A novel 1-D densely connected feature selection convolutional neural network for heart sounds classification

Background: Heart sound auscultation, due to it being a non-invasive, convenient, and relatively low-cost technique, remains the dominant method for detection of cardiovascular disease. Methods: In this paper, we present a method for identifying abnormal heart sounds based on a novel Dense Feature Selection Convolution Network framework (Dense-FSNet). The Dense-FSNet is comprised of multiple, circular dense connectivity modules, called Clique Blocks. These Clique Blocks can allow lowlevel and high-level features to stimulate each other for cyclic updates, which subsequently enhances the information flow among layers. Inspired by the channel-wise attention mechanism from Squeeze-and Excitation Networks (SENet), we adopted squeeze-and-excitation block to avoid the progressive growth of parameters. The effect of the model was assessed using the accuracy, specificity, sensitivity, and area under the curve (AUC) values. To improve model performance, in addition to the structures mentioned above, we incorporated a multi-scale attention mechanism into our framework. Results: Using this attention mechanism, our model was able to achieve adaptively spatial feature fusion by adjusting a hyper-feature that contains higher level visual information and lower-level features including edge details and context information. It is worth noting that data balance technology was also used in the process of building the model, and notable results have been achieved. Conclusions: Experience using the PhysioNet/CinC 2016 dataset shows that our proposed Dense-FSNet models achieve state of the art levels in the classification and detection of abnormal heart sounds.