Development of explainable deep learning models on ECG signals for detection of cardiac disorders

This thesis aims to develop explainable artificial intelligence (XAI) approaches for the automatic detection of cardiac disorders using ECG signals. The use of ECG signals in diagnosing cardiac disorders is particularly important for the detection of irregular heart rhythms, such as arrhythmia and atrial fibrillation. In this study, deep learning models are employed to analyze ECG data, and a Multimodal GradCAM (MM-GradCAM) method is developed to enhance the interpretability of these analyses. The hypothesis posits that a method capable of evaluating ECG data in both signal and image forms can improve diagnostic accuracy and reliability. The study involved conducting various experiments on large-scale ECG datasets using deep learning models. Findings indicate that the MM-GradCAM method provides transparency for model outputs by enabling interpretability in both time series and pattern recognition, with validations from expert cardiologists. This thesis contributes to increasing the adoption potential of XAI applications in healthcare by strengthening physicians' trust in AI-assisted systems. In conclusion, the proposed method provides an understandable and reliable framework for AI-based diagnostic models in healthcare applications.