HM-Detect: A Novel Method for Heart Murmur Detection and Classification Using Machine Learning and Sound Feature Analysis

Phonocardiography (PCG) is a non-invasive assessment for cardiac monitoring using heart sound signals. Heart murmurs are abnormal sound signals generated by turbulent blood flow and are closely associated with specific heart disorders. Current methods to detect and qualify a murmur do not recognize that an important part of diagnosing a patient depends upon classification of the heart murmur based on the location it is coming from. This research presents HM-Detect, a new machine learning-based evaluation of heart sounds for murmurs. The proposed method analyzes heart sound features, namely the filter bank energies and spectral sub-band centroids. Furthermore, HM-Detect addresses two key challenges that come with heart sound data: (a) inherent temporal dependence and (b) inaccuracy due to imbalanced classes in PCG databases. To combat these, a synthetic sampling method on heart sound data was performed. A new long short-term neural network (LSTM) was created to incorporate the temporal aspect of heart sounds. These allow for an encapsulation of the spatiotemporal profile of heart sounds and results in accurate detection and classification of heart murmurs. Experiments were done on the CirCor DigiScope Phonocardiogram dataset, which is a clinically verified dataset with thousands of heart sounds recordings from patients. Results show the proposed method outperforms the latest published methods with a performance accuracy of 87%.