Abstract:
Any disturbance in the activity of heart can cause irregular heart rhythm known
as cardiac arrhythmia. Electrocardiogram (ECG) is one of the most promising
tools for classi cation of di erent types of arrhythmia, which is necessary until it
goes fatal and causes loss of life. For ECG arrhythmia classi cation, a wide range
of signal processing techniques extracting features from time, frequency and time
frequency domains have been reported in the literature. Since, ECG is a nonstationary
signal, time frequency analysis can perform better than the conventional
time or frequency analysis methods. But, development of a multi-class arrhythmia
classi cation method, which is simple yet e ective in handling practical conditions
such as lack of enough training dataset and random selection of training and testing
dataset, is still a challenging task. ECG signals can be well modeled as self-a ned
fractal sets which vary under di erent arrhythmia. Thus local fractal dimension
(LFD) can be employed as a feature in classifying di erent ECG arrhythmia. In
the empirical mode decomposition (EMD) domain, the basic functions are directly
derived from the original signal without the knowledge of any previous value of the
signal. Therefore, the Hurst exponent (HE) required for deriving a set of LFD features
is calculated from the intrinsic mode functions (IMFs) obtained via EMD of
ECG signals. Since, for better approximation of LFD, at least three IMFs are to
be determined which is dependent on the length of the ECG signal, time-frequency
analysis in the wavelet packet decomposition (WPD) domain is performed for calculating
the HE as well as deriving a set of more e ective LFD features. Considering
the complexity and ease of implementation as an important criterion, a feature set
based on energy and entropy of only the 4th level detail WPD coe cients is found to
be simple yet the highest capable of solving a multi-class ECG arrhythmia problem.
Each of the proposed sets of feature when fed to Euclidean distance based classi er
can classify di erent arrhythmia even with reduced training dataset as well as randomly
selected training and testing dataset. Simulations are carried out to evaluate
the performance of the proposed method in terms of sensitivity, speci city and accuracy.
It is shown that the proposed method outperforms some of the state-of-the-art
methods with superior e cacy.
