A hybrid discrete wavelet transform (DWT)-principal component analysis (PCA) approach for discriminative feature selection in heart valve diseases.

Phonocardiography (PCG) plays a fundamental role in the diagnosis of heart valve diseases, but it has certain limitations. The signals are often affected by noise and variations, which makes their analysis more complex. Furthermore, human hearing does not always allow for the perception of all the sounds, thus increasing the risk of diagnostic errors. The non-stationary nature of cardiac signals also contributes to these difficulties. This paper presents a hybrid method for discriminating valvular diseases from PCG signals, using a dataset of 32 recordings divided into five categories: aortic stenosis (AS), mitral regurgitation (MR), mitral valve prolapse (MVP), and ejection click (EC), and normal cases (N). After denoising with the discrete wavelet transform (DWT), the features extracted from the PCG signals were processed using principal component analysis (PCA) to select the most relevant ones. The analysis of these features enabled the differentiation of several valvular heart diseases. The methodology achieved effective discrimination between pathological conditions using K-means clustering, with three principal components explaining 91% of the total variance. The energy ratio (ER), murmur duration (ΔTM), and inverse approximation signal ratio (InASR) emerged as the most discriminative features. The results demonstrated strong clustering performance, with a Silhouette Score of 0.5829, a Davies-Bouldin Index of 0.5798, a Within-Cluster Sum of Squares (WCSS) of 1.1403, and a Calinski-Harabasz Index of 54.17, achieving an overall accuracy of 93.3% in discriminating between the different valvular heart diseases, particularly aortic stenosis, mitral regurgitation, and mitral prolapse. Overall, this approach paves the way for the development of automated diagnostic tools, enhancing both the precision and speed of patient diagnosis.