ECG CYCLES FORMS ANALYSIS BASED ON MACHINE LEARNING TECHNIQUES 

Lagirvandze Angelina Kahayevna, student, St. Petersburg State Electrotechnical University "LETI" named after V. I. Ulyanov (Lenin) (5 Professor Popov street, St. Petersburg, Russia), E-mail: lagirvandze2016@gmail.com
Kalinichenko Aleksandr Nikolaevich, doctor of technical sciences, professor, senior researcher, sub-department of bioengineering systems, St. Petersburg State Electrotechnical University "LETI" named after V. I. Ulyanov (Lenin) (5 Professor Popov street, St. Petersburg, Russia), E-mail: ank-bs@yandex.ru
Morgunova Tatiana Valer’evna, system analyst of the department of compulsory medical insurance, CJSC "SP.ARM" (21А Gakkelevskaya street, St. Petersburg, Russia), E-mail: Morgunova-TV@yandex.ru 

621.37 

Monitoring of the heart activity and possibility of timely detection of pathologies in its work is still considered as one of the main tasks of modern medicine, since diseases of the cardiovascular system are leading in prevalence among the causes of death. This article is devoted to the development of a neural network algorithm for binary classification of ECG QRS-complexes into forms related to the background rhythm ("normal") and deviating from it ("pathology"). The vectorized ECG signal derived from several synchronous leads was used for detection of informative features. Three variants of classification analysis were investigated to improve the algorithm stability to losses of any ECG signal components: using one, two and three leads. Previously, all leads were filtered from various types of noise using spline interpolation and digital filters. For direct classification, a neural network of direct propagation with one hidden layer and using methods of thinning and regularization of weights reducing the probability of retraining was constructed. The results of the algorithm showed high accuracy rates: the highest accuracy was achieved in the case of all three leads analysis. Thus, the approach to the analysis of forms of QRS-complexes presented in the article can be taken as a basis for the development of a more stable algorithm for recognizing heart rhythm disturbances. 

electrocardiogram, vectorcardiogram, QRS-complex, classification, signal filtering, signal processing, neural networks