Classification of Phonocardiogram Recordings Using Vision Transformer Architecture

We participated in the George B. Moody Challenge 2022 to make a model which detects the presence or absence of murmurs from multiple heart sound recordings from multiple auscultation locations, as well as detecting the clinical outcomes from phonocardiogram (PCG) well. Our team, HCCL, developed a model with a visual approach for deriving a high-performance model. The model converts heart sound signals into spectrograms without requiring resampling or signal filtering. The result shows a weighted accuracy score of 0.69 (ranked 21th out of 40 teams) for the murmur detection classification on the hidden test data. For the clinical outcome identification task on the hidden test data, it shows a Challenge cost score of 11943 (ranked 6th out of 39 teams)