An improvement of unsupervised hybrid biomedical signal classifiers by optimal feature extraction in system preliminary layer

We try to place emphasis especially on the feature extraction stage of classification procedure, where new feature vectors obtained from a high-dimensional data space, which the best match the analysed classification task are proposed. Based on multilevel Mallat wavelet decomposition, parameters obtained directly from the wavelet component as well as feature resulting from energy and entropy analysis are tested. In classifier part of proposed hybrid systems, unsupervised learning systems with self organizing maps (SOM) and adaptive resonance networks (ART2) are verified. T-F methods and particularly wavelet analysis was chosen as feature extraction tool because of its ability to deal with non-stationary signals. It is important to take into consideration, that heart rate variability (HRV) signals, which were classified in elaborated systems are nonstationary and have important parameters included both in time and frequency domain. Proposed structures were tested using the set of clinically characterized heart rate variability (HRV) signals of 62 patients, as cases with a coronary artery disease of different level. Additionally similar control group of healthy patients was analyzed. Whole database was divided into learning and verifying set. Results showed, that the new HRV signal representation obtained in the space created by the feature vector based on Shannon entropy of Mallat component energy distribution gave the best classifier performance with ART2 neural structure used in classifier part of described hybrid system.