Browsing by Author "Cibil, Çağla Sarvan"

Now showing 1 - 1 of 1

Dijital sağlık hizmetleri için dalgacık üretimi
(2024) Cibil, Çağla Sarvan; Özkurt, Nalan
Cardiovascular diseases are increasingly prevalent today, emphasizing the critical importance of early detection and treatment planning for maintaining health. Specialist analysis of electrocardiogram recordings plays a pivotal role in this process. To reduce the workload on healthcare professionals, automated and semi-automated systems have been developed to rapidly and accurately detect cardiac conditions. These technological innovations represent a significant step forward in the management of cardiovascular diseases. However, many current approaches prefer simple decision algorithms to reduce computational complexity in real-time electrocardiography (ECG) applications, and some generalized cardiac arrhythmia classification methods may not meet specific diagnostic needs. The wavelet transform (WT) is one of the most common algorithms used to extract meaningful information from nonstationary signals. Although it is an indispensable tool for analyzing signals in both time and frequency at various resolutions, the main challenge lies in selecting the suitable wavelet family for analysis. Typically, all available mother wavelets are employed in the analysis, and the best wavelet is selected heuristically or through an optimization algorithm to identify the most appropriate wavelet functions from a standard wavelet library. This thesis aims to construct the appropriate wavelet family for specific applications using wavelet theory and multi-objective genetic algorithm (MOGA). The proposed method describes a new and systematic approach that can also be utilized in computationally cost-effective classification models for portable health devices. The integration of wavelet theory into the construction algorithm ensures that the wavelets satisfy properties such as minimum phase, symmetry, and orthogonality. Additionally, the desired time-frequency content of the signals is analyzed and adapted to the designed wavelets thanks to the high-resolution decomposition ability of wavelet packet transform. Moreover, the constructed wavelets not only resemble the desired signals but also have discrete wavelet filters that can be used in fast wavelet transform calculations. The thesis proposes two modifications of the wavelet construction algorithm. The first one starts the wavelet design by creating piece-wise linear functions as genotypes in the piecewise polynomial-based wavelet construction method. In the second approach, namely, the roots of unit circle-based method, the genotypes are randomly initialized roots of the unit circle. Then, MOGA produces Pareto optimal solutions for the user according to the selected fitness functions. As a result, a new application-specific wavelet is constructed by combining the first and second-generation wavelet construction models. The performance of the wavelet construction algorithm is demonstrated with a case study on atrial fibrillation (AF) detection. Recordings from public datasets and ECG signals collected from the Ege University Cardiology department were used in the experiments. It was observed that a simple multilayer perceptron network detects AF signals with better performance than standard wavelets, thereby proving that our main objective is achieved.