Entropy-Based Phonocardiogram Classification Using Continuous and Synchrosqueezed Wavelet Transforms: A Systematic Comparison

Background/Objectives: Cardiac auscultation is an important method for identifying cardiovascular abnormalities, but conventional methods are limited by examiner-dependent variability and sensitivity. The automatic classification of phonocardiogram (PCG) has the potential to be applied for standardized cardiac screening, but still has to deal with signal non-stationarity and the extraction of discriminative features. This investigation develops a computational framework integrating wavelet analysis with entropy-based features for distinguishing normal and abnormal heart sounds. Methods: The study employs the PhysioNet Computing in Cardiology Challenge 2016 database. Signal processing includes resampling, zero-phase Butterworth band-pass filtering (20–800 Hz), and median absolute deviation normalization. Time–frequency representations are generated through continuous wavelet transform (CWT) and synchrosqueezed CWT using analytic Morlet wavelets. Multiple entropy measures including Shannon, Rényi, Tsallis, spectral, permutation, and sample entropies are computed globally and across four physiologically motivated frequency bands (20–80, 80–200, 200–400, 400–800 Hz). A regularized multi-layer perceptron with dropout performs classification. Evaluation employs stratified 5-fold cross-validation with recording-level partitioning to prevent data leakage. Results: The best configuration using standard CWT with 800 Hz bandwidth achieved test-set AUROC of 0.972, balanced accuracy of 0.915, and 96% sensitivity maintained at 90% specificity. Contrary to expectations, standard CWT outperformed synchrosqueezed CWT with AUROC advantage of +0.038. Also, the band-specific entropy analysis provided the largest performance contribution with +4.1% to AUROC, confirming frequency-localized pathological signatures. Conclusions: This methodology demonstrates how the conventional wavelet analysis integrated with entropy engineering achieves state-of-the-art performance. Also, it maintains computational efficiency (1.2 s extraction and classification) and interpretability, offering practical potential for point-of-care cardiac screening in resource-limited settings.