An LSTM-based Listener for Early Detection of Heart Disease

• Main Authors: Gemke, Philip, Spicher, Nicolai, Kacprowski, Tim
• Format: Conference Proceeding
• Language: English
• Subjects: Heart beat; Machine learning; Manuals; Pipelines; Prediction algorithms; Training; Training data
• ISSN: 2325-887X
• DOI: 10.22489/CinC.2022.151

As a contribution to the George B. Moody PhysioNet Challenge 2022 we (team listNto_urHeart) propose a phonocardiogram classifier. Based on the assumption that these recordings bear similarity to music, we borrow methods from the field of computational music analysis. In contrast to end-to-end machine learning approaches, we propose a carefully-crafted processing pipeline for automatically detecting single heartbeats in phonocardiogram recordings which are then classified by a bi-directional long short-term memory network. Our approach has the advantage of not requiring manual annotations during training, therefore alleviating the lack of annotated training data. In murmur detection, we reached a weighted accuracy of 0.68 in validation, 0.668 in test (rank: 25/40) and 0.64 \pm 0.08 during training. In predicting patient outcome, we reached 10,362 in validation, 13,866 in test (rank: 27 /39) and 11, 386\pm 2,108 during training. The results indicate that borrowing algorithms from computational music analysis could bear the potential to address challenges in phonocardiogram classification successfully.