Optimizing Deep Learning Architecture for ST Elevation Detection in Myocardial Infarction Diagnosis

Myocardial infarction(MI) refers to the ischemic death of myocardial tissue, thick middle cardiac tissue. The most common manifestation of MI is acute coronary syndrome(ACS), which is usually diagnosed by the elevation of the ST segment of an electrocardiogram(ECG). This is called ST-elevation myocardial infarction(STEMI). STEMI causes the deadliest heart attacks and increases the risk of heart arrhythmia and heart failure. Existing medical diagnoses using deep learning techniques involve the categorization of ECG delineations into different feature maps to detect disease-causing abnormalities. This can be done by training a deep neural network on transforming separate windows of the best ECG features. The features are appointed based on their relevance to the diagnosed disease or class of diseases. Severe episodes of ischemic heart disease and STEMI contribute to many heart attack deaths. Immediate ST elevation heart attacks only have subtle ECG signs, which increases the likelihood of physician error. This paper aims to classify different heart attack-causing cardiovascular phenomena classes by applying feature extraction, transforms, and convolutional neural networks (CNNs). This method contributes to decreasing errors caused by patterns that may be unnoticeable by physicians. This paper proposes a diagnosis model that optimizes the search for the unhealthy class using a population-based optimization method called Particle Swarm Optimization (PSO). The model utilizes ECG data to extract features, assesses their reliability, stores the enhanced features, and feeds them into an optimized CNN to produce a desirable operational characteristic for the unhealthy class. The results of this work perform experiments on different propositions of the same general architecture, such as ConvNetQuake and ResNet50, as a comparative. The channel-split PSO ResNet50 has achieved the maximum results concerning an accuracy of 98.67%, a precision of 88%, an F1 score of 60%, and a relative operational characteristic (ROC) of 60%. This provides insights into enhancing cardiovascular health monitoring and diagnosis.