Automatic detection of heart rhythm disorders using deep learning models

This thesis focuses on artificial intelligence research into heartrelated abnormalities. Early diagnosis and assessment of heart diseases and their types play a crucial role in facilitating treatment, enabling individuals to lead a high-quality life. This study addresses heart rhythm disorders such as atrial fibrillation, ventricular arrhythmia, ventricular tachycardia and ventricular fibrillation. Furthermore, ECG beats provide preliminary information about various diseases, which is important in the diagnosis and treatment of heart conditions. In the modeling phase, different models such as ResNet50, EfficientNetB0, MobileNetV3Large, Deit Base, Swin Tiny, and Beit Base are used to support machine learning and deep learning research for the automated diagnosis of complications related to Myocardial Infarction (MI). According to the results obtained, Beit Base achieved the highest success in all evaluation criteria, yielding 98.39% accuracy and a 1.61% error rate. It is the model that provided the best result on the dataset with its F1 score. Furthermore, while Swin Tiny had the same accuracy value as the Beit model in terms of precision, it yielded lower accuracy results. Looking at the other models, ResNet50, MobileNetV3Large, and EficientNetB0 showed lower performance, respectively. ResNet50 stood out with an accuracy of 96.77%. MobileNetV3Large (95.16%) and EfficientNetB0 (94.09%) showed the lowest accuracy values, achieving the lowest success rates. In conclusion, different Transformer architectures have demonstrated superior success in distinguishing MI and related complications from ECG images. Among these models, Beit and Swin Tiny offer a highly accurate solution. This thesis demonstrates that deep learning-based diagnostic systems can be used as clinical decision support systems and offer rapidly implementable alternatives.