HsNet– Enhancing Cardiovascular Health Detection through Advanced Signal Processing and Machine Learning

Phonocardiogram (PCG) signals offer a crucial indicator of cardiac health. Our research aims to accurately classify heart sounds for robust cardiovascular health assessment. Diverse time-frequency domain features, encompassing Mel Frequency Cepstral Coefficients (MFCCs), discrete wavelet transforms (DWT), short-time Fourier transforms (STFT), spectrograms, wavelet decomposition, homomorphic filtering, Hilbert transformations, and power spectral density, are extracted through signal processing, enhancing feature representation. The study initiates with a dataset of 1000 pure, noiseless Phonocardiogram (PCG) signals, comprising four abnormal and one normal class, obtained via digital recording with an electronic stethoscope. Real-world PCG signals often encounter noise, presenting challenges in feature extraction and classification, particularly in heavily noisy scenarios. To address this, we introduce 20dB and 10dB additive white Gaussian noise (AWGN) to simulate noisy conditions. Emphasis is then placed on extracting essential features crucial for accurate classification, providing support for the proposed ensemble-based machine learning model. The performance assessment of diverse features in the classification task is conducted on the proposed HsNet model within this paper. A justification is presented for the selection of optimal features in the context of PCG classification on the proposed model. At the core of our investigation is the HsNet model, showcasing superior performance compared to other models with an impressive accuracy of 99% for noise-free signals and 97% accuracy for signals in the presence of noise when utilizing spectrogram features as input. This research presents a promising approach for enhancing cardiovascular health assessment, offering potential benefits for early diagnosis and patient care, even in noisy real-world scenarios.