Enhancing Cardiac Health Diagnoses Through Machine Learning Analysis of Phonocardiograms (PCG)

Abstract

Phonocardiograms (PCG) provide a non-invasive approach to analyzing heart sounds, making them vital for the early detection of cardiac issues. However, identifying the most effective machine learning models and feature extraction techniques for classifying PCG signals remains a challenge. This study aims to determine the most efficient and accurate combinations of machine learning models and feature engineering techniques for classifying PCG signals, with the overarching goal of enhancing diagnostic capabilities in heart health. Seven machine learning algorithms—Logistic Regression, Decision Tree, Random Forest, Naive Bayes, AdaBoost, XGBoost, and Support Vector Machine (SVM)—were evaluated. Feature extraction methods such as Mel-frequency cepstral coefficients (MFCC), Linear Predictive Coding (LPC), and Short-Time Fourier Transform (STFT) were applied. Model performance was assessed using metrics including accuracy, precision, recall, and F1-score. The study found that advanced models like XGBoost and Random Forest, particularly when combined with STFT and MFCC features, consistently outperformed others. These models demonstrated superior accuracy and F1-scores, although they also introduced higher computational complexity. The results suggest that sophisticated model-feature combinations, particularly involving XGBoost and Random Forest with STFT and MFCC, hold promise for improving cardiac diagnostics.