Abnormal heart sounds detected from short duration unsegmented phonocardiograms by wavelet entropy

Segmentation of the characteristic heart sounds is thought to be an essential requirement for the automatic classification of phonocardiograms. The aim of this work was to test the feasibility of classification using short duration, unsegmented recordings. Recordings from the 2016 PhysioNet/Computing in Cardiology Challenge were analysed. Wavelet entropy of unsegmented 5 s duration recordings was calculated and the optimum wavelet scale and wavelet entropy threshold determined from the training set. The algorithm was validated on the test set. At a wavelet scale of 1.7 wavelet entropy was significantly reduced in abnormal recordings (median (IQR), 6.3 (1.8) vs 8.0 (1.8) p<;0.0001). For a wavelet entropy threshold of 7.8 a score of 78% (sensitivity = 95%, specificity = 60%) was obtained on the training set. The robustness of this result was demonstrated on the test set which achieved a score of 77% (sensitivity = 98%, specificity = 56%). Classification of unsegmented and short duration phonocardiograms by wavelet entropy is feasible.