DETECTION OF MYOCARDIAL INFARCTION USING DEEP LEARNING TECHNIQUES

Abstract

Myocardial Infarction (MI) is a serious heart condition that occurs when blood flow to a portion of the heart is blocked long enough to cause damage or potentially fatal injury to the heart muscle. It demands prompt and precise detection for effective treatment; hence, early and accurate diagnosis of MI is essential for initiating timely intervention, which can significantly reduce the risk of heart failure. Artificial Intelligence (AI) advancements have had a notable impact on MI diagnosis. One is Deep Learning (DL), a subset of Machine Learning (ML) that focuses on using artificial neural networks with multiple layers to model and understand complex patterns and representations in data. DL algorithms are designed to automatically learn hierarchical features from raw data, such as images, audio, or text, through successive layers of abstraction without requiring manual or hand-crafted feature selection. DL algorithms have recently demonstrated significant potential in improving the accuracy of MI diagnosis using Electrocardiogram (ECG) and Echocardiography (ECHO). This thesis comprises three studies, all of which involve the detection of MI. The first study is a literature review that provides a comprehensive overview of recent innovative research on DL algorithms for analyzing ECG and ECHO in identifying MI. Relevant studies and DL models were examined primarily regarding preprocessing techniques, model architectures, performance metrics, and datasets. The results indicated that DL-based algorithms significantly enhance MI's overall diagnostic performance, particularly in accuracy, specificity, and sensitivity. These improvements can reduce the risk of complications, which enables more reliable diagnoses. In the second part, a novel method is proposed for detecting MI by leveraging the strengths of DL to utilize features from ECG signals and ECHO recordings, which is one of the pioneering efforts in this direction. The model is a two-branch Convolutional Neural Network (CNN), which was trained using the HMC-QU dataset with ECHO and ECG images for accurate MI detection. The proposed method achieved a sensitivity of 100%, specificity of 99.38%, an accuracy of 98.3%, and an F1 score of 99.77%, outperforming all other state-of-the-art methods. The results proved the proposed method's effectiveness for accurately detecting MI from ECHO and ECG signals. For the final part, clinicians and healthcare providers have raised speculative concerns about medical detectors' reliability, transparency, and trustworthiness. To facilitate its integration in real-world scenarios, Explainable AI (XAI) emerged as a promising methodology aiming to provide insights into the decisions made by this "black box"; i.e. DL. To provide such insight, a custom Deep ResNet model is created to demonstrate two XAI tools, namely Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), using one-dimensional ECG data. The model performed well during training and testing, with a specificity of 91%, recall of 85%, precision of 85%, F1-score of 86%, and accuracy of 87%. On the same tangent, the visualizations provided by SHAP and LIME show high interpretability and lucidity.