Heart sound classification from wavelet decomposed signal using morphological and statistical features

Automatic classification of heart sound recordings is one of the widely known challenges for over 50 years. The fundamental objective of this study is to evaluate a large database of heart sounds collected from a variety of clinical and non-clinical surroundings and classify them into normal and abnormal categories. Daubechis-2 wavelet transform was applied to each phonocardiogram (PCG) recording after segmenting each cardiac cycle into four windows containing first heart sound S1-Systole-Second heart sound (S2)-Diastole states of a heart cycle. Morphological, statistical and time features were extracted from each cardiac states window. Heart sound classification into normal and abnormal was based on the SVM with Gaussian kernel function. The algorithm was trained by the recordings from all available training data sets (training set A to F). The performance of the proposed prototype was evaluated by five-fold cross-validation on the available training dataset as well as on the hidden test set by PhysioNet. Overall classification accuracies of 82% during Phase I submissions and 77% during Phase II submissions were achieved of the challenge. The final score on the blind test set was 74.65%. Based on the current result, the proposed prototype could be a potential solution for a robust and automatic classification technique of normal and abnormal heart sound recordings.