Convolutional Recurrent Neural Networks for Heart Murmur Detection

Lampros Kokkalas, Nicolas Tatlas, Stelios Potirakis


Introduction: This study proposes a method for Heart Murmur Detection from Phonocardiogram Recordings using Convolutional Recurrent Neural Networks. The network architecture was created by the team Apnea-ai for the George B. Moody PhysioNet Challenge 2022. Methods: Phonocardiogram recordings were zero-padded and divided into non-overlapping frames of 5 s width. In order to determine the presence or absence of murmurs, a combination of three one-dimensional convolution layers and one Stateful Bidirectional Gate Recurrent Unit layer was used. Convolutional layers extracted local time features, while the following Recurrent layer was used for the overall time modelling. Stateful mode helped us process recordings of arbitrary length and generate a single label for each recording.
Results: Using the Challenge scoring metric, we achieved a 5-fold cross-validation score of 887, while our method exhibited an accuracy of 0.868 and a F-measure of 0.606. The method was further validated on the Challenge hidden validation data and exhibited a score of 904. Conclusion: The proposed method has the potential to be used for pre-screening and determine the presence or absence of murmurs in a patient’s recordings.