Neural Network Based Algorithm For Cardiac Cycles Classification With The Use Of Ecg Vector Representation

The heart muscle activity monitoring and possibility of timely detection of its pathologies are considered as important tasks of modern medicine since cardiovascular deceases represent the most common cause of human death. This paper is devoted to the development of neural network based algorithm for the binary classification of the ECG cardiac beats waveforms into the following two categories: background beats and the beats with considerably deviated forms. The vector transformation of the signal was used as a method of the ECG representation where synchronous values of ECG signal from different leads were interpreted as the vector coordinates. Three variants of the vector transformation were examined in order to check the robustness of the waveform classification procedure to the possible loss of some signal features: with the use of one, two and three ECG leads. Before the classification procedure, the ECG signals were preprocessed using digital filtering and spline interpolation. The feedforward neural network structure with one hidden layer was used for the waveforms classification. Some standard methods to avoid the network overfitting were applied. The results of the proposed algorithm testing showed rather good values of its accuracy estimates. As it was expected, the best results were obtained when all three synchronous ECG leads were used. Thus, the presented in this paper approach can be used as a base for the further development of a noise robust cardiac arrhythmia detection algorithm.