Hybrid wavelet-kernel based classifiers and novelty detectors in biosignal processing

The recognition of waveforms represents a major challenge in biosignal processing. In this area, the recognition scheme has to offer a particular high generalization performance and should often allow for the inclusion prior knowledge about the waveforms. In this paper, we propose a hybrid machine learning scheme as general approach to waveform recognition in the biomedical area. Our hybrid scheme is based on feature extractions by adapted filter banks and support vector machines or kernel based novelty detectors. It allows for the inclusion of a priori knowledge such as local instabilities in time and shift-variance of bioelectric waveforms. We apply our scheme for the classification of endocardial waveforms and the detection of transient evoked otoacoustic emissions. For both applications, we show that our scheme outperforms conventional methods used before.