Using spectral acoustic features to identify abnormal heart sounds

Using the Physionet challenge database we aim to determine whether a heart sound recording, a phonocardio-gram (PCG), corresponds to a "normal" or "abnormal" physiological state. Our goal is to augment the information available to a physician during auscultation of a patient's heart, ultimately assisting with clinical decision making. To that end, we first produce spectral features of the PCG, for varying windows and frequency bands. We use the resulting spectral information to identify a variety of features based on means, variance, and activity at different frequency bands. We find that much of the information corresponding to abnormalities is captured in these features, with particular good performance on murmurs. Finally, we build a discriminative model, specifically a random forest regressor, to classify new samples based on the aforementioned features. Our final performance on the challenge data received a combined score of 81%.