Precision/recall trade-off analysis in abnormal/normal heart sound classification

Heart sound analysis is a preliminary procedure performed by a physician and involves examining the heart beats to detect the symptoms of cardiovascular diseases (CVDs). With recent developments in clinical science and the availability of devices to capture heart beats, researchers are now exploring the possibility of a machine assisted heart sound analysis system that can augment the clinical expertise of the physician in early detection of CVD. In this paper, we study the application of machine learning algorithms in classifying abnormal/normal heart sounds based on the short (≤120 s) audio phonocardiogram (PCG) recordings. To this end, we use the largest public audio PCG dataset released as part of the 2016 PhysioNet/Cardiology in Computing Challenge. The data comes from different patients, most of who have had no previous history of cardiac disease and some with known cardiac diseases. In our study, we use these audio recordings to train three different classification algorithms and discuss the effects of class imbalance (normal vs. abnormal) on the precision-recall trade-off of the prediction task. Specifically, our goal is to find a suitable model that takes into account the inherent imbalance and optimize the precision-recall trade-off with a higher emphasis on increasing recall. Bagged random forest models with majority (normal) class under sampling gave us the best configuration resulting in average recall over 91% with nearly 64% average precision.