Abstracts Submitted: 510
Number of Users: 710
View Abstracts Submitted
Back to home Page
The non-invasive technique, Electrocardiography (ECG), is the principal and prominent tool used for diagnosis and prognosis of heart diseases by physicians and doctors. ECG signal has wide area of applications in biomedical field for determining the functioning of human heart. As ECG is an electrical signal, so it is susceptible to various kind of noises like low frequency noises (baseline wander) and high frequency noises (power line interference) during its acquisition. From the last few decades, denoising of ECG waveform and its classification into normal and abnormal subjects has been a challenging task in bio-medical research. Thus, ECG signal should be clean and clear to support accurate decisions by cardiologists. In this paper, discrete wavelet transform (DWT) is used for removing low and high frequency noises. The Daubechies mother wavelet (D6) is used for filtering as its shape is identical with ECG beat. The wavelet denoising is performed using D6, soft thresholding and Heursure threshold value. The denoised ECG is used for detecting R-peaks and for extracting temporal and morphological features. The temporal features includes P,Q,R,S and T waves peak amplitude, RR, PR, QRS complex interval and ST segment. The intervals and segments of ECG have fixed range value and therefore, if the extracted segments and intervals fall within this range, it is said to be a normal beat otherwise abnormal or unhealthy beat. The morphological features include wavelet coefficients of fourth level decomposition. The features from each ECG beat are fed to the classifiers for classifying beat into normal and abnormal. The classifiers used are support vector machine (SVM), random forest (RF), adaptive boosting (AD) and decision tree (DT). The performance of method is evaluated by parameters like sensitivity, positive predicitivity and accuracy for records of ECG waveforms obtained from MIT-BIH arrhythmia database. The classifiers based on five-fold cross validation have an average positive predictivity of 100%, an average accuracy of 93%, 99%, 95% and 98.2% and sensitivity of 93.2%, 98.9%, 95.1% and 98.5% for DT, AD, RF and SVM respectively. The classification accuracy of AD with the extracted parameters proves better for the proposed system in comparison to other classifiers used.
© Copyright 2017 All Rights Reserved