Pulse Waveform Classification using Neural Networks with Cross-validation Techniques: Some Signal Processing Considerations

R. Folland, A. Das, R. Dutta, E.L. Hines, N.G. Stocks, and D. Morgan (UK)


Signal processing, pulse oximetry, signal classification,signal reconstruction.


Arterial oxygen saturation measured by pulse oximetry (SpO2) has long been established as a technique for monitoring critical care patients. Motion artifacts (MA) are physical disturbances of the patient that detrimentally affect the measured arterial pulse waveform. Identification and discrimination of these motion artifacts is fundamental to improving sensitivity towards genuine clinical events such as hypoxemia. In this paper we investigate different Artificial Neural Network implementations in the discrimination between normal and distorted waveforms in the Bayesian framework. We compare this against a neural network method of reconstructing the arterial pulse waveform from a resampled version. This method uses the neural network as a means of filtering out MA corrupted waveforms from normal waveforms. These two approaches provide methods for discriminating between normal and abnormal (MA corrupted) arterial pulse waveforms. We investigate the application of these techniques to identify the characteristics of normal arterial pulse waveforms in the development of an effective classification paradigm. We conclude that although the waveform classification system can successfully discriminate between 83.2% of normal and MA waveforms, the method of signal reconstruction offers an attractive means of identifying MA by attaining accuracies of 97% on average.

Important Links:

Go Back