Heartbeat murmurs detection in phonocardiogram recordings via transfer learning

Heart murmurs are abnormal heartbeat patterns that could be indicative of a serious heart condition, which can only be detected by trained specialists with the use of a stethoscope. However, it is occasionally the case that those specialists are not available, resulting in the need for a machine-automated system for murmur detection. Many methods might be used to produce such a system, one of which is the utilization of transfer learning. A recent machine learning method that saw popularity due to the little time it needs for training and the boosted accuracy it produces. This paper aims at testing the performance of transfer learning when detecting murmurs of the heart, by evaluating three transfer learning models, namely, VGG16, VGG19, and ResNet50, trained on a database of phonocardiogram (PCG) heartbeat recordings, i.e., PASCAL CHSC database. The data is cleansed, processed, and converted into images using two signal representation methods; Spectrograms and Mel Frequency Cepstral Coefficients (MFCCs). The paper compares the results of each model, using metrics of accuracy and loss, where the use of Spectrograms proved to yield the best results with 83.95%, 83.95%, and 87.65%, classification accuracy for VGG16, VGG19, and ResNet50, respectively. Based on the findings of the paper, it is evident that the Spectrogram-ResNet50 transfer learning pipeline could further facilitate the detection of heart murmurs with less time spent on training.