Characterising the Discrete Wavelet Transform of an ECG Signal with Simple Parameters for use in Automated Diagnosis

The spectral distribution of energy varies between normal ECGs and those from patients post infarct or with ventricular hypertrophies. This suggests that discriminating between normal and abnormal conditions may be possible on the basis of differences in the distribution of spectral energy. We compare a reduced Discrete Wavelet Transform characterisation of an ECG QRS complex using three different wavelets. The wavelet transforms are based on dyadic scales and decompose the ECG signals into four detail levels and one approximation level with each decomposition being characterised by a mean and a standard deviation value. Our results indicate that, even after reducing the information in each level of decomposition of the wavelet transform to these two simple values, the discriminating power between normal and abnormal cases, calculated using receiver operator curve (ROC) analysis, exceeds 75%. This improves on the results obtained for scalar parameters such as QRS duration, areas and cardiac axis. INTRODUCTION Wavelet transforms (WT) offer an alternative representation of signals to the more traditional Fourier methods. In wavelet analysis and synthesis the basis functions, unlike the complex exponential in Fourier methods, are localised in both time and frequency’ making them very suitable for analysing non-stationary signals. The Continuous Wavelet Transform (CWT) characterises the spectral content of a signal by correlating the input signal with a number of scaled and shifted versions of what is known as a ‘Mother Wavelet’. The various correlation figures give an indication of the spectral spread of the signal. The Discrete Wavelet Transform (DWT), which is a discrete time version of the CWT, is implemented by a filter bank that decomposes the signal using successive low pass and high pass filtering operations [ 11. We compared three wavelets, the Haar wavelet [2 ] , a Daubechies wavelet [3] and a bi-orthogonal wavelet [4]. The three wavelets we have chosen have very different characteristics. The filters for the Haar wavelet are symmetrical and orthogonal. The Daubechies filters are maximally flat, orthogonal but not symmetrical and the bi-orthogonal wavelet filters are symmetrical but not orthogonal. We wish to determine how important these filter characteristics are when trying to extract pertinent features from an ECG signal in order to diagnose a particular cardiac pathology. We chose a four level dyadic decomposition of the ECG based on the fact that the signal is sampled at 500Hz and the resulting bandwidths coincide quite well with known spectral properties of the QRS complex [ 5 ] . At each level of decomposition the mean and standard deviation of the respective reconstructed signal was taken. Simple parameters like this are ideal inputs for a newal network discriminator.