Diagnosis of Arrhythmia Using Neural Networks

This dissertation presents an intelligent framework for classification of heart arrhythmias.
It is a framework of cascaded discrete wavelet transform and the Fourier transform as
preprocessing stages for the neural network. This work exploits the information about
heart activity contained in the ECG signal; the power of the wavelet and Fourier
transforms in characterizing the signal and the power learningpower of neural networks.
Firstly, the ECG signals are four-level discrete wavelet decomposed using a filter-bank
and mother wavelet 'db2'. Then all the detailed coefficients were discarded, while
retaining only the approximation coefficients at the fourth level. The retained
approximation coefficients are Fourier transformed using a 16-point FFT. The FFT is
symmetrical, therefore only the first 8-points are sufficient to characterize the spectrum.
The last 8-points resulting from theFFTare discarded during feature selection.
The 8-point feature vector is then used to train a feedforward neural network with one
hidden layer of 20-units and three outputs. The neural network is trained by using the
Scaled Conjugate Gradient Backpropagation algorithm (SCG). This was implemented in
a MATLAB environment using the MATLAB GUINeural networktoolbox..
This approach yields an accuracy of 94.66% over three arrhythmia classes, namely the
Ventricular Flutter (VFL), the Ventricular Tachycardia (VT) and the Supraventricular
Tachyarrhythmia (SVTA). We conclude that for the amount of information retained and
the number features used the performance is fairly competitive.