Development of advanced artificial intelligence techniques for the detection of myocardial infarction ECG signals in clinical settings

• Main Author: Jahmunah, V.
• Other Authors: Ng Yin Kwee
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Science::Medicine::Biomedical engineering
• Online Access: https://hdl.handle.net/10356/170476
• Institution: Nanyang Technological University

Coronary artery disease occurs when plaque is accumulated in the walls of the artery. This causes the artery to narrow, reducing blood flow to the heart. Coronary artery disease is globally identified as the most predominant and lethal cardiovascular disease. Furthermore, undiagnosed coronary artery disease may progress and lead to complications such as myocardial infarction and congestive heart failure. Hence there is a compelling need for the prompt and unerring detection of coronary artery disease, myocardial infarction, and congestive heart failure using automated systems. The electrocardiogram (ECG) is the most preferred method of detecting cardiovascular diseases as it is easily available and economical compared to imaging methods. Hence, this thesis describes the development of advanced models using ECG signals for the detection of coronary artery disease, myocardial infarction, and congestive heart failure, focusing on the detection of myocardial infarction. Also, this thesis contributes to the medical field as this offers some level of explainability of the inner workings of the deep models that clinicians may relate to. The reliability of the developed deep model used in healthcare applications such as emergency diagnosis of different types of myocardial infarction contributes significantly to clinicians. The thesis has three contributing chapters, which are given below. In the first chapter, the development of convolutional neural network (CNN) and GaborCNN (with a unique Gabor layer) models for rapidly classifying coronary artery disease, myocardial infarction, congestive heart failure, and healthy ECG signals is discussed. The ECG signals which were acquired from the Physikalisch- Technische Bundesanstalt database, were fed to the two models for classification. The GaborCNN was affirmed to be the better model for the classification task due to its high overall accuracy of 98.74% and lower computational demand. This is the first study to integrate the Gabor filter into the CNN model to automatically classify normal, coronary artery disease, myocardial infarction, and congestive heart failure classes using ECG signals. Despite the surge in the development of robust models for the automated detection of cardiovascular diseases, these are often not trusted by clinicians due to the lack of explainability of models’ mechanisms. Hence, in the second chapter, the development of the CNN and DenseNet models with the application of an advanced and unique GRAD- CAM technique to both models’ output is discussed. ECG beats were extracted from the healthy and ten myocardial infarction classes using the R peak detection algorithm and fed to the developed CNN and DenseNet models. Application of the GRAD-CAM technique enabled visualization of ECG leads and portions of ECG waves that influenced the models’ predictive decisions. DenseNet was identified as a better model due to its low computational complexity and higher classification accuracy of 98.9% due to feature reusability. Lead V4 was the most activated lead in both models. The DenseNet model with the Grad-CAM technique enables clinicians to determine the type of myocardial infarction based on explainability and, thus, has the potential to boost clinicians’ confidence in using it in hospital settings. This is the first study to report features that influenced the classification decisions of deep models for multi-class classification of myocardial infarction and healthy ECGs. Current diagnostic models for cardiovascular diseases have been primarily developed using public databases and are thus unsuitable for hospital settings, where the uncertainty of models is predominant. In the third chapter, a unique Dirichlet DenseNet model was trained with pre-processed myocardial infarction ECG signals and tested with noisy myocardial infarction signals. The predictive entropy was used as an uncertainty measure to determine the misclassification of normal and myocardial infarction signals. The misclassification of signals was determined based on the computation of four uncertainty metrics; uncertainty sensitivity, specificity, accuracy and precision. The proposed method demonstrates that the developed model is reliable as it is able to convey when it is not confident in the diagnostic information its presenting, having the potential to make a significant contribution to clinicians, especially in emergencies such as urgent diagnosis of myocardial infarction. This is the first work to have explored uncertainty quantification of a deep model using multi-class myocardial infarction ECG signals. In summary, the models proposed in the three chapters have great potential to contribute significantly to healthcare in areas such as the emergency diagnosis of acute myocardial infarction.