الفهرس 
ABSTRACTOnly 14 pages are availabe for public view
 |  | from | 206 |  |  |
| | | from | 206 | | |
Abstract
Over the past two decades, the analysis of heart rate variability (HRV) has gained considerable traction, serving as a pivotal tool in studying various disease pathologies. HRV analyses encompass methodologies aimed at quantifying heart rate variations non-invasively.
Accurately analyzing an electrocardiogram (ECG) signal holds significant importance in this particular application, aiming to extract specific characteristics within the ECG signals crucial for identifying potential cardiovascular irregularities. However, this objective is notably challenging due to the frequent corruption or presence of noise in the desired ECG signals. Wavelet analysis emerges as a comprehensive solution to address these challenges. The thesis presents a comprehensive analysis aimed at automatically detecting R-peaks in single-lead digital ECG data. Utilizing wavelet transforms proves highly effective in examining such signals. This study harnesses wavelets as a method to filter and scrutinize noisy ECG signals, employing them specifically for identifying the positions of the QRS complex occurrences within the analysis period. Moreover, inter-beat intervals (IBIs) computation and visualization based on peak detection within an ECG signal have been presented for heart rate variability (HRV) analyses.
Moreover, this thesis aimed to conceive, assess, and apply an accessible HRV analysis. The presented analysis integrates four primary categories of HRV techniques: statistical and time-domain analysis, frequency-domain analysis, nonlinear analysis, and time-frequency analysis. Evaluations of the presented analysis were conducted by conducting HRV analysis on simulated data. The results obtained from simulations indicated the reliability of the presented analysis as an HRV analysis procedure. The presented analysis is a valuable resource, offering researchers an effective tool for conducting HRV analysis.
In conjunction with the notion and recognizing the significance of employing the introduced HRV analysis, a novel investigation was conducted to employ the analysis outcomes in indices for heart rate classification. Various algorithms based on artificial intelligence techniques were employed to classify a heart rate database using the results obtained from the HRV analysis. A comprehensive analysis was conducted to compare the accuracy of the AI-based classifier implemented using the ECG signal and the HRV analysis indices. This thesis delineates a methodology that integrates transfer learning and continuous wavelet analysis for classifying two distinct classes of electrocardiogram (ECG) signals. Long short-term memory (LSTM) networks, a subset of recurrent neural networks (RNNs), are adopted for their aptitude in analyzing sequential and time-series data. Noteworthy for their ability to recognize long-term dependencies across various time steps within a sequence, LSTM networks serve as a pivotal tool for such investigations. Within the LSTM network domain, the lstmLayer facilitates the examination of temporal sequences solely in the forward direction, while the bidirectional LSTM layer (bilstmLayer) allows for analysis in both forward and backward directions. For this study, a bidirectional LSTM layer is employed. Additionally, this approach harnesses pre-trained convolutional neural networks (CNNs), namely GoogLeNet and SqueezeNet. By utilizing wavelet-based time-frequency representations, scalograms of the ECG signals are generated, followed by the creation of RGB images derived from these scalograms. The fine-tuning of both deep CNNs is conducted utilizing these images. Furthermore, an exploration into the activations of various network layers is undertaken. The findings indicate that the accuracy of the BiLSTM network and BiLSTM with feature extraction, GoogLeNet, and SqueezeNet are 60.7143%, 90.6122%, 97.1602%, and 98.9858%, respectively, thereby confirming the superior performance of SqueezeNet.