Enhanced Classification of Phonocardiograms Using Modified Deep Learning

Cardiovascular diseases (CVD) are the foremost cause of death globally, highlighting the importance of effective diagnostic techniques. Phonocardiograms (PCG), known for their affordability and simplicity, are pivotal in assessing heart anomalies and identifying CVDs. Cardiac auscultation, while commonly employed for cardiac assessment, heavily relies on the clinician's expertise, leading to a growing need for automated and objective cardiac sound analysis methods. This research focuses on developing an automated PCG classification system. Since the data is imbalanced, first, the data set was balanced using the random oversampling method and then the data audio augmentation method for the publicly accessible PhysioNet/CinC 2016 Challenge dataset. Instead of handicraft features, we converted the speech files into spectrograms and then fed them to the Convolutional neural network (CNN) model as images. The innovative approach involves a modified CNN integrated with dual classifiers: a SoftMax classifier and a Support Vector Machine (SVM), The proposed model demonstrates remarkable proficiency, achieving 97.85% accuracy with the SoftMax classifier and 98.28% accuracy with the SVM, surpassing the former. This model not only outperforms existing methods in PCG signal classification but also enhances computational efficiency and accuracy.