MYOCARDIAL INFARCTION DETECTION IN PHONOCARDIOGRAM SIGNAL

• Main Author: Puspasari, Ira
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/87581
• Institution: Institut Teknologi Bandung

Myocardial Infarction (MI) is a critical condition that must be identified promptly to prevent cardiac muscle cell death within a window of 80-90 minutes. Based on Electrocardiogram (ECG) findings, MI is categorized into two types, including STSegment Elevation Myocardial Infarction (STEMI) and Non ST-Segment Elevation Myocardial Infarction (NSTEMI). Currently, the diagnosis of MI relies on 12-lead ECG and Troponin testing. However, the ECG has significant limitations, including a false positive rate for STEMI cases as high as 42%. Furthermore, Troponin levels typically peak only 12-24 hours after the onset of myocardial injury. Given these challenges, this study proposes using a phonocardiogram (PCG) to diagnose MI. PCG captures the acoustic signals generated by the mechanical activity of the heart, providing real-time information that could enhance diagnostic accuracy. By adopting this approach, the diagnosis of MI can be improved, leading to timely and effective patient management. The study involved data acquisition from 72 normal subjects, 30 with STEMI and 30 with NSTEMI. Data was acquired using the 3M Littmann Cardiology IV electronic stethoscope, with each auscultation position recorded for 30 seconds. The following process involves applying a bandpass filter and segmenting each cycle using the Shannon energy envelope method. The next phase was extracting 61 features across time, frequency, time-frequency, and statistical domains. Feature elimination was conducted utilizing a minimum variance threshold and correlation analysis through Pearson Distance Correlation. Subsequently, mutual information was applied to select the most essential features and rank using the Select K-best method. Ultimately, 18 significant features were identified for the classification process. This study has developed a detection method for normal, STEMI, and NSTEMI signals using bagging techniques, specifically employing K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Random Forest (RF). The best classification results were achieved using RF, with an accuracy of 96%, precision of 95%, sensitivity of 95%, and an F1-score of 95%. Additionally, MI detection was performed based on the position of each auscultation, with feature selection conducted manually and classification implemented via RF. The classification system yielded an accuracy of 86%, precision of 84%, sensitivity of 85%, and an F1-score of 84%. Furthermore, this research has developed methods to enhance the accuracy of MI detection at each auscultation position through automatic feature selection, parameter tuning, and boosting methods, namely AdaBoost and Gradient Boosting. This detection system was tested on a new dataset combining samples from Indonesia and Japan. The highest accuracy was observed at the LUSB position, with AdaBoost achieving 94% and Gradient Boosting reaching 98%. This study demonstrates that eight PCG signal features—autocorrelation, negative turning, mean absolute difference, zero cross, interquartile range, minimum, entropy frequency, and spectral distance, exhibit significant correlations with troponin levels. Principal Component Analysis (PCA) revealed a total contribution of 87.5% across all auscultation sites and 90.6% specifically at the LUSB position, highlighting the potential of these features as supplementary indicators for assessing the severity of STEMI and NSTEMI. The application of the Mann- Whitney U test and Spearman correlation analysis further supports the relevance of these features in cardiovascular condition detection using PCG signals. This research proposes using PCG signals as a substitute for Troponin and ECG in detecting MI. The study demonstrates that analyzing specific features in PCG signals can contribute to developing a noninvasive MI detection system. However, the study still encounters misclassifications, where abnormal signals are classified as normal signals.