abstract

• Lung sounds represent a relevant source of information about the state of the lung parenchyma. Accordingly, several efforts have been made to obtain a quantitative characterization as well as a classification of respiratory information. This paper proposes a methodology for the characterization of multichannel lung sounds (LS), based on the feature extraction from the joint time-frequency domain through the short-time Fourier transform to contend with the non-stationary nature of LS. Subsequently the classification into two classes, healthy and sick, was performed by an artificial neural network (ANN). The average percentage of correct classification for new subjects, whose information was not used in ANN design, was 79.68%25 for healthy subjects and 81.88%25 for patients diagnosed with diffuse interstitial pneumonia. Results indicate that the average percentage of correct classification for healthy subject improved when compared to results obtained in previous efforts of our research group where other techniques, such as univariate and multivariate techniques, that assumed stationarity of data were used. © 2018 IEEE.

authors

• Aljama-Corrales T.
• Charleston-Villalobos S.
• Del-Rio A.
• González-Camarena R.
• Mejía-Ávila M.
• Reyes Bersain Alexander

publication date

• 2019-01-01

published in

• 2018 IEEE International Autumn Meeting on Power, Electronics and Computing, ROPEC 2018  Proceedings

keywords

• Artificial Neural Networks; Gray Level Co-occurrence Matrix; Joint Time-Frequency Representation; Lung Sounds; Multichannel Acquisition Biological organs; Frequency domain analysis; Neural networks; Patient monitoring; Gray level co-occurrence matrix; Lung sounds; Multi-channel acquisition; Multivariate techniques; Quantitative characterization; Short time Fourier transforms; Time frequency domain; Time-frequency representations; Classification (of information)

Digital Object Identifier (DOI)

• 10.1109/ROPEC.2018.8661356

start page

end page