Phonocardiography-Based Automated Detection of Prosthetic Heart Valve Dysfunction Using Persistence Spectrum and Interpretable Deep CNN

Cardiac prosthetic valve dysfunction (PVD) is a life-threatening complication of valve replacement surgery (VRS). It is therefore crucial to monitor the mechanical prosthetic heart valve (MPHV) functioning regularly. The standard diagnostic method, cine fluoroscopy (CF) involves X-ray exposure and may not be available to a large population. Therefore, a wearable modality like phonocardiogram (PCG) seems to be a promising alternative. The proposed work is a novel method to automate PCG-based PVD detection. 2-D convolutional neural network (CNN) is explored toward the automated classification of persistence spectrum images of the PCG. Persistence spectrum, a time-frequency representation, displays the duration for which a particular frequency is present. It enables the identification of the hidden components of a signal. This work explores persistence spectrum for PCG analysis. In all, 4215 PCG samples (2127 normal and 2088 PVD) were used for training and testing the CNN. Two AI interpretation techniques, occlusion maps and deep dream images (DDIs), are used to introduce interpretability in the DL model’s decision-making. The overall accuracy of 95.73 (SD = 7.62)% is achieved during fivefold cross-validation (CV) with the highest accuracy of 100% for three out of five folds. The performance during the leave-one-subject-out CV (LOSOCV) is 90.64 (SD = 27.98)%. Through AI interpretation, novel findings of MPHV’s PCG characteristics in the spectral domain, corresponding to cardiac events of normally functioning MPHV and PVD, are revealed, making the CNN decision more transparent. The novel explainable DL model may potentially address PVD-induced clinical burden in resource-constrained settings with no radiation exposure and can be used for screening.