Кластеризація медико-біологічних сигналів в розпізнавальних системах, що навчаються з учителем

Abstract

Recognition of biomedical signals by the characteristics of their shape using supervised learning algorithms in patient diagnostic systems is considered. The classification of signals is assumed to be known and training samples are assumed to be available. A statistical assessment of the decision fidelity of the types of signals in the control samples by indicators of sensitivity, specificity, general validity is assumed. The principle and procedure for improving the basic signal recognition procedure are disclosed. A software toolkit is developed and researched to identify and implement possible reserves in increasing the fidelity of decisions from a more detailed consideration of the nature and characteristics of the probability distributions of signals in their phase space. For this purpose, clustering of training samples is additionally carried out at locations of their aggregation, separately according to signal classes. Families of cluster standards attached to these locations are formed and they are used to determine the types of incoming signals. The procedures for their processing are getting more complicated. The complexity of training and decision-making is growing. Signal recognition training becomes combined. However, their initial classification, which carries a certain medical meaning, is not violated. The emphasis in decision-making is shifted to comparing recognized signals with nearest standards in their vicinity. Distant standards lose their influence. The development and research are illustrated by examples of recognition of three types of QRS complexes (N, A, and V) in the patient’s electrocardiogram record. The effectiveness of the developed tools is tested on concrete training and control samples in comparison with the basic algorithm.