes long and it was segmented two minutes before the onset of ventricular fibrillation (VF). Consequently, second segment is one minute long and it was segmented five minutes before the onset of VF. As for normal subjects, these segmentations were done at random intervals. Besides, these segmentations were done to achieve two and five minute prediction of imminent SCA, respectively. Both HRV signal segments were pre-processed to remove and interpolate ectopic beats. Then, time and non-linear domain features were extracted. Next, HRV signals were detrended and frequency domain features were extracted. Feature selection method is different for each time segment. For features of five minutes HRV signal, sequential forward selection (SFS) was used to select optimal features while in one minute HRV analysis, feature selection using principal component analysis (PCA) and correlation based feature selection (CFS) were experimented in addition to SFS. Optimal features selected using each methods were analyzed for its statistical significance using analysis of variance (ANOVA) test. Based on literature, four machine learning classifiers (support vector machine (SVM), probabilistic neural network (PNN), K-nearest neighbour (KNN) and classification tree (CTree)) were used for prediction in both analyses. In contrast, one minute ECG, which is five minutes before the onset of VF, was extracted from the database. Then, it was pre-processed to eliminate power line interference and high frequency noises. S-Transform (ST) based novel noise removal method was used for removing zero energy noises. Then, segment from R wave until the end of T wave (RT ABSTRACT end) was extracted from each ECG trace. Two groups of features (G1 and G2) were extracted from this novel ECG segment. G1 consists of four non-linear features (Hurst exponent, largest Lyapunov exponent, approximate entropy and sample entropy) while G2 consists of four higher order statistic features (mean, variance, skewness and kurtosis) and proposed angle of elevation/depression (AED) feature. The proposed AED feature is statistically significant (ANOVA) with p < 0.05. In this analysis, three classifiers (SVM, subtractive fuzzy clustering (SFC) and neuro-fuzzy classifier (NFC)) were used for SCA prediction. Through these analyses, maximum prediction accuracy of 97.37% was achieved in both two and five minutes SCA prediction using HRV signals. In addition, 100% prediction accuracy was produced in one-minute ECG analysis. The proposed AED feature produced 86.84% prediction accuracy.