Deep attention-based neural networks for explainable heart sound classification

Cardiovascular diseases are the leading cause of death and severely threaten human health in daily life. There have been dramatically increasing demands from both the clinical practice and the smart home application for monitoring the heart status of individuals suffering from chronic cardiovascular diseases. However, experienced physicians who can perform efficient auscultation are still lacking in terms of number. Automatic heart sound classification leveraging the power of advanced signal processing and deep learning technologies has shown encouraging results. Nevertheless, a lack of explanation for deep neural networks is a limitation for the applications of automatic heart sound classification. To this end, we propose explaining deep neural networks for heart sound classification with an attention mechanism. We evaluate the proposed approach on the heart sounds shenzhen corpus. Our approach achieves an unweighted average recall of 51.2% for classifying three categories of heart sounds, i.e., normal, mild, and moderate/severe. The experimental results also demonstrate that the global attention pooling layer improves the performance of the learnt representations by estimating the contribution of each unit in high-level features. We further analyse the deep neural networks by visualising the attention tensors.