The Choice of Boundaries for Scales of Biomedical Signals in the Recognition Algorithms

Abstract

The work is aimed at improving the recognition procedures of biomedical signals in supervised learning systems in the direction of increasing the fidelity level of decisions. The prerequisites of the work is the increasing data flow during medical diagnostics and deepening search of informative features of signals of patient’s state. The results of the work are applied to the field of medical diagnostic software.The work is based on using shape characteristic of a signal and scalar product as criteria of several signals’ similarity. The reasoning for the work is to provide signal shape characteristic algorithms with the methods to manipulate the standard for more tailored supervised learning and effective recognition. While simultaneously maintaining correctness of comparing incoming signals with the heavily modified standards with custom composition of features.Six modifications of recognition algorithms with the representation of signals in normalized measurement scales are proposed, in which joint and separate options for choosing the boundaries of these scales on training samples of signals of different classes are implemented. Procedures have been built for comparing signals with standards and making decisions with and without fitting standards to signals.Modifications to recognition procedures are implemented and explored in the MATLAB® environment. Their performance has been confirmed. The half-hour ECG recording from an open access database has been used for signal recognition for the normal and two types of abnormal cardiac beats.The viability of the development has been confirmed with the control samples in test example. An increase was found in the values of statistical estimates of the sensitivity, specificity, and general validity of solutions for recognizing signals when switching to normalized measurement scales in comparison with the option of using ordinary scales of measurements in describing the observed processes. The modification with variable boundaries of the scales has been found to be more effective for the used signal types while learning all three types of signals simultaneously.Additional modifications with variable standards and discarding worse features allowed further increase in correct decisions rate. As the result, the statistical values of recognition has been calculated for all six algorithms with variable calculation complexity and effectiveness.Special consideration has been paid to the consistency of signals and their standards for the cases of variable scales, boundaries, and composition during learning and recognition phases.