Classification of arrhythmias based on artificial intelligence from electrocardiography signals

Heart rhythms that differ from the normal sinus rhythm occurs as a result of the heart beating in a certain order and at regular intervals in electrocardiography (ECG) signals, are called arrhythmias. Irregularities in the rhythm can be an indicator of abnormal condition or even vital risks. Consequently, the detection and accurate classification of arrhythmias are critically important. The classification of arrhythmias using machine learning models trained with features obtained from ECG signals has been a subject of interest for researchers for many years. However, the application of deep learning techniques which eliminate the need for signal processing and feature extraction steps is also prominent research topic. Comparing the performance of these methods and determining the model parameters are essential for better results. In this thesis, two different rhythm classification methods are developed with the analysis of ECG signals and the performance of these methods are evaluated. The first method involves the rhythm classification with using machine learning models trained with features obtained from Heart Rate Variability (HRV) analysis. The second method classify rhythms by spectrogram and scalogram images as input to Convolutional Neural Networks (CNNs) without requiring signal preprocessing and feature extraction. The performance results of these two classification methods were compared. MIT-BIH Arrhythmia Database is used in the study. Normal Sinus Rhythm (N), Atrial Fibrillation (AFIB), Left Bundle Branch Block (LBBB), Right Bundle Branch Block (RBBB), Ventricular Bigeminy (B), Pacemaker Rhythm (P) are classified. In the classification performed with HRV-based features, firstly the noise removal process is applied to the signal. R wave peaks are detected on the filtered signals. HRV analysis is performed with R wave peaks and features are extracted. These features are then used to train machine learning models, including Decision Trees (DT), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and Ensemble Learning methods. In deep learning, CNN-based classification, spectrogram and scalogram images are generated from the denoised ECG signals. These images are then input into CNN architectures such as GoogleNet, ResNet-50, AlexNet, and SqueezeNet to assess their classification performance. It is determined how the performance of different neural networks performed when trained with scalogram and spectogram images. In addition, the rhythm classification performance performed with feature extraction and machine learning models is compared with the classification performance performed using deep learning models. In the HRV analysis, the highest accuracy value in arrhythmia classification performed with a two-minute signal length is obtained with the ensemble learning model with 87.91%. In the five-minute signal length, 90.48% accuracy is achieved with the SVM. In the classification performed using spectrogram images, the maximum accuracy is 91.90% with AlexNet. In the classification performed with scalogram images, the maximum accuracy is obtained as 92.85% with the GoogleNet. These findings demonstrate that scalogram-based CNN models achieve superior classification performance, highlighting their potential for accurately detecting and classifying arrhythmias in ECG signals. The observed findings show that these evaluated models provide high performance in classifying arrhythmias in ECG signals.