Respiratory Sound Classification: From Fluid-Solid Coupling Analysis to Feature-Band Attention

Abstract

Based on respiratory sound production mechanisms, we study the relationship between airflow characteristics in the bronchi and the sound pressure spectrum curves to implement an end-toend respiratory sound classification system with a feature-band attention module. First, we analyse fluidsolid coupling simulations of the bronchi and execute acoustic simulations to obtain spectrum curves of the bronchi at the sound pressure level. Then, based on the spectrum characteristics of the bronchi, we propose an attention strategy to refine acoustic features with adaptive weights. In addition, we introduce a featureband attention module to ResNet-based networks with a squeeze-and-excitation block. Finally, we perform experiments on the ICBHI public database to classify respiratory sounds as belonging to one of four classes: normal, wheezes, crackles, and both (wheezes and crackles). The results show that our proposed system produces superior performance compared to the baseline system. This type of feature learning strategy is useful for exploring distinct characteristics of different types of respiratory sounds.INDEX TERMS fluid-solid coupling, attention learning, end-to-end system, respiratory sound classification, squeeze-and-excitation.

Respiratory Diseases Diagnosis Using Audio Analysis and Artificial Intelligence: A Systematic Review

“…This results in an increase in the models’ generalization capability, while also expanding the field of application where the trained algorithm can be used. Data constructed by this methodology are employed in [ 44 , 51 , 54 , 55 , 57 , 61 , 62 , 67 , 68 ]. On the other hand, only a limited number of studies have employed smartphone recordings as a means of data collection, with one notable example being [ 63 ], which utilized various smartphones to record lung sounds.…”

“…Nevertheless, most studies are focused on respiratory sounds, since they can be correlated with almost all diseases or at least provide useful information for a preliminary diagnosis. The studies [ 47 , 51 , 52 , 53 , 54 , 55 , 56 , 57 , 58 , 59 , 60 , 61 , 62 , 64 , 65 ] provide different techniques and methodologies for the identification of respiratory sounds, like crackles and wheezes, with promising results. The accuracy achieved in these studies is between 73% and 98%.…”

“…Delving into the data processing field that is followed for implementation, the studies [ 43 , 45 , 46 , 47 , 48 , 49 , 51 , 54 , 55 , 60 , 61 , 62 ] perform Feature Extraction on the data; however, the Feature Extraction comes after several preprocessing steps, namely filtering, upsampling or downsampling, or other transformations. The studies [ 45 , 46 , 49 ] extract several statistical and spectral features, such as spectral bandwidth, zero-crossing rate, spectral roll-off, etc.…”

“…Moreover, in [ 48 ], both statistical features and feature maps, extracted from a Convolutional Bidirectional LSTM Network, are combined in order to provide more descriptive features of the audio data. Some interesting approaches are utilized by [ 54 , 55 ]. The first study proposed a method of feature-band attention by analyzing the fluid-solid coupling simulation of a bronchial model.…”

“…Other studies that employ ANNs with features as inputs to perform the training procedure are [ 47 , 48 , 60 ]. The studies [ 49 , 53 , 54 , 58 , 65 ] use some of the state-of-the-art models. In Ref.…”