Automatic Wheeze Detection and Lung Function Evaluation

A Preliminary Study

Ana Oliveira

, Cátia Pinho

1,2

, João Dinis

1,2

, Daniela Oliveira

and Alda Marques

School of Health Sciences, University of Aveiro (ESSUA), Campus Universitário de Santiago, Aveiro, Portugal

Institute of Electronics and Telematics Engineering of Aveiro (IEETA), University of Aveiro,

Campus Universitário de Santiago, Aveiro, Portugal

Keywords: Time Frequency Analysis, Wheezing, Lung Function Testing, Lower Respiratory Tract Infection.

Abstract: The automatic detection of wheeze offers the potential for diagnosing and monitoring respiratory diseases,

e.g., lower respiratory tract infection (LRTI). By determining the relationship between wheeze detection and

other lung function data, it is possible to develop a more sensitive tool for detecting respiratory conditions.

This pilot study aimed to: i) explore the robustness of a time frequency wheeze detector (TF-WD) and ii)

describe the correlation between wheezing and spirometry parameters. Lung sounds and spirometry parame-

ters were acquired from six outpatients with LRTI (five with right lung infection). Number, fundamental

frequency and duration of wheezes were obtained through a TF-WD algorithm. The performance of the TF-

WD algorithm was evaluated by comparing its findings in 40 files with those annotated by two experts. Re-

sults suggest that the TF-WD algorithm is an efﬁcient and robust method for computerised wheeze detection

in LRTI (SE=72.5%; SP=99.2%). Furthermore, significant correlations were found between the percentage

predicted of forced expiratory volume in 1 second and forced vital capacity (FEV

pp and FVCpp) and

wheeze duration at lateral (rs=-0.9, p=0.03) and posterior (rs=-0.9, p=0.01) right regions respectively. These

results support the use of pulmonary auscultation and spirometry to detect areas of obstruction in LRTI.

1 INTRODUCTION

Lower respiratory tract infection (LRTI) are among

the most common infectious diseases, with an annu-

al incidence of approximately 429 million cases

worldwide (World Health Organization, 2008).

Currently, health professionals use respiratory

function tests, such as spirometry and standard pul-

monary auscultation to diagnose and monitor pa-

tients with these respiratory conditions (Marques et

al., 2006). However, standard auscultation has been

reported as a subjective process (Sovijärvi et al.,

2000) and therefore, many research efforts are being

conducted to automatically detect, quantify and

characterise respiratory sounds (Dinis et al., 2012).

Wheezes have been the most common type of

ALS investigated for diagnostic purposes, using the

stethoscope (Earis and Cheetham, 2000). They are

clinically defined as musical sounds characterised by

their location, intensity, pitch (frequencies above

100Hz) and duration (longer than 100 ms) (Sovijärvi

et al., 2000). These respiratory sounds can be classi-

fied as monophonic or polyphonic, and are mainly

associated with diseases that structurally involve the

narrowing of airway calibre such as bronchospasm

or airway obstruction (Meslier et al., 1995). Their

presence have proved to have a significant contribu-

tion in the diagnosis and monitoring of LRTI (Paciej

et al., 2004). Taplidou and Hadjileontiadis (2007)

wheeze detection algorithm has been shown to be

reliable and valid for cystic fibrosis (Oliveira et al.,

2011), however it requires further validation for

different respiratory diseases.

It has been suggested that the combination of spi-

rometry with computer aided lung sound analysis

increases the sensitivity for detecting early signs of

respiratory diseases (Marques et al., 2009), however

this fact needs to be investigated.

This preliminary study aimed to explore the ro-

bustness of a TF-WD algorithm and describe the

correlation between wheezing and spirometry pa-

rameters in patients with LRTI.

323

Oliveira A., Pinho C., Dinis J., Oliveira D. and Marques A..

Automatic Wheeze Detection and Lung Function Evaluation - A Preliminary Study.

DOI: 10.5220/0004191903230326

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2013), pages 323-326

ISBN: 978-989-8565-37-2

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

2 METHODS

2.1 Participants

Participants were eligible for the study if they pre-

sented at the emergency department of the Hospital

Infante D. Pedro, (Aveiro, Portugal) with cough and

at least one of the following symptoms: sputum,

dyspnoea, wheezes or chest pain, according to the

guidelines for the diagnosis of LRTI (Woodhead et

al., 2011) and were ≥18 years old. Ethical approval

was obtained from the Ethics Committee of Hospital

Infante D. Pedro. All participants gave written in-

formed consent prior to any data collection.

2.2 Data Collection

Data were collected by two researchers in a clinical

setting within 24 hours of hospital presentation.

Socio-demographics and anthropometric data were

collected first. Then, respiratory sound recordings

for short-term acquisitions were performed accord-

ing to the Computerized Respiratory Sound Analysis

(CORSA) guidelines (Rossi et al., 2000), i.e., partic-

ipants were in a seated-upright position and lung

sound data was collected with a digital stethoscope

(Thinklabs® Rhythm: ds32a, Colorado, US) in sev-

en chest locations (trachea, left and right: anterior,

lateral and posterior). Respiratory sounds were rec-

orded three times for each location with 25 seconds

duration each to assure that 7-10 respiratory cycles

were recorded (Rossi et al., 2000). Finally, forced

expiratory volume in 1 second (FEV

), forced vital

capacity (FVC) and peak expiratory flow (PEF)

were acquired with the spirometer MicroLab Micro

Medical 36-ML3500-MK8, UK, following the Eu-

ropean Respiratory Society guidelines (Miller,

2005).

Patients received standard medical treatment for

LRTI and after three weeks, the time taken to recov-

er from a LRTI (Woodhead et al., 2011), all the pre

intervention measurements were repeated.

2.3 Wheeze Detection

Wheezes were detected using the algorithm de-

scribed by Taplidou and Hadjileontiadis (2007). This

algorithm is based on a time-frequency analysis

technique: the Short-time Fourier transform (STFT),

proposed by Gabor (1946). In the implemented algo-

rithm, the signal was digitally filtered (band pass

60–2100Hz, order-8 Butterworth) and resampled (to

5512s

-1

) before the STFT calculation. To remove

noise from the signal, a smoothing procedure based

on box filtering, also known as mean-filtering, was

applied. Peaks higher than a specific magnitude

threshold were then selected and classified as

wheezes or non-wheezes, according to a set of crite-

ria that includes: local maxima, peak coexistence

and continuity in time (Taplidou and

Hadjileontiadis, 2007). For each wheeze duration

and fundamental frequency were obtained.

2.4 Algorithm Robustness Evaluation

Sounds annotation by respiratory experts is the most

common and reliable method to assess the robust-

ness of algorithms to detect ALS (Guntupalli et al.,

2008). Nevertheless, annotation is a time–consuming

process, being difficult to conduct in a large amount

of sound files. To overcome this difficulty, 40 res-

piratory sound files were randomly selected from a

total of 252, through a simple randomisation. Two

respiratory physiotherapists, with experience in vis-

ual-auditory wheeze recognition, independently

annotated the selected sound files in terms of pres-

ence, number and duration of wheezes. For the an-

notation, the Respiratory Sound Annotation Soft-

ware v1.1 was used (Dinis et al., 2012).

2.5 Statistical Analysis

The robustness evaluation of the algorithm was ob-

tained by comparing its wheeze detection with the

assessment of the two expert respiratory physiother-

apists. Cohen’s kappa coefficient was used to assess

the inter-rater agreement between the physiothera-

pists. The performance of the TF-WD algorithm was

calculated through the sensitivity (SE) and specifici-

ty (SP) of the algorithm. True positives/negatives

and false positives/negatives were counted by com-

paring each point of the sound file.

The univariate relationships between gender and

the participants’ individual characteristics were ex-

amined by the Mann–Whitney U-test. Differences

between calculated parameters in the first (base) and

the second acquisition (post) were explored with

paired sample t-test or Wilcoxon Signed Ranks test

when the data did not follow a normal distribution.

Spearman's correlation coefficient was used to

correlate number, duration and fundamental fre-

quency of wheezes with spirometry parameters.

Data were expressed as number, mean and stand-

ard deviation (Mean±SD). Analysis was performed

using PASW® Statistics 18.0 software (SPSS Inc,

Chicago, IL, USA) and Matlab®R2009a (The

MathWorks, Inc, Natick, MA, USA). Significance

level was set at p<0.05.

HEALTHINF2013-InternationalConferenceonHealthInformatics

324

3 RESULTS

The proportion of agreement between the two physi-

otherapists was very good (k=0.804). The SE of the

automated method was 99.2%, SP was 72.5% and

the performance was 84.8%. The total number of

wheezes detected by the algorithm was 37 at base-

line and 89 at post treatment.

Six participants (three males) diagnosed with

LRTI enrolled in this pilot study, five of whom pre-

sented the focus of the infection in the right lung.

Their characteristics are summarised in table 1. The

sample at baseline was generally homogeneous.

Table 1: Sample characterisation.

Variables Base Post p

Patients (n) 6 6 -

Age (years) 44.5 ± 4.5 44.5 ± 11.1 -

Height (m) 1.6 ± 0.0 1.6 ± 0.1 -

Weight (Kg) 79.5 ± 8.1 79.7 ± 8.5 0.92

BMI (Kg/m

) 29.3 ± 2.6 29.3 ± 2.7 0.92

FEV

pp 84.0 ± 3.9 101.2 ± 5.0 0.01*

FVCpp 84.7 ± 5.7 99.3 ± 5.6 0.05

PEFpp 50.7 ± 5.0 78.7 ± 8.9 0.01*

Nº of wheeze 6.0 ± 0.9 14.8 ± 2.3 0.03*

Dt of wheeze (s) 0.21 ± 0.0 0.22 ± 0.0 0.63

Fq of wheeze (Hz) 365 ± 37.0 363 ± 29.1 0.92

BMI: body mass index; FEV

pp: percentage predicted of forced

expiratory volume in 1 second; FVCpp: percentage predicted of

forced vital capacity; PEFpp: percentage predicted of peak expir-

atory flow; nº: number; Dt: duration; Fq: frequency. Results are

shown in mean ±standard deviation.

*p < 0.05.

The lung function evaluation showed no differ-

ences between genders and all participants presented

a statistically significant increase in FEV

(p=0.01), PEFpp (p=0.01) and in the number of

wheezes (p=0.03), from the baseline to post treat-

ment (table 1).

Table 2: Correlation between the duration of wheeze and

lung function parameters at post treatment.

Wheeze FEV

pp FVCpp PEFpp

Duration

rs p* rs p* rs p*

Tc 0.1 0.87 0.5 0.32 -0.8 0.06

Ar 0.8 0.14 0.6 0.29 0.1 0.94

Al 0.6 0.28 0.2 0.74 0.2 0.74

Lr -0.9 0.04* -0.5 0.39 0.1 0.94

Ll 0.2 0.75 -0.1 0.87 -0.5 0.40

Pr -0.6 0.2 -0.9 0.02* 0.8 0.07

Pl 0.0 1 0.3 0.62 0.2 0.75

Tc: trachea; Ar: anterior right; Al: anterior left; Lr: lateral right ; Ll:

lateral left; Pr: posterior right; Pl: posterior left; rs: correlation coeffi-

cient.

* p < 0.05.

The results obtained from the wheeze evaluation

(number, duration and fundamental frequency) of

the 252 sound files were not significant correlated

with spirometry parameters (FEV

pp, FVCpp and

PEFpp) at the baseline assessment. However, signif-

icant correlations were found post treatment between

FEV

pp and wheeze duration at lateral right region

(rs=-0.9, p=0.04), and between FVCpp and wheeze

duration at posterior right region (r

s=-0.9, p=0.01)

(table 2).

4 DISCUSSION

This preliminary study demonstrates the robustness

of the proposed TF-WD algorithm for computerised

wheeze detection in LRTI. Since wheezing are often

present in patients with LRTI (Woodhead et al.,

2011), this technique can constitute a valuable and

practical non-invasive tool for diagnosing and moni-

toring patients with this respiratory conditions. The

overall results of the algorithm (SE=72.5%;

SP=99.2%) showed similar performance in the sen-

sitivity and specificity of the wheeze detection com-

pared to previous studies (Taplidou and

Hadjileontiadis, 2007); (Oliveira et al., 2011).

A reduced number of wheezes were observed at

baseline, probably due to the presence of air or fluid

in/or around the lung, causing a decrease in number

of breath sounds (Lieberman et al., 2002). This re-

duced number of wheezing may have contributed to

the absence of baseline correlations. Furthermore, a

significant increase of wheezing post treatment was

observed (p=0.03).

Significant and strong negative correlations were

found in the post treatment assessment between: i)

FEV

pp and duration of wheeze at lateral right re-

gion, and ii) FVCpp and duration of wheeze at pos-

terior right region. These results suggest that by

combining lung sounds with spirometry it may be

possible to predict and assess the location of the

obstruction with more accuracy.

Oud et al. (2000) found that 60-90% of lung

sound data can classify FEV

-values by using com-

puted spectral sound data. Leuppi et al. (2006), in a

study conducted in the emergency setting, reported

that airways obstruction may often be overestimated

by auscultation and when combining auscultation

with spirometry the accuracy of the diagnosis could

be increased by approximately 8%.

Additionally to previous research, our study pro-

poses a further analysis of wheezes and lung func-

tion data, with the evaluation of more than one loca-

tion of auscultation and its correlation with spirome-

AutomaticWheezeDetectionandLungFunctionEvaluation-APreliminaryStudy

325

try. This correlation gave relevant information to

detect respiratory obstructed areas.

As verified only the right areas of the lung had

correlation with the wheezes parameters, which can

be explained by the fact of the majority of partici-

pants enrolled (n=5) presented with right lung infec-

tion. Nevertheless, the small sample size (6 partici-

pants with LRTI), limits and decrease the statistical

power and may have polarised the results.

Currently, there is a lack of published data as-

sessing the correlation between wheeze and spi-

rometry parameters and therefore, it is believed that

these findings provide a significant contribution for

research and clinical practice.

5 CONCLUSIONS

This study suggests that the TF-WD algorithm is a

robust method for computerised wheeze detection in

patients with LRTI. Furthermore, the use of compu-

terised auscultation and spirometry as outcome

measures to detect the area of obstruction in patients

with this respiratory condition is also supported.

However, further studies with larger samples are

needed to fully confirm the presented results.

ACKNOWLEDGEMENTS

The authors gratefully acknowledge the funding

provided to the project “Adventitious lung sounds as

indicators of severity and recovery of lung pathology

and sputum location- PTDC/SAU-

BEB/101943/2008” by Fundação para a Ciência e a

Tecnologia (FCT). The authors would also like to

thank to Vânia Rocha for contributing to the annota-

tion process.

REFERENCES

Dinis, J., Campos, G., Rodrigues, J. & Marques, A. Year.

Respiratory Sound Annotation Software. In:

International Conference on Health Informatics, 2012

Vilamoura, Portugal. Proceedings of HEALTHINF

2012 - International Conference on Health

Informatics, 183-188.

Earis, J. & Cheetham, B. 2000. Future perspectives for

respiratory sound research. European Respiratory

Review, 10 641-646.

Gabor, D. 1946. Theory of communication. IEEE, 93,

429-457.

Guntupalli, K. K., Alapat, P. M., Bandi, V. D. & Kushnir,

I. 2008. Validation of automatic wheeze detection in

patients with obstructed airways and in healthy

subjects. Journal of asthma, 45, 903-7.

Leuppi, J. D., Dieterle, T., Wildeisen, I., Martina, B.,

Tamm, M., Koch, G., Perruchoud, A. P. &

Leimenstoll, B. M. 2006. Can airway obstruction be

estimated by lung auscultation in an emergency room

setting? Respiratory Medicine, 100, 279-285.

Lieberman, D., Korsonsky, I., Ben-Yaakov, M.,

Lazarovich, Z., Friedman, M. G., Dvoskin, B.,

Leinonen, M., Ohana, B. & Boldur, I. 2002. A

comparative study of the etiology of adult upper and

lower respiratory tract infections in the community.

Diagnostic microbiology and infectious disease, 42,

21-8.

Marques, A., Bruton, A. & Barney, A. 2006. Clinically

useful outcome measures for physiotherapy airway

clearance techniques: a review. Physical Therapy

Reviews, 11, 299-307.

Marques, A., Bruton, A. & Barney, A. 2009. Reliability of

lung crackle characteristics in cystic fibrosis and

bronchiectasis patients in a clinical setting.

Physiological Measurement, 30, 903-912.

Meslier, N., Charbonneau, G. & Racineux, J. L. 1995.

Wheezes. European respiratory journal, 8, 1942-8.

Miller, M. R. 2005. Standardisation of spirometry.

European Respiratory Journal, 26, 319-338.

Oliveira, D., Pinho, C., Marques, A. & Dinis, J. 2011.

Validation of a time-frequency wheeze detector in

cystic fibrosis: a pilot study. European Respiratory

Journal, 38, 237s.

Oud, M., Dooijes, E. H. & van der Zee, J. S. 2000.

Asthmatic airways obstruction assessment based on

detailed analysis of respiratory sound spectra. IEEE

transactions on bio-medical engineering, 47

, 1450-5.

Paciej, R., Vyshedskiy, A., Bana, D. & Murphy, R. 2004.

Squawks in pneumonia. Thorax, 59, 177-178.

Rossi, M., Sovijärvi, A. R. A., Piirilä, P., Vannuccini, L.,

Dalmasso, F. & Vanderschoot, J. 2000. Environmental

and subject conditions and breathing manoeuvres for

respiratory sound recordings. European Respiratory

Review, 10, 611-615.

Sovijärvi, A., Dalmasso, F., Vanderschoot, J., Malmberg,

L., Righini, G. & Stoneman, S. 2000. Definition of

terms for applications of respiratory sounds. European

Respiratory Review, 10, 597-610.

Taplidou, S. A. & Hadjileontiadis, L. J. 2007. Wheeze

detection based on time-frequency analysis of breath

sounds. Computers in Biology and Medicine, 37,

1073-1083.

Woodhead, M., Blasi, F., Ewig, S., Garau, J., Huchon, G.,

Ieven, M., Ortqvist, A., Schaberg, T., Torres, A., van

der Heijden, G., Read, R. & Verheij, T. J. 2011.

Guidelines for the management of adult lower

respiratory tract infections--full version. Clinical

Microbiology and Infection, 17 Suppl 6, E1-59.

World Health Organization 2008. The global burden of

disease - 2004 update Switzerland: World Health

Organization.

HEALTHINF2013-InternationalConferenceonHealthInformatics

326