Two-stage Detection of Murmurs from Phonocardiograms using Deep and One-class Methods

Sara Summerton1, Danny Wood1, Darcy Murphy1, Oliver Redfern2, Matt Benatan3, Matti Kaisti4, David Wong1
1University of Manchester, 2University of Oxford, 3Independent Researcher, 4University of Turku, Department of Computing


Abstract

Detection of heart murmurs from stethoscope sounds is a key clinical technique used to identify cardiac abnormalities. We describe the creation of an ensemble classifier using both deep and hand-crafted features to screen for heart murmurs and clinical abnormality from phonocardiogram recordings over multiple auscultation locations. The model was created by the team \textit{Murmur Mia!} for the George B. Moody PhysioNet Challenge 2022.

  Methods: Recordings were first filtered through a gradient boosting algorithm to detect \textit{Unknown}. We assume that these are related to poor quality recordings, and hence we use input features commonly used to assess audio quality. Two further models, a gradient boosting model and ensemble of convolutional neural networks, were trained using time-frequency features and the mel-frequency cepstral coefficients (MFCC) as inputs, respectively. The models were combined using logistic regression, with bespoke rules to convert individual recording outputs to patient predictions.

  Results: On the challenge validation set, our classifier scored 0.737 for the weighted accuracy and 11828 for clinical outcome challenge metric. This placed 28/305 and 188/305 on the challenge leaderboard, for each scoring metric, respectively.