Classification of Murmurs in PCG Using Combined Frequency Domain and Physician Inspired Features

Abstract

Physiological machine learning methods have a unique opportunity to augment deep-learning engineered features with additional features derived from prior pathological knowledge. We propose an phonocardiogram (PCG) classifier that combines raw spectrogram features with crafted, physician-inspired features with an end-to-end neural network architecture. Learned spectrogram features were obtained by training a convolutional neural network (CNN) directly on the raw mel-spectrogram representation of the PCG time-series. Crafted features were based on the four stages of the cardiac cycle (S1, systole, S2, and diastole). The spectrogram features have the advantage of introducing flexibility for the model to learn abstract, low-level information that captures a variety of different rhythmic abnormalities and the latter has the advantage of using segmentation to elucidate specific, high-level, human-interpretable information. Combined features are fed into a fully connected neural network which is able to learn the relationship between the two feature types. In the George B. Moody PhysioNet Challenge 2022 test set, our team (“lubdub”) received a weighted accuracy score of 0.835 with a cost of 14905 in the clinical outcome task (ranked 31/39). For the murmur prediction task, our model received a weighted accuracy score of 0.525 and a cost of 15083 (ranked 33/40).