| Abstract [eng] |
Heart failure is the main condition of death in the world, and this is the reason why detecting of this disease as early as possible is important. Efforts are being made to aid cardiologists in diagnosing the disease. Real electrocardiogram heart failure and normal sinus rhythm signals were processed in the Master’s final degree project. Lagrange first and second descriptors were computed, and two dimensional digital images were created and prepared to be used for classification of patients into heart failure and normal sinus rhythm classes using artificial neural networks. The two-dimensional digital images were constructed by reconstructing the attractor in three-dimensional phase space. Using different combinations of two time delays, in this respect the position of the digital image element was determined. The corresponding attractor's parameter, which was calculated using Lagrange descriptors, is an element of this matrix. This way of constructing two-dimensional digital image is novel feature extraction approach. The objective of the project is to create and evaluate a Lagrangian-based feature extraction method that uses two-dimensional digital images to detect heart failure in electrocardiogram segments. The ResNet and AlexNet artificial neural network architectures were utilized to classify patients into belonging to heart failure and normal sinus rhythm classes. Various parameters, such as the number of epochs, learning rate, batch size, and optimization functions (Adam and SGD), were tested. The best ResNet architecture model achieved an overall accuracy of 99,198 % using two-dimensional digital images of the first Lagrange descriptor. The top performing AlexNet architecture model also used two-dimensional digital images of the first Lagrange descriptor, with an overall accuracy of 99,104 %. |
