Classification of cardiac conditions from ecg signals using hybrid feature extraction and deep learning methods

Cardiovascular diseases are among the leading causes of death worldwide, highlighting the critical importance of early diagnosis and accurate detection. This study focuses on the classification of various cardiac conditions, including arrhythmias, congestive heart failure (CHF), and normal sinus rhythm (NSR). Using two distinct datasets containing signals of various arrhythmias, CHF, and NSR, three separate ECG classification applications were performed. To achieve this, innovative signal processing methods were combined with deep learning models, and OD-1D-LBP, CI-1D-LBP, and Continuous Wavelet Transform (CWT)-based Motif Transformation (MT) techniques were applied to analyze the local and global features of ECG signals. In the first application, OD-1D-LBP was used to classify CHF, atrial fibrillation (AF), and NSR signals. The features extracted using this method were evaluated with LSTM and 1D-CNN models, with the highest accuracy of 98.97% achieved using the LSTM model. The 1D-CNN model also demonstrated competitive performance with an accuracy of 98.86%. The success of OD-1D-LBP in highlighting local features of signals has proven it to be an effective tool for distinguishing cardiac conditions. In the second application, CI-1D-LBP was employed to classify four types of arrhythmias: ventricular beats, supraventricular beats, fusion beats, and unidentified arrhythmias. Features extracted through this method were analyzed using LSTM, GRU, and 1D-CNN models, with the GRU model achieving the highest accuracy of 98.59%. The LSTM and 1D-CNN models followed with accuracies of 98.02% and 97.17%, respectively. The ability of CI-1D-LBP to differentiate between various arrhythmia types demonstrates its effectiveness as a robust analytical framework. In the third application, MT and CWT techniques were combined to analyze CHF, AF, and NSR signals in the time-frequency domain. The resulting scalogram images were classified using DenseNet models. The DenseNet169 model achieved the highest accuracy of 99.31%, while DenseNet121 and DenseNet201 models yielded accuracies of 98.30% and 98.97%, respectively. The integrated use of Motif Transformation and CWT has proven to be an effective approach for time-frequency analysis of cardiac signals. The results demonstrate that the methods used in this thesis outperform those reported in the literature and provide significant advancements in the diagnosis of cardiac conditions. These methods are projected to enhance accuracy and reliability when integrated into clinical applications for the diagnosis of cardiovascular diseases. Furthermore, the comparison of different models and datasets shows that these methods provide a generalizable and reliable framework not only for a specific signal type but also for the analysis of various cardiac conditions. The high accuracy achieved through the extraction of distinctive features and the application of deep learning models highlights the potential of these techniques to set new standards for cardiac signal analysis. This study suggests that the integration of these methods into clinical settings can improve the accuracy and efficiency of diagnostic processes, thereby enhancing the quality of patient care. The findings of this work not only surpass existing approaches in the literature but also provide more reliable and effective solutions for the diagnosis of cardiovascular diseases.