Ensemble methods with outliers for phonocardiogram classification

Heart sound classification and analysis play an important role in the early diagnosis and prevention of cardiovascular disease. To this end, this paper introduces a novel method for automatic classification of normal and abnormal heart sound recordings.

Signals are first preprocessed to extract a total of 131 features in the time, frequency, wavelet and statistical domains from the entire signal and from the timings of the states. Outlier signals are then detected and separated from those with a standard range using an interquartile range algorithm. After that, feature extreme values are given special consideration, and finally features are reduced to the most significant ones using a feature reduction technique. In the classification stage, the selected features either for standard or outlier signals are fed separately into an ensemble of 20 two-step classifiers for the classification task. The first step of the classifier is represented by a nested set of ensemble algorithms which was cross-validated on the training dataset provided by PhysioNet Challenge 2016, while the second one uses a voting rule of the class label.

The results show that this method is able to recognize heart sound recordings efficiently, achieving an overall score of 96.30% for standard signals and 90.18% for outlier signals on a cross-validated experiment using the available training data.

The approach of our proposed method helped reduce overfitting and improved classification performance, achieving an overall score on the hidden test set of 80.1% (79.6% sensitivity and 80.6% specificity).