Heart Murmur Detection Using Wavelet Time Scattering and Support Vector Machines

Adrian Cornely and Grace Mirsky
Benedictine University


Abstract

The lack of infrastructure and specialized clinicians have led to reduced diagnoses and treatment of cardiac conditions for children in developing countries. This limited access to care has been further exacerbated as a result of the COVID-19 pandemic, which has resulted in many individuals choosing to forgo the screening that is necessary for early detection. An important tool to facilitate improved screening for cardiac abnormalities is the phonocardiogram (PCG). The PCG non-invasively records heart sounds, allowing for the observation of murmurs, which are abnormal heart sounds that may indicate underlying cardiac disease.

The objective of the 2022 George B. Moody PhysioNet Challenge is to utilize multiple PCG recordings taken from a single patient to determine the presence, absence, or unclear case of murmurs for each patient. The training set consisted of 3163 recordings obtained from 942 patients. Common auscultation locations for the PCG include the pulmonary valve (PV), aortic valve (AV), mitral valve (MV), and tricuspid valve (TV), though other recording locations (Phc) were allowed for in the dataset. To that end, we employed wavelet time scattering and support vector machines to determine the classification of each PCG recording based on the features from the signals alone. The classification for each of the recording locations as well as the demographic information (height, weight, age, sex) was then fed into a naïve Bayes classifier to determine the patient’s overall classification.

In the Unofficial Phase, our best performing entry for team Eagles received a score of 1413.752 using the Challenge evaluation metric on the validation set. In the Official Phase, we plan to further investigate additional features as input to the classifier as well as to fine-tune the hyperparameters of the model in order to increase classification accuracy.