<p>Phonocardiogram (PCG) signals contain valuable spectral features that can be levenraged to analyze heart sounds effectively. This study investigates the discriminative power of low-dimensional Spectral Basis Vectors (SBVs) extracted from PCG spectrograms using deep Orthogonal Non-negative Matrix Factorization (deep ONMF) to distinguish normal heart conditions from pathological cases.Unlike shallow ONMF mezthods, deep ONMF leverages a multi-layer structure to extract hierarchical spectral features, enhancing feature representation and discrimination. The proposed methodology consists of data preprocessing, time–frequency (TF) matrix transformation, deep ONMF-based SBVs extraction, feature integration, and performance evaluation. The effectiveness of the extracted features was assessed using statistical metrics, including euclidean distance, P-value, and the Kruskal–Wallis test, demonstrating significant separability and robustness, even in noisy conditions. Comparative analysis with conventional feature extraction methods—Discrete Wavelet Transform (DWT), Mel-Frequency Cepstral Coefficients (MFCC), and Convolutional Neural Networks (CNN)—was conducted on the Yaseen and PhysioNet databases. On the Yaseen database, the proposed method achieved a highly significant P-value of 1.18 × 10⁻⁹³ with a 95% Confidence Interval (CI) of 91.9% to 95.1%. Similarly, on the PhysioNet database, the method attained a P-value of 1.28 × 10⁻¹⁰⁰ with a CI range of 95.2% to 97.8%. These results underscore the superior discriminative capability of deep ONMF in feature extraction, establishing it as a reliable and effective approach for HVD diagnosis.</p>