Classification of Normal and Abnormal Heart Sounds Using Deep Learning on Phonocardiogram Signals

Cardiovascular diseases (CVDs) are causing the world a significant rate of death thereby requiring early and precise detection techniques. Phonocardiogram (PCG) signals are non-invasive and affordable method of diagnostic method, nevertheless, manual interpretation is subjective and requires much clinical knowledge. In this paper, the researcher suggests a deep learning-based framework to classify the heart sounds as normal or abnormal based on the PhysioNet/CinC 2016 dataset. An extensive pipeline of bandpass filtering, normalization, segmentation and feature extraction (MFCCs and mel spectrograms) is used. The advantages of SMOTE are used to balance data, whereas data augmentation is used to enhance robustness. Various architectures, such as CNN, LSTM, CNNLSTM, EfficientNet-LSTM, and EfficientNet-CNN are tested. Among them, EfficientNet-CNN model has the highest accuracy of 95%, and EfficientNet-LSTM is also a competent model, notwithstanding the importance of spatial and temporal model combination. The outcome indicates that real-time, automated diagnosis of the heart sounds are possible in health care settings that have limited resources.