Abstract Body (Do not enter title and authors here): Introduction
Heart murmurs are abnormal sound signals generated by turbulent blood flow and are closely associated with specific heart disorders. Current methods to detect and qualify murmurs do not recognize that an important part of diagnosing a patient depends upon murmur location in the heart valves and timing between fundamental heart sounds. This research presents HM-Detect, a novel algorithm to provide accurate heart murmur characteristics to cardiologists, specifically in telemedicine. In particular, can we use signal processing and machine learning techniques to detect and classify heart murmur characteristics (e.g. timing and location) from heart sound signals?
Approach
HM-Detect analyzes heart sound features, namely the filterbank energies and spectral sub-band centroids. Four separate Random Forest Classifiers (RFCs) were built for each of the four valves of the heart using statistical moments of these features. Additionally, an undersampling technique on heart sound data and a novel “multi-modal” long short-term memory (MMLSTM) neural network incorporate the temporal aspect of heart sounds. The algorithms were validated on signals from the CirCor DigiScope dataset, which is a clinically verified dataset. Three different neural network architectures with varying number of cells and MMLSTM layers were compared.
Results and Data Analysis
The feature importance of the first four statistical moments from the RFCs show that the variance has the lowest importance (max value < 0.010) while the mean, skewness, and kurtosis, have higher importance (max value > 0.020). All RFC models were evaluated using the area under their receiver operating characteristic curves (AUROCs). The RFCs were much more successful at classifying murmurs from the pulmonary and tricuspid valves (AUROC = 0.83 and 0.78, respectively) when compared with the aortic and mitral valves (AUROC = 0.72 and 0.65, respectively). Furthermore, the proposed method for murmur timing can achieve a performance accuracy of around 90%, with the best of the three MMLSTM-based models having an F1 score of 0.91 and a test accuracy of 87%.
Conclusions
Heart sound features, namely filterbank energies and sub-band spectral centroids, were successfully used to build RFC algorithms which accurately classify murmur locations. The combination of undersampling and MMLSTMs allow for an encapsulation of the spatiotemporal profile of heart sounds and accurately detect and classify heart murmurs based on murmur timing.