Analysis of Heart Sound for Biometric Identification system using Bayesian Optimized Support vector machine (SVM)

Abstract

Researchers have found that the distinctiveness of the phonocardiogram (PCG) signal can be used for biometric identification, unlike advances in heart sound analysis for screening and diagnosing different heart problems. Biometrics, functioning as a measure and statistical test, proves invaluable for representing the distinctive behavioral and intellectual characteristics of individuals. Auscultation through a stethoscope is a non-invasive, and cost-effective technique for heart sound analysis in both diagnostic and identification applications. However, the non-stationary nature of heart sounds renders exclusive spectral or time domain analysis ineffective. That's why it's important to make the transition to time-frequency analysis. This study introduces a method for heart sound analysis aimed at identifying subjects based on the uniqueness of their heart sounds. An electronic stethoscope was developed for signal acquisition, 560 heart sound recordings were obtained from 20 participants, At a uniform sampling frequency of 44.1 KHz and 16-bit depth. For machine learning based classification, data augmentation was performed using amplitude scaling techniques within a scaling range of [0.8, 1.2]. Discrete Wavelet Transform (DWT) based denoising of the raw PCG signal employed a Db10 wavelet family at 5th level of decomposition and soft thresholding method. Feature extraction utilized DWT-based and Mel-Frequency Cepstral Coefficients (MFCC) techniques, followed by the application of a feature reduction algorithm to eliminate irrelevant features. The selected features underwent training using a cubic Support Vector Machine (SVM), chosen for its superior classification accuracy among various machine learning classifiers. Bayesian optimization was applied during the training phase. System performance was evaluated on an independent dataset during the testing phase, resulting in a notable accuracy, precision, recall, and F1 score of 96.6%, 96.7%, 96.7%, and 96.6%, respectively. The proposed method demonstrates significant enhancement in representing distinct features crucial for identity recognition. As a future extension, the method will be refined using a larger dataset on heart sound to enhance its further capabilities on classification.