Transfer learning based cardiac murmur detection in phonocardiogram signals using spectrograms.

This study uses transfer learning architectures to detect cardiac murmurs in phonocardiogram signals by denoising the signal, extracting relevant features for spectrograms generation. The Short-Time Fourier Transform, Mel-Frequency Cepstral Coefficients, and Continuous Wavelet Transform techniques were applied on Physionet's CirCor Digiscope PCG dataset. VGG16, VGG19, ResNet50, and InceptionV3 models, were trained on these spectrograms for binary classification. Fourth-order Butterworth bandpass filter, used with Savitzky-Golay filtering, gave the best results. The CWT Spectrogram and VGG19 combination yielded best accuracy of 89.44%. Different combinations of spectrograms and transfer learning architectures performed better on performance metrics of precision, recall, F1-score, and ROC-AUC.