Detection of pulmonary hypertension associated with congenital heart disease based on time-frequency domain and deep learning features

•The original heart sounds were pre-processed first, in which a double-threshold adaptive segmentation method was used to segment the 20 s-long signal into each cardiac cycle.•The pathological information of CHD-related pulmonary hypertension is concentrated in S2, so we extract both the time–frequency domain features of the entire cardiac cycle and the time–frequency domain features of S2. Then the fusion features were extracted. It includes the time–frequency domain features of the entire cardiac cycle and S2 and the depth features which are extracted by convolutional neural network (CNN). The fusion features consist a feature vector which will be input into a classifier.•The XGBoost was used as three-class classifier for the classification of normal, CHD and CHD-PAH. Finally, the majority voting algorithm was used to obtain the optimal classification results. A classification accuracy of 88.61% was achieved using this novel method. The heart sounds reflect the health of the heart. Its recording is the phonocardiogram (PCG). Pulmonary hypertension associated with congenital heart disease (CHD-PAH) is a serious heart disease and is often associated with severe disability and death. The disease is not well characterized onset. The most patients are severe when they have been diagnosed and miss the best time to treat them. The objective of this study was to develop a computer aided diagnosis, which based on single cycle with multiple features, for detecting pulmonary hypertension associated with congenital heart disease. It is a non-invasive and simple method which may be hopeful at early diagnosis of CHD-PAH. The original heart sounds were pre-processed first, in which a double-threshold adaptive segmentation method was used to segment the signal into each cardiac cycle first. Then the time–frequency domain features and wavelet packet energy features of cardiac cycle and S2 component are extracted. And convolutional neural network (CNN) is used to extract the depth features of cardiac cycle. The above features were combined into a fused feature vector. Normal, CHD and CHD-PAH were classified using XGBoost as the classifier. Finally, the majority voting algorithm is used to obtain the best classification result for multiple results corresponding to multiple cardiac cycles of the same person. Using this new method, a classification accuracy of 88.61% was achieved.