Machine learning based identification of pathological heart sounds

Automated interpretation of heart sounds holds great promise in increasing the diagnostic accuracy and consistency of cardiac auscultation and allowing for use in remote, tele-health settings. However, existing algorithms for classification of hearts sounds have been constrained by limited idealized training sets and methodological issues with validation. As part of the 2016 PhysioNet Challenge competition, we present an algorithm for automated heart sound classification sthat uses Hilbert-envelope and wavelet features to attempt to capture the qualities of the heart sounds that physicians are trained to interpret. We perform a two-step classification of heart sounds into poor quality, normal or abnormal with sensitivity of 0.7958 and specificity of 0.7459.