Abstract

Info

Author(s):: REN Ling; HUANG Yudan; CHEN Ying; School of Electrical Science and Engineering, Nanjing University, Nanjing 210023, China

Keywords:: Keywords: heart sound classification cross-contrast neural network information-based similarity measurement theory deep learning

Abstract:: Abstract: Objective To automatically classify heart sound signals by cross-contrast neural network (CCNN), thereby realizing the early diagnosis of cardiovascular diseases. Methods The experiment was carried out based on PhysioNet/Cinc 2016 heart sound database. The training set and test set data which came from mutually exclusive healthy subjects/pathological patients were divided at a ratio of 4:1 and then input into CCNN. Finally, CCNN used deep convolutional neural network for feature extraction and was combined with information-based similarity measurement theory to measure and classify the similarity between feature vectors. Results The sensitivity and specificity of CCNN for heart sound classification in the experiment were 0.834 6 and 0.962 3, respectively, and the overall score reached 0.898 5. Conclusion By expanding the amount of data using a cross-contrast input mode, introducing contrast information between signals and applying statistical ideas in the training process of neural networks, CCNN has good generalization and is more suitable for small medical data, having a good performance in heart sound classification.