Cardiac sound classification using a hybrid approach: MFCC-based feature fusion and CNN deep features

The detection of cardiovascular diseases through the analysis of phonocardiograms (PCGs), which are digital recordings of heartbeat sounds, is crucial for early diagnosis. Conventional feature extraction methods often face challenges in distinguishing non-stationary signals like healthy and pathological PCG signals. Our research addresses these challenges by adopting a hybrid feature extraction scheme that leverages deep learning and handcrafted techniques. This approach allows for a more effective analysis and classification of PCG signals. This paper presents a novel approach to PCG signal classification, leveraging a fusion of deep learning features and handcrafted features based on mutual information measurements. High-level features are obtained through a pretrained deep network applied to time-frequency representations of PCG signals. Additionally, Mel-Frequency Cepstral Coefficients of empirical wavelet subbands serve as handcrafted features. Canonical correlation analysis is employed for feature fusion, effectively combining crucial information from both feature types. Classification is performed using support vector machines, k-nearest neighbor, and multilayer perceptron (MLP) classifiers with a fivefold cross-validation approach. Evaluation using the Physionet Challenge 2016 database demonstrates the superior performance of our proposed approach compared to existing state-of-the-art studies, showcasing its efficacy in PCG signal classification.