Abstract:
The work in this thesis is a development of a QRS detector working on large amplitudes. A nonlinear whitening and optimized adaptive filter based on an RBF neural network whose activation functions are sigmoidal is used to intelligently remove noise affecting the ECG signal and at the same time to enhance the peaks corresponding to QRS complexes. The behavior of the filter is controlled with a cost function based on M-estimators and a new technique for contextual centering with which the filter completely removes the lowfrequency components and preserves the high frequency components. The filter residual signal then passes through a matched filter that improves the signal to noise ratio (SNR). The matched filter increases the detection performance. The signal from the matched filter bears other nonlinear transformations namely squaring and moving average filtering. To validate the performance of the detector, we applied the algorithm over all files in the MIT/BIH database. We obtained an error rate of 0.28%, sensitivity of 99.82% and a specificity of 99.91%. These results are comparable even better than those in the literature.
