Improving classification accuracy of heart sound recordings by wavelet filter and multiple features

This study aimed to improve the accuracy for normal/abnormal classification of heart sound recordings from PhysioNet/Computing in Cardiology Challenge 2016. In order to get the main elements of the first heart sound (S1) and the second heart sound (S2) for segmentation, the Butterworth filter with a pass-band of 25-400 Hz was replaced by the wavelet filter with the pass-band of 31.25-250 Hz. The pre-process in the example entry was modified to improve the accuracy of heart sound segmentation. The re-sampled heart sound was segmented into S1, systole, S2 and diastolic using a duration dependant logistic regression-based hidden semi-Markov model (HSMM). Then, twenty basic time domain features were calculated. Based on the above twenty features, four frequency domain features, four entropy features and two time domain features were added to improve the classification accuracy. Using the logistic regression method, the heart sound recordings were classified into normal and abnormal ones based on the obtained features. To evaluate the modified program, the sensitivity (Se) and specificity (Sp) of the classification results were presented. When performing on the hidden test set, we got the best results as Se of 71.6%, Sp of 78.2%, and the overall score of 74.9%.