Phonocardiogram classification based on machine learning and deep learning using hybrid features extraction techniques

Cardiovascular ailments require early and accurate diagnosis. Auscultation with electronic stethoscope records phonocardiogram (PCG) signals useful for analysis. PCG signals contain information about heart function that can be analyzed with signal processing (SP) and machine learning (ML). Utilizing the PCG signal, cardiac sounds can be dichotomized into two distinct classes labelled as normal and abnormal. This study utilized a database comprising four distinct cardiac sound signal categories: one normal and three pathological, derived from multiple sources. Phonocardiographic waveforms were processed to extract features using Discrete Wavelet Transform (DWT) and the Filter Bank Coefficients (FBC). For learning and classification, four machine learning models were utilized: Support Vector Machine (SVM), Naïve Bayes (NB), K-Nearest Neighbor (KNN), and Gated Recurrent Unit (GRU). To augment results and enhance classification accuracy, FBC and DWT features were amalgamated for model training, followed by classification using SVM and GRU. Empirical evidence reveals that the combination of FBC and DWT features significantly improves classification outcomes across all evaluated models, including NB, KNN, SVM, and GRU. The approach delineated in this paper has the potential to accurately diagnose heart disorders in patients with an efficacy of up to 97.8%.