Comparing and Combining Audio Processing and Deep Learning Features for Classification of Heartbeat Sounds

Abstract

Machine learning has been increasingly employed in healthcare. Considering the alarming number of deaths caused by cardiovascular diseases globally, tackling problems involving heart-related data is particularly important. This paper investigates how feature engineering influences classification performance. We used a support vector machine with three different feature extraction techniques: firstly, audio signal processing features; secondly, deep learning features from a VGG-like architecture pre-trained on Google’s AudioSet; lastly, concatenated deep learning features from the VGG16 and VGG19 architectures pre-trained on the ImageNet dataset. Finally, we combined all approaches through majority voting or feature concatenation. We tested our methods on two datasets from the PASCAL Classifying Heart Sounds Challenge and compared them with previous methods in the literature. Experimental results show how audio processing and deep learning features through spectro-grams might interchangeably hold the same relevant information for this application, regardless of the pre-training dataset, and how experimentation is still recommended.