A Novel Method for Automatic Identification of Respiratory Disease from Acoustic Recordings

Abstract

This paper evaluates the use of breath sound recordings to automatically determine the respiratory health status of a subject. A number of features were investigated and Wilcoxon Rank Sum statistical test was used to determine the significance of the extracted features. The significant features were then passed to a feature selection algorithm based on mutual information, to determine the combination of features that provided minimal redundancy and maximum relevance. The algorithm was tested on a publicly accessible respiratory sounds database. With the testing dataset, the trained classifier achieved accuracy of 87.1%, sensitivity of 86.8% and specificity of 93.6%. These are promising results showing the possibility of determining the presence or absence of respiratory disease using breath sounds recordings.

Automatic Lung Health Screening Using Respiratory Sounds

“…We also compared the performance of our system with reported works by Kok et al [12] and Chambers et al [6]. The average accuracies for both the systems along with the proposed system are provided in Table 10.…”

“…The average accuracies for both the systems along with the proposed system are provided in Table 10. Kok et al [12] 87.10 Chambers et al [6] 85.00 Proposed technique 99.22…”

“…In their study, the reported accuracy was 98.79%. Kok et al [12] employed multiple features, such as MFCC, DWT, and time domain metrics to distinguish healthy and nonhealthy sounds. In their study, they reported accuracy, specificity, and sensitivity values of 87.1%, 93.6%, and 86.8%, respectively on the ICBHI17 dataset.…”

CNN-MoE Based Framework for Classification of Respiratory Anomalies and Lung Disease Detection

“…Research into the automated detection or analysis of respiratory sounds has some precedents [2], [3], [4], but has drawn increasing attention in recent years as robust machine hearing methods have been developed, leveraging on ever more capable deep learning techniques. Most existing respiratory sound analysis systems tend to rely upon frame-based feature representations such as Mel-Frequency Cepstral Coefficients (MFCC) [5], [6], borrowed from the Automatic Speech Recognition (ASR) and Speaker Recognition (SR) fields. However, Grønnesby et al [7] found that MFCCs did not represent crackles well.…”

“…Next, they applied a Maximal Information Coefficient (MIC) [11] to score each feature, selected only the most influencing, before feeding into a classifier to improve performance and reduce complexity. Similarly, Kok et al [6] applied the Wilcoxon Sum of Rank test to indicate which features among MFCCs, Discrete Wavelet Transform (DWT) and a set of time domain features (namely power, mean, variance, skewness and kurtosis of audio signal) mainly affected final classification accuracy. Image processing techniques were then tried by Sengupta et al [12], who employed Local Binary Pattern (LBP) analysis on mel-frequency spectral coefficients (MFSCs) to capture texture information from the MFSC spectrogram, thus obtained an LBP spectrogram.…”