Myocardial Infarction Detection using Variational Mode Decomposition with Fuzzy Weight Particle Swarm Optimization and Depthwise Separable Convolutional Network.
Journal:
Computers in biology and medicine
Published Date:
May 21, 2025
Abstract
The challenge of precisely recognizing myocardial infarction (MI) from electrocardiographic (ECG) readings stems from the complex nature of these signals.ECG data exhibit both nonlinear and non-stationary properties, making interpretation difficult. In addition, the presence of noise in these signals can obscure vital cardiac information, further complicating the detection process. Conventional machine learning and deep learning techniques, including typical Convolutional Neural Networks (CNN), often struggle to detect nuanced differences that are crucial for distinguishing between normal and abnormal heartbeats. This study addresses these shortcomings by introducing an advanced framework that combines Variational Mode Decomposition with Fuzzy Weight Particle Swarm Optimization (VMD-FWPSO) for enhanced noise elimination, Principal Component Analysis (PCA) to reduce feature dimensionality, and a Depthwise Separable Convolutional Network (DwSCN) for precise classification of heartbeats. The VMD-FWPSO technique enhances signal decomposition by optimizing the selection of Intrinsic Mode Functions (IMFs) and preserving critical cardiac features while filtering out noise.PCA refines the extracted features by highlighting the essential morphological patterns, which are then processed using the DwSCN model. DwSCN is particularly adept at capturing spatial and temporal dependencies in ECG data, thereby improving the classification accuracy. An in-depth evaluation of the model's capabilities was performed using the PTB-ECG and MIT-BIH Arrhythmia datasets. The results were impressive across both the datasets. For the MIT-BIH dataset, the model exhibited exceptional performance metrics: 99.06 % accuracy, 99.01 % precision, 99 % recall, 99.05 % F1-score, and 99.05 % specificity. The PTB dataset yielded equally remarkable results, with the framework achieving 99.25 % accuracy, 99.1 % precision, 99.15 % recall, 99.2 % F1-score, and 99.3 % specificity. These outstanding performance indicators demonstrate the reliability and effectiveness of the framework, presenting a promising solution for automated MI detection that could significantly impact clinical cardiology practices.