Deep learning based murmur detection from PCG signals collected at four valve locations using joint optimization and decision fusion

Detecting a murmur at an early stage potentially saves lives due to many adverse cardiac complications. Phonocardiogram (PCG) signals, collected from the heart valves' neighborhood can be used to detect murmurs. In this paper, a deep learning based murmur detection scheme is proposed that introduces parallel networks to utilize four valve-specific PCG signals under a joint optimization framework. Here, each network incorporates a temporal pattern-extracting convolutional block and a nonlinear activation-based self-attention block with dilated convolution. The network can efficiently detect murmurs by utilizing the characteristics of the four valve-specific information of the PCG signals. It is shown that the use of a decision fusion mechanism, combining detection results obtained by jointly optimizing four separate deep learning networks, trained by using valve-specific PCG signals, can significantly enhance the overall murmur detection performance. Moreover, the proposed scheme can also detect the murmur by using a single network without explicitly utilizing the location information of the heart sounds. For the purpose of performance evaluation, a publicly available dataset containing four valve-specific PCG signals is used. A very satisfactory murmur detection performance, such as an accuracy of 92.21% is achieved when four heart valve-specific PCG signals are used, and 88.76% is obtained when no valve-specific information is used. In comparison to other methods, our network is a very competitive candidate for murmur detection. • This manuscript introduces a novel end-to-end architecture that uses valve-specific PCG signals to detect murmur. • The proposed deep learning model employs four parallel networks to process multi-location PCG signals, where a self-attention block with dilated convolution is incorporated as a key feature. • The performance of the model is compared with some state-of-the-art techniques that valve-specific PCG signals enhance murmur detection accuracy over methods without valve location information.