Learning Time-Frequency Representations of Phonocardiogram for Murmur Detection

As part of the George B. Moody PhysioNet Challenge 2022, our team (amc-sh) developed new computational approaches to diagnose cardiac abnormality or heart murmur. The proposed deep learning models for detecting heart murmur were based on EEGNet and temporal convolutional networks to employ learning frequency-temporal-specific representation from phonocardiogram. To learn patient-specific representation of phonocardiogram, we also utilized demographic information: age, sex, BMI, pregnancy status. From view of frequentist inference, we extracted statistical features and trained two separate Random Forest models to classify heart murmur and patient outcome. The weighted accuracy of murmur detection in 5-fold CV in public training set was 0.727 and challenge cost for outcome detection was 11,886. Our team (amc-sh) recorded 0.688 (rank: 22/40) weighted accuracy in murmur detection, 13,002 (rank: 20/39) challenge cost for outcome detection as final score.