INDICON_Presentation_LungSound_COPD.pdf

Full text

Automated Severity Detection Of Chronic
Obstructive Pulmonary Disease Using Lung
Sounds
by:
Arka Roy, and Udit Satija
Department of Electrical Engineering
Indian Institute of Technology Patna, Bihar, 801106
November 20, 2022
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 1 / 19

Content
Introduction to lung sound
Chronic obstructive pulmonary disease (COPD)
Motivation for lung sound based diagnostic modality
Description of database
Proposed framework
Performance comparison
Conclusion
References
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 2 / 19

Introduction to lung sound
Produced due to tubulan air flow in trachea-bronchial tree
Can be heard at the course of chest auscultation
Linked to structural faults occurred in lungs due to disease condition
Table 1: Description of different lung sounds
Lung
sound
Frequency
range
Inspiratory/
expiratory
Associated
disease
Normal bellow
1000Hz Both Heathy
Wheeze 200-2500Hz expiration or
both in severe condition
COPD,
bronchial asthma
Crackle 100-500Hz Inspiratory Pneumonia.
pulmonary edema
Pleural rub 200-4000Hz Inspiratory Lung tumor, infection
in inner linings
Figure 1: Human respiratory system
Figure 2: Cross sectional view of (a) healthy bronchous and (b) COPD diseased bronchous
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 3 / 19

Chronic obstructive pulmonary disease (COPD)
COPD is associated two airway obstruction defect:
Emphysema: destruction of the elastic fibers of the alveoli.
Chronic bronchitis: bronchial tubes become inflamed and narrowed.
Cause: Cigarette smoking, secondhand smoke, pipe smoke, α1-antitrypsin deficiency.
Alarming statistics:
According to WHO, COPD is the world’s third leading cause of mortality[1]
COPD accounted for 75.6 percent of all chronic respiratory illness in India in 2016 [2].
a b
Figure 3: (a) Airway obstructions in COPD, (b) World map of the prevalence of clinical COPD [1]
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 4 / 19

Motivation for lung sound based diagnostic modality
Gold standard modality: Spirometry or lung function test based parameters (FEV1, FVC
measure).
Disadvantage of spirometry:
Highly dependent on patient efforts, cooperation with the technician.
Laborious procedure, especially for younger and the elderly
Costly
Why intervention of lung sounds:
Observable wheezing sounds during auscultation.
Low-cost non-invasive modality.
Table 2: Different COPD severity levels described by GOLD [3]
COPD
severity
qualitative
severity level
Gold standard
measures (FEVR)
Physiological
effect
COPD0 Under Risk >85% non-chronic symptoms,
persistent cough
COPD1 Mild Level >80% Light wheezing
COPD2 Moderate level 50% - 80% chronic symptoms,
wheezing
COPD3 Severe Level 30% - 50% all chronic symptoms,
pulmonary infections
COPD4 Very severe level <30% bedridden case with
respiratory machine Figure 4: Spirometric measurement
apparatus
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 5 / 19

Database
Database used: RespiratoryDatabase@TR [3]
Table 3: Description of RespiratoryDatabase@TR[3]
Recording device Littmann3200 digital stethoscope
Sampling rate 4000 Hz
Acquisition sites 4 anterior, 8 posterior (Fig.5)
Recording length Uneven, atleast 17 sec
Distribution
of classes
COPD0: 5, COPD1: 7,
COPD2:7, COPD3:7, COPD4: 10
Figure 5: Lung auscultation positions on the anterior and posterior side of body.
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 6 / 19

Proposed framework for automated severity detection of COPD using lung sounds
Objective: To develop an automated COPD severity detection system using the lung sound
signal based on signal processing and machine learning technique.
Figure 6: Block diagram of the proposed COPD severity classification framework
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 7 / 19

Preprocessing
Temporal snippet generation:
Let, s[n] be the raw lung sound signal, being segmented into 10 sec. snippets (st[n])
keeping 50% overlap with adjacent frame.
DFT based baseline wander (BW) component removal [4]:
DFT of the tth temporal snippet (st(n)) is calculated as:
St(K) =
N−1
X
n=0
st(n)e
−j2πnK
N(1)
Frequency range of BW component is 0-1 Hz. Thereby, remove DFT coefficients which
are smaller than 1 Hz. DFT coefficient index for the fHz: K=⌈fN
fs⌉where, fsdenotes
the sampling rate of lung sound. Threshold the DFT coefficient:
˜
St(K) = [0, ...., 0,St[K+ 1], ...St[N−K−1],0, ...., 0]
Baseline wander removed signal:
LSt
bwf (n) = 1
N
N−1
X
K=0
˜
St(K)ej2πnK
N(2)
Normalization:
LSt
norm(n) = LSt
bwf (n)
max|(LS t
bwf (n))|(3)
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 8 / 19

Preprocessing (Contd.)
012345678910
-1
-0.5
0
0.5
s[n]
012345678910
-0.05
0
0.05
s[n]-LSt
bwf(n)
012345678910
-1
-0.5
0
0.5
LSt
bwf(n)
012345678910
Time (in secs)
-1
0
1
LSt
norm(n)
(c)
(a)
(b)
(d)
Figure 7: (a) Lung sound snippet, (b) BW component, (c) BW removed signal, (d) Normalised lung sound snippet.
012345678910
-1
0
1
Amplitude
012345678910
Time in sec
-1
0
1
(b)
(a)
Figure 8: Preprocessed lung sound signals of (a) COPD-0 and (b) COPD-4 subject
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 9 / 19

Variational mode decomposition (VMD) [5]
VMD [5] decomposes a 1-D signal: v(t), into Knumber of modes {hk}, each with a different
center frequency {ωk}with compact bandwidth. The IMF extraction process is described as:
calculate analytic version of each mode {hk}to get unilateral frequency spectrum
shift the frequency component to base-band
compute the bandwidth by considering L2norm of the gradient resulting in the following
equality constrained variational problem:
minhk,ωknPk
∂t[(δ(t) + j/πt)∗hk(t)] e−jωkt

2
2o
s.t. Pkhk(t) = v(t)(4)
Unconstrained optimization is formulated by adding augmented Lagrangian function:
L(hk, ωkλ) = αX
k



∂tδ(t) + j
πt∗hk(t)e−jωkt



2
2
+




v(t)−X
k
hk(t)




2
2
+*λ(t),v(t)−X
k
hk(t)+(5)
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 10/ 19

VMD (Contd.)
The minimization problem is solved by using ADMM optimization which results in the following
update equations for modes and center frequencies:
ˆ
hq+1
k(ω)←ˆv(ω)−Pi<kˆ
hq+1
i(ω)−Pi>kˆ
hq
i(ω) + ˆ
λq(ω)
2
1+2αω−ωq
k2(6)
ωq+1
k←R∞
0ω
ˆ
hq+1
k(ω)
2
R∞
0
ˆ
hq
k(ω)
2(7)
0246810
-1
0
1
v(t)
0246810
-0.5
0
0.5
h1(t)
0246810
-0.2
0
0.2
h2(t)
0246810
-0.05
0
0.05
h3(t)
0246810
-0.02
0
0.02
h4(t)
0246810
-0.5
0
0.5
h5(t)
0246810
Time in sec
-0.5
0
0.5
h6(t)
0246810
Time in sec
-0.2
0
0.2
h7(t)
(b)
(d)
(f)
(c)
(e)
(g) (h)
(a)
Figure 9: (a) Lung signal of COPD-0 subject, (b-h) IMFs extracted using VMD
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 11/ 19

VMD based feature extraction
Mean= 1
NPN
n=1 h[n] [9] St. dev=q1
NPN
n=1(h[n]−¯
h)2[9]
skewness 1
NPN
n=1
(h[n]−¯
h)
σ
3[9] Kurtosis= 1
Nσ4PN
n=1(h[n]−¯
h)4[9]
RMS=q1
NPN
n=1 h[n]2[9] Energy=PN
n=1 h[n]2[9]
Shannon entropy=−Pih(i)log(h(i)) [9] Log energy entropy=Pilog(h2(i)) [9]
ZCR= 1
2NPN
n=1 |sgn(h[n]) −sgn(h[n−1])|[9]
Approximate entropy(M,R)=limn→+∞ΨM(R)−ΨM+1(R)
where, ΨM(R) = 1
N−M−1PN−m+1
i=1 logCM
i(R), and CM
i(R) provides count in resolution R,Mis
embedded dimension, Rdenotes a threshold value.
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 12/ 19

Feature Selection using analysis of variance (ANOVA) test
Analysis of variance (ANOVA) test-based feature selection
Considered the features which have p−value <0.05
Table 4: Selected Features By Employing ANOVA Test
IMF No. Selected features with p−value <0.05
IMF1 ShEn, RMS, st. dev, energy, ZCR, ApEn, LogEn
IMF2 ZCR, ApEn, LogEn
IMF3 ZCR, ApEn, LogEn
IMF4 Kurtosis, ApEn, LogEn, ZCR
IMF5 ApEn, ZCR, LogEn
IMF6 Mode, ZCR, ApEn
IMF7 St. dev, ShEn, RMS, energy, mode, ApEn, LogEn
Total 30 features have been selected by ANOVA test.
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 13/ 19

Box plot visualization of the selected features
COPD-0 COPD-4
0.05
0.1
0.15
COPD-0 COPD-4
6
8
10
12
14
10-3
COPD-0 COPD-4
0.05
0.1
0.15
0.2
COPD-0 COPD-4
0
0.1
0.2
0.3
0.4
COPD-0 COPD-4
0
500
1000
1500
2000
COPD-0 COPD-4
-8
-6
-4
-2
105
COPD-0 COPD-4
0.2
0.4
0.6
COPD-0 COPD-4
0
200
400
600
800
COPD-0 COPD-4
0
0.1
0.2
0.3
0.4
COPD-0 COPD-4
-8
-6
-4
-2
105COPD-0 COPD-4
0
0.1
0.2
0.3
0.4
COPD-0 COPD-4
0.2
0.4
0.6
COPD-0 COPD-4
0.2
0.4
0.6
COPD-0 COPD-4
-8
-6
-4
105
COPD-0 COPD-4
-0.4
-0.2
0
COPD-0 COPD-4
-6
-5
-4
-3
-2
105
COPD-0 COPD-4
-4
-3
-2
105
COPD-0 COPD-4
0.2
0.4
0.6
COPD-0 COPD-4
0
0.05
0.1
0.15
COPD-0 COPD-4
-0.6
-0.4
-0.2
0
COPD-0 COPD-4
0
0.05
0.1
0.15
COPD-0 COPD-4
0.2
0.4
0.6
COPD-0 COPD-4
0
500
1000
COPD-0 COPD-4
0
1000
2000
COPD-0 COPD-4
-8
-6
-4
-2
105
COPD-0 COPD-4
0
2000
4000
COPD-0 COPD-4
0
0.1
0.2
COPD-0 COPD-4
0
0.05
0.1
0.15
COPD-0 COPD-4
0
0.05
0.1
0.15
COPD-0 COPD-4
0.2
0.4
0.6
(i) (ii) (iii) (iv) (v) (vi)
(vii) (xi)
(viii)
(xiii) (xv) (xvii)
(xxiii) (xxiv)
(xxvi) (xxvii) (xxviii)
(xviii)
(xxx)
(xxv)
(ix) (x) (xii)
(xxi)
(xx)
(xix) (xxii)
(xvi)
(xiv)
(xxix)
Figure 10: Illustrate box plots of different features: (i) IMF1 ApEn, (ii) IMF1 ZCR,(iii) IMF1 st. dev,(iv) IMF1 RMS,(v) IMF1
Energy,(vi) IMF1 LogEn,(vii) IMF1 ShEn,(viii) IMF2 LogEn,(ix) IMF2 ApEn,(x) IMF2 ZCR,(xi) IMF3 LogEn,(xii) IMF3
ApEn,(xiii) IMF3 ZCR,(xiv) IMF4 ApEn,(xv) IMF4 LogEn,(xvi) IMF4 ZCR,(xvii) IMF4 Kurtosis,(xviii) IMF5 ZCR,(xix) IMF5
ApEn,(xx) IMF5 LogEn,(xxi) IMF6 mode,(xxii) IMF6 ZCR,(xxiii) IMF6 ApEn,(xxiv) IMF7 RMS,(xxv) IMF7 ApEn,(xxvi) IMF7
energy,(xxvii) IMF7 ShEn,(xxviii) IMF7 LogEn,(xxix) IMF7 mode,(xxx) IMF7 st dev, from COPD-0 and COPD-4 class
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 14/ 19

Classification using machine learning (ML) classifier
ML classifiers:support vector machine (SVM),K- nearest neighbor (KNN),decision
trees (DT),shallow neural network (2 hidden layers having 20, 15 neurons)
Performance metrics: classification accuracy (C a), sensitivity (S e), specificity (S p).
Where Se=tp
tp+fn
Sp=tn
tn+fp
Ca=tp+tn
tp+tn+fp+fn
Table 5: Performance of Different Classifiers Evaluated on Test Set
Classification Scheme Classifiers S e S p C a
COPD-0 versus COPD-4
KNN [10] 0.9375 1 0.9743
SVM [8] 0.9538 0.9607 0.9615
DT [7] 0.8518 0.96 0.923
Shallow NN 0.9259 0.9572 0.949
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 15/ 19

Sensitivity with respect to processing length and
performance comparison
Tested the algorithm with respect to varying processing lengths: 5 sec, 10 sec, and 15 sec
5 10 15
Processing length in sec
0.6
0.7
0.8
0.9
1
Accuracy
SVM KNN Shallow NN
Figure 11: Effect of processing length with respect to accuracy
Table 6: Performance Comparison with Other Existing Works
Sl.
No.
Work
reference
Data
used
Proposed
technique
C a
(%)
S e
(%)
S p
(%)
1 Altan et al. [6] Lung
sound 3D-SODP based features, DBN 95.84 93.65 93.34
2 Proposed Lung
Sound
VMD based feature
extraction and ML classifier 97.43 93.75 100
3D-SODP: 3-dimensional second-order difference plot, DBN: deep belief networks
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 16/ 19

Conclusion
Establishes the fact that lung sounds play an important role in COPD severity detection
This shows importance towards the early detection of COPD due to the incorpopration of
COPD0 class
This work highlights the feasibility of automatic COPD severity detection using machine
learning while emanating complex deep learning
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 17/ 19

References
References I
A. A. Cruz, ”Global surveillance, prevention and control of chronic respiratory diseases: a comprehensive approach”,
World Health Organization,2007.
Salvi, Sundeep, et al. ”The burden of chronic respiratory diseases and their heterogeneity across the states of India: the
Global Burden of Disease Study 1990–2016.” The Lancet Global Health 6.12 (2018): e1363-e1374.
ALTAN, G¨okhan, Yakup Kutlu, Yusuf Garbi, Adnan ¨
Ozhan Pekmezci, and Serkan Nural. ”Multimedia respiratory
database (RespiratoryDatabase@ TR): Auscultation sounds and chest X-rays.” Natural and Engineering Sciences 2, no. 3
(2017): 59-72.
Satija, U., Ramkumar, B. and Manikandan, M.S., 2017. Real-time signal quality-aware ECG telemetry system for
IoT-based health care monitoring. IEEE Internet of Things Journal, 4(3), pp.815-823.
K. Dragomiretskiy and D. Zosso, ”Variational Mode Decomposition,” in IEEE Transactions on Signal Processing, vol. 62,
no. 3, pp. 531-544, Feb.1, 2014.
Altan G, Kutlu Y, Pekmezci A¨
O, Nural S, ”Deep learning with 3D-second order difference plot on respiratory sounds.”
Biomedical Signal Processing and Control 45 (2018): 58-69.
S. R. Safavian and D. Landgrebe, ”A survey of decision tree classifier methodology,” in IEEE Transactions on Systems,
Man, and Cybernetics, vol. 21, no. 3, pp. 660-674, May-June 1991.
Kamruzzaman, Joarder Begg, Rezaul. (2007). ”Support Vector Machines and Other Pattern Recognition Approaches to
the Diagnosis of Cerebral Palsy Gait.” IEEE Transactions on Biomedical Engineering, 53. 2479 - 2490.
G. Sahu and K. C. Ray, ”An Efficient Method for Detection and Localization of Myocardial Infarction,” in IEEE
Transactions on Instrumentation and Measurement, vol. 71, pp. 1-12, 2022.
Cunningham, P., Delany, S. J.. ”K-nearest neighbor classifiers-a tutorial” ACM Computing Surveys (CSUR).
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 18/ 19

References
Thank You
Indicon 2022, PID: 421, Roy et al. Severity Detection Of COPD by lung sounds November 20, 2022 19/ 19