Detection of abnormal phonocardiograms through the Mel-frequency ceptrum and convolutional neural networks

Heart auscultation is the main method routinely used to diagnose cardiovascular disease. However, even when auscultation is performed by a knowledgeable and experienced physician, the error rate remains high, and accurate computational tools for detecting abnormalities in heart sounds would be of great aid in everyday clinical practice. Still, previous attempts have evidenced how this classification problem is extremely hard, possibly due to the vast heterogeneity and low signal-to-noise ratio of noninvasive heart sound recordings. We present an accurate classification strategy for diagnosing heart sounds based on 1) automatic heart phase segmentation, 2) state-of-the art filters drawn from the filed of speech synthesis (mel-frequency cepstrum rapresentation), and 3) convolutional layers followed by fully connected neuronal ensembles. We achieve an overall area under the receiver operating characteristic curve of 0.77, demonstrating the possibility of designing a machine-learning-based tool for heart sound classification which could serve as a diagnostic as well as screening tool in a variety of situations including telemedicine applications.