Detection of occluded coronary arteries in non-ST-elevation myocardial infarction (NSTEMI) patients with deep learning models and ReliefF-based weighted subspace SVM ensembles (RBWSSE) algorithm

Abstract

Background/Objectives: Heart attacks are the leading cause of sudden death worldwide. Early diagnosis of the main coronary arteries is vital to prevent sudden death. While occluded coronary arteries are easy to detect from electrocardiogram (ECG) signals in ST-elevation myocardial infarction (STEMI), they are not in Non-ST-Elevation Myocardial Infarction (NSTEMI) cases. This is because ECG abnormalities in NSTEMI cases are visually indistinguishable to physicians. This study aims to develop a deep learning-based method that can directly detect occluded main coronary arteries in NSTEMI patients from 12-lead ECG data without performing coronary angiography (CAG). Methods: In this study, 12-channel digital ECG signals of NSTEMI cases were collected by expert physicians and labelled by CAG processing. Thus, a unique dataset has been created in this field. In the proposed approach, the labelled signals are used to predict the occluded main coronary artery or arteries with deep learning-based models. A hybrid classification strategy has been developed. Results: A classification accuracy of 73% was achieved with the Residual & Attention & Long Short-Term Memory (LSTM)-Convolution Neural Network (CNN) (RAL-CNN) model developed on the seven-class data set. With the classification strategy ReliefF-Based Weighted Subspace—Support Vector Machine (SVM)—Ensembles (RBWSSE), the classification performance was significantly increased (about 8%), and 81.3% classification accuracy was achieved. Classification accuracies obtained with popular pre-trained CNN models are also given. Although the classification results are not top-level, they serve as a baseline model for future studies. With newer models to be developed in the future, the classification performance will be further improved for this unique data set.