OF AN ELECTROCARDIOGRA PHIC WAVEFORM
A. K. Valuzhis, A. P. Rashimas · 1979
The principal cause of failure lies in the lack of optimization of the solution. The present study represents an attempt at a mathematical formulation of the recognition problem for ECG elements {waves, complexes and segments.) as an optimization problem, and a method of solution is proposed. The description of the traces and waveforms as parts of a whole is usually referred to as structural [6], and so we refer to the proposed algorithm as a statistical algorithm for the structural analysis of ECG waveforms. As a rule, the real ECG does not have distinct boundaries between its separate elements, making it difficult to hope for the development of relatively simple algorithms for the identification of those elements and for the delineation of their boundaries. In the present article, we propose the following procedure for the description of the ECG: Among all possible schemes for partitioning of the ECG waveform, the one is chosen which ensures the optimal, in some definite sense, element recognition. Exhaustive scanning of the partition schemes is realized directly in the recognition process. This is accomplished by the application of a dynamicprogramming computational procedure, which determines a recognition procedure whereby the results of partitioning the ECG waveform are repeatedly adjusted. The same problems as in the recognition of ECG elements have arisen and continue to arise in other recognition-iden tification problems, for example, in the recognition of flowing speech [7], cursive typewritten text [8, 9], and other graphical patterns [10] under random noise conditions. The optimality criterion in all these problems is the likelihood function. The dynamic-programm ing approach for complex pattern recognition was first used by Kovatevskii for subdivision of the image of a typewritten line into individual characters [8t and by Vintsyuk for the recognition of speech waveforms [7]. In these studies, the observed patterns are treated as the result of noise distortion of certain standard patterns. It is assumed that the standards are complex patterns composed of predetermined unit patterns according to definite rules, while the random noise is described by a known probability distribution function. Then the problem of describing the given pattern is the problem of determining the unit patterns and a sequence of rules such that the composed complex pattern is most like the given pattern. The application of the method of standard sequences for the recognition of ECG elements is obstructed by computational difficulties arising from the enormous diversity of the ECG elements. For the unit standards it would be necessary to choose P-wave standards (of which there would have to be approximately 10), ST-segment standards (13), T-wave standards (12) [11], and QRS-complex standards (55) [12]. Moreover, the majority of the standard waveforms in this case are functions of two parameters: width (duration) and amplitude, which must be varied in synthesizing the standard sequence.