In this paper we analyze algorithms for filtering white gaussian noise and 50 Hz power line noise p o m ECG signals. We used several wavelets to study their efect and efficiency in the filtering process. To deal with these dferent kinds of noises we used two distinct sopthresholding techniques: the Donoho 's statistical threshold estimator and a method developed by us. This last method exploits one of the wavelet processing main features: time-Pequency relation.The de-noising methods led to good filtering results: noise reduction with only minor change of the ECG waveforms as confirmed by the high correlation values between the processed signal and the original one. These good results are due to a wavelet advantage over classical filtering -time-frequency relation -enabling the possibility offiltering noise in the same frequency band of the ECG signal with minimal interference. IntroductionECG noise removal has always been a subject of great study. The purpose of the de-noising filtering process is to reduce noise level in the signal and, simultaneously, prevent waveform distortion. This last characteristic is of vital importance in order that wrong diagnosis or analysis of the ECG signal will not be made.The ECG signals used in our study are from a European Database (CSE, [ 5 ] ) and exhibit a broad spectrum of QRS morphologies. These signals are contaminated by random noises uncorrelated with the ECG signals (myoelectric, thermal, etc.), that can be approximated by a white noise source, as well as the periodic 50 Hz power line noise.In order to remove these two types of noise, using wavelets, we had to select those ones similar to specific ECG waveforms. These included wavelets developed by Ingrid Daubechies [7], Coiflets [7] (compactly supported wavelets) and other with bi-orthogonal properties [7]. This served the purpose of analyzing the effects of de-noising when applying different wavelets to different ECG 0276-6547/99 $10.00 0 1999 IEEE waveform morphologies. After the wavelets selection, the noise removal process involves a soft-thresholding operation on the coefficients of the ECG wavelet decomposition.Details concerning this soft-thresholding operation and the performance of this wavelet de-noising approach are presented and discussed next. Wavelet de-noisingWavelet signal decomposition can be seen has an iterative process whereby a signal is decomposed into finer resolution signals in time and frequency. First of all, two symmetric filters are created fiom a "mother" wavelet and a scaling function associated to that wavelet. These filters will provide an orthogonal basis dividing the signal fiequency spectrum and generating high and low frequency signals in each iterative step. These signals are decimated by two before the next iterative step. Details on wavelet decomposition can be found elsewhere [ 1,7]. Figure 1 illustrate this wavelet decomposition tree of an ECG signal, where the A bpproximations) boxes represent the low frequency components obtained by the low pass filter (LPF), and the D Uetail...

No abstract

No abstract

FHR rFetal Heart Rate'? signals analysis is an important diagnostic tool in the assessment of the fetus well being. One of the most important FHR features is its baseline. Visual evaluation of FHR baseline reveals however a large inter and intraobserver variability.In this paper a new FHR base line determination method using artificial neural networks (ANN) is presented.Two base line determination methods with multilqer perceptron ANNs (namely base line estimation and base line classijkation) are described and compared based on their practical application results.

A computer program for ECG analysis and interpretation developed at the University of Porto, Portugal is presented. The program runs on a microcomputer and employs the three-lead Frank VCG. The signals are sampled at 250 Hz with 8-bit precision during 5.5 s. Details on signal conditioning and wave recognition and measurement techniques are given. The diagnostic part of the program uses decision-tree logic. The decision rules are mainly derived from the Washington Code. The diagnostic accuracy of four classes (normal, left and right ventricular hypertrophies and myocardial infarction), evaluated in a sample of 1,075 pediatric and adult patients and classified by ECG-independent means was 76%. The program is currently being evaluated on the CSE database.