A Feature Polymerized Based Two-Level Ensemble Model for Respiratory Sound Classification
Lin Zhang, Yangxin Zhu, Shikui Tu, Lei Xu
Abstract
Accurate analysis and classification of respiratory sound play an important role in the diagnosis of respiratory diseases. Most of the current methods for classifying respiratory sounds use deep neural networks to classify the spectrograms generated from the audio. However, the parameters of the generated spectrograms are difficult to adjust and the features obtained by these methods are rather homogeneous. In this paper, we propose a Feature Polymerized (FP) based Two-Level Ensemble Model (TLEM) for respiratory sound detection, shortly called FP-TLEM to make effective use of auscultation recordings. In FP-TLEM, considering the expiration and inspiration or the steady and non-steady stage in signal, we divide the signals to conquer and aggregate rather than extract from the whole breathing process or the whole signal. Instead of using just spectrogram features, FP extract various features such as MFCCs, spectrograms, chromagrams features from each segment. We also use the technique of oversampling to solve the problem of unbalanced sample size between normal and adventitious samples. Considering gender is essential in data distribution and is the main factor influencing building the modeling, a two-level ensemble model technique is adopted to build a robust and promising classifier. Experiments on newly released SPRSound database demonstrate the effectiveness of our method.