Acoustic Feature Representation Comparison for Blood Pressure Classification from Phonocardiogram Signals Using Deep Learning

Abstract

Hypertension affects over one billion people worldwide yet remains underdiagnosed because conventional cuff-based screening is episodic, operator-dependent, and unsuitable for continuous monitoring. Phonocardiogram (PCG) signals offer a promising cuffless alternative by capturing subtle cardiac mechanical changes associated with elevated blood pressure. This paper investigates the effect of acoustic feature representation on ResNet50-based binary hypertension detection using HSBP-DB $(n=400)$. Four two-dimensional representations are evaluated: Spectrogram, Scalogram, Cepstral Coefficient (CC), and MFCC + Delta, using stratified train, validation, and test splits based on the 2017 AHA/ACC blood pressure schema. CC achieves the highest accuracy (64.44%) and AUC (0.7647), indicating stronger overall discrimination. In contrast, MFCC + Delta achieves perfect Hypertension recall (1.0) and the highest F1-score (0.7445), making it more suitable for sensitivity-critical screening. The Scalogram shows the weakest generalization, with class-biased predictions and overfitting. These findings show that cepstral and mel-frequency representations outperform time-frequency energy maps for PCG-based hypertension classification.