Automatic Heart Sound Recording Classification using a Nested Set of Ensemble Algorithms

Automated phonocardiogram (PCG) analysis may provide better clinical information to physicians for analyzing and diagnosing different heart abnormalities. However, despite recent advances in PCG analysis methods, it is still a challenging task to extract accurate and useful information from contaminated heart sound recordings. The main objective of this paper is to introduce a new approach for classification of normal and abnormal heart sound recordings using a nested ensemble of algorithms that includes Random Forest, LogitBoost and a Cost-Sensitive Classifier. The approach consisted of three stages: preprocessing, classification and evaluation. In the preprocessing stage, PCG signals were first downsampled to 1 kHz using a polyphase antialiasing filter. Next, each heart sound was segmented using Springer's improved version of Schmidt's method to identify four states; S1, S2, systole and diastole. Thereafter, 131 features in time, frequency, wavelet and statistical domains were extracted from the entire signal and from the timings of the states. In the classification stage, the meta-classifier was cross validated on the entire training dataset provided by Physionet Challenge 2016. In the evaluation stage, the sensitivity and specificity of the trained algorithm was tested with unseen signals selected randomly by the Challenge testing environment. Experimental results showed that the proposed approach achieved an overall score of 84.48%, ranking 5th. The use of a nested set of ensemble classifier with a combined set of features extracted from different domains helped reduce overfitting and improved classification performance.