Enhanced Myocardial Infarction Identification in Phonocardiogram Signals Using Segmented Feature Extraction and Transfer Learning-Based Classification

Myocardial Infarction (MI), commonly known as a heart attack, is a type of cardiovascular disease characterized by the death of heart muscle cells. This condition occurs due to the blockage of blood vessels around the heart, inhibiting blood flow and causing an insufficient oxygen supply to the body. Typically, cardiovascular disease tests involve electrocardiogram (ECG) and photoplethysmogram (PPG) signals. In recent years, researchers have explored the application of Phonocardiogram (PCG) signals for cardiovascular detection due to their non-invasive, efficient, accessible, and cost-effective nature. While deep learning has been successful in object detection in digital images, its application to PCG signals for heart attack detection is rare. This study bridges this gap by introducing an enhanced technique called the Myocardial Infarction Detection System (MIDs). In contrast to previous deep learning research, this study employs a transfer learning algorithm as a classifier for MI feature datasets. Feature extraction is performed in segments to obtain more accurate MI features. Six feature extraction methods and transfer learning models based on Convolutional Neural Networks (CNN) using the VGG-16 architecture were selected as the primary components for MI identification. Additionally, this study compares these models with other CNN transfer learning models, such as VGG-19 and Xception, to assess their performance. Two experimental scenarios were conducted to evaluate MIDs performance in MI detection: experiments without hyperparameter tuning and with hyperparameter tuning. The results indicate that MIDs with CNN (VGG-16) after tuning exhibited the highest detection performance compared to other transfer learning CNN models, both with and without tuning. The accuracy, specificity, and sensitivity of MIDS detection with this configuration were 96.7%, 96.0%, and 97.4%, respectively. This research contributes to the development of an enhanced MI detection technique based on PCG signals using a transfer learning CNN.