Research of heart sound classification using two-dimensional features

Heart sound plays a vital role to achieve an accurate diagnosis of cardiovascular diseases, and its auxiliary diagnosis methods have become a hotspot. Aim: In this paper, novel classification algorithms that transfer heart sound classification into image classification are proposed to select better features. The features used were all important in clinical diagnosis. Method: First, four open datasets are used to construct an integrated dataset. Second, the data is preprocessed. Third, two-dimensional features are extracted. In the end, different methods like traditional machine learning, deep learning, and transfer learning are applied to classify heart sounds. Results: The results show that logmel and logpower can achieve a better effect than envelope and waveform, and the average accuracy is improved by 6–10%, which can achieve around 94%. F1 score shows a trend consistent with accuracy. This is verified by both machine learning and deep learning methods. Under the experimental conditions in this paper, transfer learning can promote the effect of Xception and MobileNet, the accuracy can improve by about 2% on time-domain features. The results of transfer learning are comparatively more stable, and more results are within the 95% confidence interval. Conclusion: This paper uses different methods to systematically compare the effects of different two-dimensional features in heart sound classification, and explains why different features achieve different effects from different perspectives such as clinical, and provides new insights like the application of feature fusion in it.