bs_bs_banner
Editor's Choice
ORIGINAL ARTICLE
Novel algorithm to identify and differentiate specific digital signature of breath sound in patients with diffuse parenchymal lung disease PARTHASARATHI BHATTACHARYYA,1 ASHOK MONDAL,2 RANA DEY,1 DIPANJAN SAHA1 AND GOUTAM SAHA2 1
Institute of Pulmocare and Research, Kolkata, and 2Department of Electronics and Electrical Communication Engineering, Indian Institute of Technology, Kharagpur, India
ABSTRACT Background and objective: Auscultation is an important part of the clinical examination of different lung diseases. Objective analysis of lung sounds based on underlying characteristics and its subsequent automatic interpretations may help a clinical practice. Methods: We collected the breath sounds from 8 normal subjects and 20 diffuse parenchymal lung disease (DPLD) patients using a newly developed instrument and then filtered off the heart sounds using a novel technology. The collected sounds were thereafter analysed digitally on several characteristics as dynamical complexity, texture information and regularity index to find and define their unique digital signatures for differentiating normality and abnormality. For convenience of testing, these characteristic signatures of normal and DPLD lung sounds were transformed into coloured visual representations. The predictive power of these images has been validated by six independent observers that include three physicians. Results: The proposed method gives a classification accuracy of 100% for composite features for both the normal as well as lung sound signals from DPLD patients. When tested by independent observers on the visually transformed images, the positive predictive value to diagnose the normality and DPLD remained 100%. Conclusions: The lung sounds from the normal and DPLD subjects could be differentiated and expressed according to their digital signatures. On visual transformation to coloured images, they retain 100% predictive power. This technique may assist physicians to diagnose DPLD from visual images bearing the digital signature of the condition. Key words: diffuse parenchymal lung disease, digital signature, lung fibrosis, pulmonary fibrosis, X-ray. Correspondence: Parthasarathi Bhattacharyya, Institute of Pulmocare and Research, Kolkata.CB 16, Salt Lake, Kolkata 700 064, India. Email:
[email protected] Received 8 April 2014; invited to revise 4 August 2014; revised 17 September 2014; accepted 15 October 2014 (Associate Editor: Yuben Moodley). © 2015 Asian Pacific Society of Respirology
SUMMARY AT A GLANCE We developed an instrument and technology to capture lung sounds and filter off the admixed heart sounds. We utilized lung sounds from 8 controls and 20 patients with diffuse parenchymal lung disease to identify the accurate digital signature of normal and interstitial lung sounds and transform them into characteristic visual images.
Abbreviations: DAS, data acquisition system; DPLD, diffuse parenchymal lung disease; EMD, empirical mode decomposition.
INTRODUCTION Traditionally, a physician’s exercise to diagnose a lung disease starts with history taking and clinical examination. Auscultation forms an important component of clinical examination and it has remained a widely used practice for nearly two centuries since its introduction by Laennec in 1816.1 While it allows a subjective interpretation of the intensity and quality of the transmitted lung sounds, the success depends largely upon the clinician’s experience. Although attempts have been made to develop a technology to assist clinicians through capturing lung sounds and their subsequent qualitative identification, the translation such efforts into clinical practice has largely remained elusive. None of the proposed developments could meet the demand of being simple, easily interpretable (preferably visible), objective, pure (specific) and yet highly predictive of normal and interstitial lung sounds. Beyond the traditional phonopneumography, several developments employed advancement in the computer technologies, statistical signal processing and machinelearning algorithms. Several methods based on time, frequency and the analyses of time–frequency domain2–16 have been attempted earlier, but none of them could be reasonably successful. Respirology (2015) 20, 633–639 doi: 10.1111/resp.12529
634
P Bhattacharyya et al.
Since the demand of a tool to assist clinical practice through objective analysis and automatic interpretation of lung sounds appears high, we attempted to develop an instrument and algorithm to serve the purpose. In this, normal and pathological lung sounds were collected from control and diffuse parenchymal lung disease (DPLD) subjects. The digital signatures of both of these sounds were determined using their morphological features in terms of its skewness, texture information and complexity index. We then applied a visual transform of these signatures to generate images easily interpretable by any person.
Microphone (Audio band 25 kHz)
Microphone pre-amplifier
Low pass filter with cut off frequency of 10 kHz
Medium Thin
Low pass filter with cut off frequency of 2 kHz Fatty
METHODS The project was approved by the institutional ethics committee of the Institute of Pulmocare and Research, and signed informed consent was obtained from all participants.
Preparing the data acquisition system A single-channel data acquisition system (DAS) was constructed by circuiting the active electronic devices (transistors, operational amplifiers) and passive elements (resistors, inductors and capacitors). This was fitted to a stethoscope to capture lung sound using the diaphragm mode. The DAS has two output connections: one connected to stethoscope through a microphone transducer and the other being connected to the storage device to display lung sound, as well as a power supply. A functional block diagram of the DAS system has been described in17 and represented by Figure 1. The output of the stethoscope is connected to a transistor-based pre-amplifier and passed through two successive op-amp-based Butterworth filters and a frequency gain amplifier. The outcome of the DAS system is stored in a buffer. It can be replayed via a speaker or visualized in a display device. It can be subjected to digital data analysis. Collection of lung sounds from normal and DPLD subjects The lung sound signals were recorded at the Institute of Pulmocare and Research, Kolkata, in both healthy control and DPLD participants. The whole group was evaluated with clinical examination, spirometry and chest x-ray. The DPLD group also had high-resolution computed tomography. We preferred a tracheal position for lung sound collection to standardize data acquisition and to reduce the level of artefacts. The acquired data were stored as .wav files in 16 bit, pulse code modulation, mono audio format at sampling frequency of 8 kHz. The whole analysis was implemented on a standard PC. The MATLAB (R2008a, The Mathworks, Inc., Natick, MA, USA) tool was used for conducting the all experiments. Analysis of the morphological characteristics of lung sound signals and identification of characteristic digital signatures The method included analysis of the morphological characteristics of lung sound signals to discriminaRespirology (2015) 20, 633–639
Gain selector
Frequency gain amplifier (0-2 kHz)
Audio amplifier
Buffer
Computer sound card or external sound card
Speaker Signal out for display and storing
Figure 1
Block diagram of the data acquisition system.
Figure 2
Different stages of decision-making system.
tion of control from DPLD subjects. A block diagram (see Fig. 2) represents the steps that starts with processing the raw lung sound data to obtain improved signal-specific information and filtering off the admixed heart sound (de-noising) followed by extraction of features from the enhanced signals. This was done through using the collected sound by forming a training set and then validating it by a disjointed test set, for individual features as well as one composite feature. Finally, a visual transformation was done on composite features, and those images were tested by blinded observers. In the de-noising phase, we eliminated the heart sound from lung sound signals using a novel algorithm based on an empirical mode decomposition (EMD) technique developed by us,17 for which an Indian patent (Ref. No. 515/KOL/2011) has been filed. The results of the EMD-based technique are shown in Figure 3. Next to de-noising, in the stage of feature extraction for determination of the digital signatures of normal © 2015 Asian Pacific Society of Respirology
Amplitude
Amplitude
Amplitude
Amplitude
© 2015 Asian Pacific Society of Respirology
−1
0
1
−1
0
1
−1
0
1
−1
0
1
−1
0
0
0
0
0
0
0.5
0.5
0.5
0.5
0.5
1
1
1
1
1
1.5
1.5
1.5
1.5
1.5
Sample Number
2
(e) Residual HS signal
2 Sample Number
(d) Reconstructed LS signal
2 Sample Number
(c) Mixed signal (80% LS & 20% HS)
2 Sample Number
(b) Mixed signal (50% LS & 50% HS)
2 Sample Number
(a) Mixed signal (20% LS & 80% HS)
2.5
2.5
2.5
2.5
2.5
3
3
3
3
3
3.5
3.5
3.5
3.5
3.5 x 10
4
4
x 10
x 10
4 4
x 10
4
4
4
4
4
x 10
4
Figure 3 Separated lung sound (LS) signal and heart sound (HS) signal using our proposed method based on empirical mode decomposition (EMD) technique:17 (a),(b) and (c) LS and HS mixed with proportions 20:80, 50:50, 80:20 respectively; (d) Reconstructed LS signal, and (e) Residual HS signal for case (a) where HS interference is maximum (similar separation result for other two cases).
Amplitude
1
Digital signature of lung sound in DPLD
635
Respirology (2015) 20, 633–639
636
P Bhattacharyya et al.
and abnormal lung sounds, we chose the statistical parameters to measure the morphological characteristics of the normal as well as pathological signals in terms of three parameters. They are (i) skewness (ε), (ii) lacunarity (ς), and (iii) sample entropy (α). The skewness index (a measurement of asymmetry) is derived from the probability density function that measures how frequently a signal parameter changes in the sample space. Mathematically, it is defined as a ratio given by Equation 1.
skewness ( ε ) =
E [( y (n ) − μ )3 ]
( E [ y (n ) − μ)2 ])
1.5
(1)
where y (n) is the lung sound signal, μ represents the mean of the signal amplitude distribution. The lacunarity (ς) parameter that measures the texture information of any visual object was calculated using a gliding box algorithm.18 The sample entropy (SampEnt) measures the complexity of a signal.19 It is defined as the negative logarithm of an estimate of the conditional probability. Conditional probability refers to the occurrence of one with respect to the occurrence of the other. In this, the two states that are matched point wise for a dimension m within a matching tolerance ‘r’ remain matched in dimension m+1 and is expressed as:
m +1 ⎡ C (r ) ⎤ SampEnt (m, r , N ) = − ln ⎢ m ⎣ C (r ) ⎥⎦
(2)
where Cm+1 (r) is the probability that two templates will match for m + 1 points. We observed that these three features give distinctive values to differentiate normal and DPLD lung sounds. Validation was performed by dividing the data in training and test set as explained in Results section.
Conversion of the digital signature to visual displays of normal and abnormal lung sounds For a visual display of the digital signatures of the sounds, we used transformation to convert them to an image format using a three-dimensional surface plot algorithm (Fig. 4). The display of the inherent morphological properties of respiratory sound looks like a sheet of paper with different colours (red, green, blue and pink) being sprayed over it in different fashions at different places in different proportions. The picture characteristics are used for normal and DPLD identification. In control subjects, the shape appears folded transversely with the acute angle looking up, while in DPLD subjects, it appeared looking down to make a tent-like shape. In addition, the upper and right lateral point of the display touches the right
Figure 4 The visual display of normal and the abnormal lung sounds with differentiating features have been elaborated. The upper row (a) depicts normal and lower row (b) diffuse parenchymal lung disease (DPLD) lung sounds. Respirology (2015) 20, 633–639
© 2015 Asian Pacific Society of Respirology
637
Digital signature of lung sound in DPLD
border above a blue bar in control but below the bar in DPLD subjects. The distribution of colours also appeared different in the two groups; the blue colour crosses the fold in DPLD subjects in the form of making the boarder of a wedge shape where the tip of the wedge falls on the fold.
RESULTS In the study, the clinician’s diagnosis with standard evaluations were taken as the ‘gold standard’, and the experimental results were validated by comparing the outcomes with that clinical diagnosis. The effectiveness of individual features (skewness, lacunarity, sample entropy) and composite feature (incorporating all these three features together) were evaluated in terms of sensitivity, specificity and accuracy. A total of 28 subjects are used in the study that consists of 8 normal and 20 DPLD subjects. In five rounds of testing where, in each round, randomly chosen 50% of data was used for training and 50% was used for testing. Each of 28 subjects became part of test data in one or more rounds. The final result was obtained by averaging the results of these five rounds. Tables 1 and 2 provide a quantitative comparison of strengths of individual features as well as the composite feature. It shows that individual features carry important information about lung status. The composite feature gives the best result with an accuracy of 100%. For visual examination-based classification, six mutually exclusive sets of data were prepared from test data set. Each set consisted of four subjects with both normal and DPLD cases. These sets of pictures were examined by three blinded physicians and three blinded non-professionals. They were asked to mark the pictures independently as normal or abnormal according to the understanding from the visually transformed digital signature characteristics presented in Figure 4. The result (Table 3) presents predictive power. The visual test also gives an accuracy of 100% for both normal and abnormal subjects.
DISCUSSION Respiratory sounds are complex to analyse due to their generation from vibration of the airway walls and turbulent flow of air through the conducting system. The morphological complexity of lung sounds alters with the pathological conditions of the lungs. It appears that our method of determining the distinctive features of normal and abnormal lung sounds using the morphological parameters has been successful in differentiating normal subjects from DPLD patients. We used skewness, lacunarity and sample entropy to represent embedded morphological characteristics of lung sound. These, being statistical domain features, are robust in nature. The composite feature gives higher sensitivity, specificity and accuracy compared with individual features probably because of higher information content. The accuracy of 100% when tested on 8 normal and 20 DPLD subjects shows the strength of this method and the potential for adoption into clinical practice. In this work, we have chosen tracheal sound for the experiments because it provides several advantages as it is easy to collect and shows higher amplitude, intensity and frequency components compared with other respiratory sounds.9 In addition, tracheal sound also contains additional information about the airflow signal and is less affected by artefacts from clothes or hair which produce interferences. A major problem in the lung sound analysis is the interference of heart sounds that may obscure the lung sound information and lead to misinterpretation. We addressed this by considering: (i) EMD-based technology to filter out the associated heart sounds17; (ii) lower degree of variability in data acquisition by focusing on tracheal sounds; and (iii) provision of adjustment parameters for recording such as gain selection, zooming up and audio-video playing of the recorded sounds etc. The visual representations of normal and abnormal sounds seem useful, easy and user friendly. The digital signatures, transformed into
Table 1 Performance of the proposed method tested on 8 normal subjects Feature Skewness Lacunarity SampEnt Composite
Sensitivity (%)
Specificity (%)
Accuracy (%)
85.71 85.71 95.86 100
80.00 87.50 92.24 100
84.21 86.11 93.88 100
Table 2 Performance of the proposed method tested on 20 abnormal patients with DPLD diseases Feature Skewness Lacunarity SampEnt Composite
Sensitivity (%)
Specificity (%)
Accuracy (%)
82.33 82.56 92.86 100
83.47 87.50 95.00 100
82.50 83.33 93.75 100
© 2015 Asian Pacific Society of Respirology
Respirology (2015) 20, 633–639
638 Table 3
P Bhattacharyya et al. Testing results of visual displays (mixed with normal and abnormal) Decision
Observers Physician 1 Physician 2 Physician 3 Layman 1 Layman 2 Layman 3
Digital signatures Shape LETB Shape LETB Shape LETB Shape LETB Shape LETB Shape LETB
Set 1 (4)
Set 2 (4)
Set 3 (4)
Set 4 (4)
Set 5 (4)
Set 6 (4)
Predictive power (%)
4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4
4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4
4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4
4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4
4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4
4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4 4/4
100 100 100 100 100 100 100 100 100 100 100 100
The results of the response of three physicians and three untrained individuals on six different sets of visual displays of lung sounds are elaborated in terms of predictive power. Shape refers to overall shape of the image; LETB = lateral edge touching border characteristics of the image. Each set consists of four subjects including normal and abnormal. A total of 18 abnormal and 6 normal subjects have been used for this exercise from test data set. The predictive power is the ratio of correctly identified subjects and total participating subjects.
visual images, retained the same predictive power when presented to a group of physicians and untrained people. The journey to understand the lung sounds with technology is long. The literatures record and list several attempts as using the Mahalanobis distance,2 autoregressive coefficients, k-nearest neighbour and quadratic classifier,3 amplitude and frequency parameters with regression line slopes,4 neural networkbased classification technique using subphase features like autoregressive coefficients, prediction error and the ratio of expiration and inspiration duration,5 stochastic classification using time and frequency domain features,6 maximum likelihood approach7 and classifier for classification of normal and crackle sounds8 with host of other efforts. They include work with neural network classifier and wavelet domain features,9,10 computer-aided system,11 pattern recognition-based model matching algorithm12 and a multisensor breath sound mapping device,13 spectral analysis method14 and also a vibration response imaging-based method.15 Computer technology has been in use for this purpose too. A 16-channel computer-aided system is reported to discriminate certain types of crackles.16 Other personal computer-based methods for lung sound analysis primarily emphasizes on modern signal processing and signal machine learning techniques.20–23 A meta-analysis in 2010 revealed that the application of Fourier transform and neural network algorithms on lung sound achieved an overall sensitivity of 80% (95% confidence interval (CI): 72–86%) and specificity of 85% (95% CI: 78–91%) in detecting wheezes or crackles.24 The weakness of our effort has been that the suitability of this technique to differentiate other conditions such as bronchiectasis or disease states with crackles such as congestive heart failure has not been tested. Further studies in this area and also in the field Respirology (2015) 20, 633–639
of obstructive lung diseases, pleural and mediastinal diseases, remains a promising area to explore in future. In conclusion, we presented an innovative method of capturing the distinctive digital signatures of lung sounds of control subjects and participants with DPLD. We further transformed these signatures to characteristic images. This technique may be implemented to build an instrument that can assist medical students to acquire clinical acumen quickly, primary healthcare workers to identify abnormality in remote and resource constrained situations, and physicians to diagnose DPLD subjects easily.
REFERENCES 1 Laennec R. Invention of the stethoscope. In: R. Bruce Lindsay (ed.) Acoustics: Historical and Philosophical Development. Dowden, Hutchinson and Ross Inc, Stroudsburg, Pennsylvania, 1973; 166–71. 2 Cohen A, Landsberg D. Analysis and automatic classification of breath sounds. Biomed. Eng. IEEE Trans. 1984; 31: 585–90. 3 Sankur B, Kahya YP, Agatay Gler E, Engin T. Comparison of ar-based algorithms for respiratory sounds classification. Comput. Biol. Med. 1994; 24: 67–76. 4 Gavriely ARDN, Nissan M. Spectral characteristics of chest wall breath sounds in normal subjects. Int. J. Respir. Med. 1995; 50: 1292–300. 5 Yeginer M, Ciftci K, Cini U, Sen I, Kilinc G, Kahya YP. Using lung sounds in classification of pulmonary diseases according to respiratory subphases. Eng. Med. Biol. Soc. 2004; 1: 482–5. 6 Zheng H, Wang H, Wang LY, Yin G. Lung sound pattern analysis for anesthesia monitoring. Am. Control Conf. 2005; 3: 1563–8. 7 Matsunaga S, Yamauchi K, Yamashita M, Miyahara S. Classification between normal and abnormal respiratory sounds based on maximum likelihood approach. Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, 2009; 517–20. 8 Marshall A, Boussakta S. Signal analysis of medical acoustic sounds with applications to chest medicine. J. Frankl. Inst. 2007; 344: 230–42. © 2015 Asian Pacific Society of Respirology
639
Digital signature of lung sound in DPLD 9 Kandaswamy A, Kumar CS, Ramanathan RP, Jayaraman S, Malmurugan N. Neural classification of lung sounds using wavelet coefficients. Comput. Biol. Med. 2004; 34: 523–37. 10 Sello S, Strambi SK, Michele GD, Ambrosino N. Respiratory sound analysis in healthy and pathological subjects: a wavelet approach. Biomed. Signal Process. Control 2008; 3: 181–91. 11 Murphy R, Vyshedskiy A, Power-Charnitsky V, Bana D, Marinelli P, Wong-Tse A, Paciej R. Automated lung sound analysis in patients with pneumonia. Res. Care 2004; 49: 1490–7. 12 Kahya YP, Yeginer M, Bilgic B. Classifying respiratory sounds with different feature sets. Conf. Proc. IEEE Eng. Med. Biol. Soc. 2006; 1: 2856–9. 13 Mor R, Kushnir I, Meyer J, Ekstein J, Ben-Dov I. Breath sound distribution images of patients with pneumonia and pleural effusion. Res. Care 2007; 52: 1753–60. 14 Hiroshi O, Yasuyuki T, Kinya S, Tetsu S, Shoji K, Akihiko G. Evaluation of the usefulness of spectral analysis of inspiratory lung sounds recorded with phonopneumography in patients with interstitial pneumonia. J. Nippon Med. Sch. 2009; 76: 67–75. 15 Zeng W, Smith JTB. Lung sound analysis in the diagnosis of obstructive airway disease. Respiration 2008; 77: 134–8. 16 Flietstra B, Markuzon N, Vyshedskiy A, Murphy R. Automated analysis of crackles in patients with interstitial pulmonary fibrosis. J. Pulm. Med. 2010; 2011: 1–7. 17 Mondal A, Bhattacharya PS, Saha G. Reduction of heart sound interference from lung sound signals using empirical
© 2015 Asian Pacific Society of Respirology
18 19
20
21 22
23
24
mode decomposition technique. J. Med. Eng. Technol. 2011; 35: 344–53. Allain C, Cloitre M. Characterizing the lacunarity of random and deterministic fractal sets. Phys. Rev. 1991; 44: 3552–8. Al-Angari HM, Sahakian AV. Use of sample entropy approach to study heart rate variability in obstructive sleep apnea syndrome. Biomed. Eng. IEEE Trans. 2007; 54: 1900–4. Sovijärvi AR, Malmberg LP, Paajanen E, Piirilä P, Kallio K, Katila T. Averaged and time-gated spectral analysis of respiratory sounds. Repeatability of spectral parameters in healthy men and in patients with fibrosing alveolitis. Chest 1996; 109: 1283–90. Piirilä P, Sovijärvi AR. Crackles: recording, analysis and clinical significance. Eur. Respir. J. 1995; 8: 2139–48. Murphy R. Computerized multichannel lung sound analysis. Development of acoustic instruments for diagnosis and management of medical conditions. IEEE Eng. Med. Biol. Mag. 2007; 26: 16–19. Ellington LE, Gilman RH, Tielsch JM, Steinhoff M, Figueroa D, Rodriguez S, Caffo B, Tracey B, Elhilali M, West J et al. Computerised lung sound analysis to improve the specificity of paediatric pneumonia diagnosis in resource-poor settings: protocol and methods for an observational study. BMJ Open 2012; 2: 1–7. Gurung A, Scrafford CG, Tielsch JM, Levine OS, Checkley W. Computerized lung sound analysis as diagnostic aid for the detection of abnormal lung sounds: a systematic review and meta-analysis. Respir. Med. 2011; 105: 1396–403.
Respirology (2015) 20, 633–639
