Systems and methods for detecting pulmonary abnormalities using lung sounds

Patent metadata

Field	Value
Publication number	US-2019008475-A1
Application number	US-201815912234-A
Country	US
Kind code	A1
Filing date	Mar 5, 2018
Priority date	Jul 4, 2017
Publication date	Jan 10, 2019
Grant date	—

Abstract

Identification of pulmonary diseases involves accurate auscultation as well as elaborate and expensive pulmonary function tests. Also, there is a dependency on a reference signal from a flowmeter or need for labelled respiratory phases. The present disclosure provides extraction of frequency and time-frequency domain lung sound features such as spectral and spectrogram features respectively that enable classification of healthy and abnormal lung sounds without the dependencies of prior art. Furthermore extraction of wavelet and cepstral features improves accuracy of classification. The lung sound signals are pre-processed prior to feature extraction to eliminate heart sounds and reduce computational requirements while ensuring that information providing adequate discrimination between healthy and abnormal lung sounds is not lost.

First claim

What is claimed is:

1 . A processor implemented method ( 200 ) comprising:
receiving a plurality of auscultation sound signals to be classified from a plurality of subjects ( 202 ); pre-processing the received plurality of auscultation sound signals, wherein the pre-processing comprises resampling of the plurality of auscultation sound signals and removing heart sound signals to obtain a plurality of lung sound signals ( 204 ); extracting one or more of spectral features, spectrogram features, wavelet features and cepstral features from the plurality of lung sound signals ( 206 ); selecting a plurality of discriminative features from the extracted features ( 208 ); and classifying the plurality of lung sound signals based on the selected discriminative features ( 210 ).

2 . The processor implemented method of claim 1 , wherein the resampling comprises down-sampling the plurality of auscultation sound signals to a pre-defined frequency based on a range of frequencies that provide optimum discrimination between healthy and abnormal lung sounds.

3 . The processor implemented method of claim 1 , wherein the step of removing heart sound signals is based on Empirical Mode Decomposition method.

4 . The processor implemented method of claim 1 , wherein the step of extracting comprises dividing the plurality of lung sound signals into a plurality of overlapping windows.

5 . The processor implemented method of claim 1 , wherein the step of extracting spectral features comprises:
computing periodograms for each of the plurality of overlapping windows; and averaging the periodograms to obtain a Power Spectral Density (PSD) estimate curve for the plurality of lung sound signals.

6 . The processor implemented method of claim 5 , wherein the extracted spectral features are areas under a normalized Power Spectral Density (PSD) estimate curve corresponding to a pre-defined number of frequency bands from 0-1.5 kHz (PS 1 to PS 15 ), ratio of the spectral power below 500 Hz (P 1 ) to that from 500 Hz to 1500 Hz (P 2 ), spectral centroid (S cent ), spectral flux (S flux ), spectral rolloff (S roll ) and spectral kurtosis (S kurt ), wherein the pre-defined number of frequency bands is selected such that optimum discrimination between healthy and abnormal lung sounds is achieved.

7 . The processor implemented method of claim 4 , wherein the extracted spectrogram features are median spectral powers (SP 1 to SP 58 ) across the plurality of overlapping windows, wherein the median spectral powers are computed for a predefined number of equally spaced frequencies between 0.15 kHz-1.5 kHz of the plurality of lung sound signals, the pre-defined number of equally spaced frequencies being based on the computational complexity and frequency resolution.

8 . The processor implemented method of claim 4 , wherein the step of extracting wavelet features comprises:
selecting a best mother wavelet in each window of the plurality of overlapping windows based on maximum energy and minimum Shannon entropy criteria; decomposing the plurality of lung sound signals using the best mother wavelet into decomposition levels; and computing median of absolute values of approximation and detail coefficients for the decomposition levels.

9 . The processor implemented method of claim 8 , wherein the extracted wavelet features (W 1 to W 21 ) are (i) the median of absolute values of approximation and detail coefficients for the decomposition levels and (ii) ratios thereof across sub-bands of the plurality of overlapping windows.

10 . The processor implemented method of claim 4 , wherein the extracted cepstral features are mean (mfccm i and lfccm i ) and standard deviation (mfccsd i and lfccsd i ) of Mel Frequency Cepstral Coefficients (MFCC) and Linear Frequency Cepstral Coefficients (LFCC).

11 . The processor implemented method of claim 1 , wherein the step of selecting a plurality of discriminative features from the extracted features comprises:
ranking the extracted features in decreasing order of importance; and selecting a plurality of discriminative features based on the optimal no. of features that result in a high performance value and a low standard deviation of a set of pre-defined performance metrics.

12 . The processor implemented method of claim 11 , wherein the set of pre-defined performance metrics comprise accuracy, sensitivity, specificity and area under the receiver operating characteristic curve.

13 . A system ( 100 ) comprising:
one or more data storage devices ( 102 ) operatively coupled to one or more hardware processors ( 104 ) and configured to store instructions configured for execution by the one or more hardware processors to: receive a plurality of auscultation sound signals to be classified from a plurality of subjects; pre-process the received plurality of auscultation sound signals by resampling the plurality of auscultation sound signals and removing heart sound signals to obtain a plurality of lung sound signals, wherein the resampling is performed by down-sampling the plurality of auscultation sound signals to a pre-defined frequency based on a range of frequencies that provide optimum discrimination between healthy and abnormal lung sounds; extract one or more of spectral features, spectrogram features, wavelet features and cepstral features from the plurality of lung sound signals; select a plurality of features from the extracted features; and classify the plurality of lung sound signals based on the selected discriminative features.

14 . The system of claim 13 , wherein the extracted spectral features are areas under a normalized Power Spectral Density (PSD) estimate curve corresponding to a pre-defined number of frequency bands from 0-1.5 kHz (PS 1 to PS 15 ), ratio of the spectral power below 500 Hz (P 1 ) to that from 500 Hz to 1500 Hz (P 2 ), spectral centroid (S cent ), spectral flux (S flux ), spectral rolloff (S roll ) and spectral kurtosis (S kurt ), wherein the pre-defined number of frequency bands is selected such that optimum discrimination between healthy and abnormal lung sounds is achieved.

15 . The system of claim 13 , wherein the extracted spectrogram features are median spectral powers (SP 1 to SP 58 ) across a plurality of overlapping windows obtained by dividing the plurality of lung sound signals, wherein the median spectral powers are computed for pre-defined number of equally spaced frequencies between 0.15 kHz-1.5 kHz of the plurality of lung sound signals, the pre-defined number of equally spaced frequencies being based on the computational complexity and frequency resolution.

16 . The system of claim 13 , wherein the extracted wavelet features (W 1 to W 21 ) are (i) median of absolute values of approximation and detail coefficients for decomposition levels obtained from the plurality of lung sound signals using a best mother wavelet and (ii) ratios thereof across sub-bands of a plurality of overlapping windows obtained by dividing the plurality of lung sound signals.

17 . The system of claim 13 , wherein the extracted cepstral features are mean (mfccm i and lfcccm i ) and standard deviation (mfccsd i and lfccsd i ) of Mel Frequency Cepstral Coefficients (MFCC) and Linear Frequency Cepstral Coefficients (LFCC).

18 . The system of claim 13 , wherein the one or more hardware processors are further configured to select a plurality of discriminative features from the extracted features by:
ranking the extracted features in decreasing order of importance; and selecting a plurality of discriminativ

Assignees

• Tata Consultancy Services Ltd

Inventors

• Datta Shreyasi
• Dutta Choudhury Anirban
• Deshpande Parijat
• Bhattacharya Sakyajit
• Pal Arpan

Classifications

• A61B5/7267

  involving training the classification device · CPC title

• A61B5/0816

  Measuring devices for examining respiratory frequency (measuring frequency of electric signals G01R23/00) · CPC title

• A61B7/003 Primary

  Detecting lung or respiration noise · CPC title

• A61B7/04

  Electric stethoscopes · CPC title