Session P76.2

A New Adaptive Approach to Remove Baseline Wander from ECG Recordings using Madeline Structure

J Mateo*, J Rieta, R Cervigon, C Vayá, C Sánchez

Universidad de Castilla-La Mancha
Cuenca, Spain

Nowadays, there are some approaches on handling baseline wander of ECG, MEG and impedance cardiograms but these techniques present different limitations; nonlinear phase distortion (high pass filter), convergence’s problems (adaptive filter) and poor behaviour in the absence of prior knowledge of the physiological signal besides a lot processing time (time-varying filter, Wavelets and Wiener filters). A neural network filter can improve the base line cancellation with low computational load, as this paper shows.
The proposed system consists initially of a simple structure similar to the neural network ADALINE (ADAptive LINear Element), which is used like initial structure because its simplicity to optimize using the Rule Widrow-Hoff Delta algorithm. It has initially an input layer, one hidden layer (with three neurons in the initial stage and the possibility of increase this number) and an exit layer. When network has converged, if the operation obtained by the system is not the required one, adds a neuron in the hidden layer.
For this study, two types of signals have been used: real recordings from the PhysioNet Database, and synthetic signals (ECGSyn software). Using these sources, 550 recordings with different pathologies and different baselines have been obtained.
The performance of the neural network approach was compared with the standard filtering techniques using cross correlation coefficient (CCC), in the case of synthetic signals, and the signal to interference ratio (SIR) in the case of real signals.
Even in the worst conditions SNR=-2.9 dB and very low frequencies, the obtained results show the efficiency of the Neural filter (CCC=0.97+-0.02 SIR=18.7+-0.3) in comparison with the classical technique with the best performance (CCC=0.91+-0.03 SIR=15.4+-0.3). In addition, this system has important advantages: to reduce the processing time, to provoke low signal distortion and can be applied to a wide range of biomedical signals.

(Abstract Control Number: 206)