The Machine Learned Stethoscope Provides Accurate Operator Independent Diagnosis of Chest Disease

Figures & data

Figure 1 (A) Chest sound auscultation and recording areas. (B) MFCC feature extraction steps. (C) Computation windowing and frame continuing.

Table 1 Frequency Range and Different Diagnoses of Studied Records

Chest Sound	Number of Records		Frequency Range (Hz)		Creation Models A & B		Validation of Models A & B
	N	%	Min	Max	N	%	N	%
Normal	147	31.6	68	133	87	33.3	60	29.6
Bronchial breathing	68	14.7	110	158	47	18.0	21	10.3
Crepitation	70	15.1	161	205	33	12.7	37	18.2
Wheezes	179	38.6	133	265	94	36.0	85	41.9
Total	464	100	–	–	261	100	203	100

Notes: Bronchial breathing caused by pneumonia. Crepitation caused by Bronchitis. Wheezes caused by bronchial asthma.

Abbreviations: N, total number of records; Min, minimum frequency; Max, maximum frequency.

Table 2 Chest Model CCR, Sensitivity, and Specificity Using Different Number of MFCC Coefficients and Different Frame Durations & Chest Models A and B Confusion Matrix, CCR, Sensitivity, and Specificity Using 13 MFCCs and Frame Duration 40 ms

						CCR %	Overall Sensitivity %	Overall Specificity %
Number of MFCCs
13 MFCCs						96.7	100	100
14 MFCCs						94.8	100	100
15 MFCCs						94.5	99.6	100
16 MFCCs						93.8	99.6	100
17 MFCCs						94.9	100	100
18 MFCCs						94.5	100	96.6
19 MFCCs						95.6	100	98.3
Frame duration
30 ms						95.5	99.6	100
35 ms						95.5	99.6	100
40 ms						96.7	100	100
45 ms						93.8	99.6	93.3
50 ms						94.5	99.6	95.0
55 ms						92.6	99.6	88.3
60 ms						91.6	98.6	93.3
	Normal	Bronchial Breathing	Crepitation	Wheezes	Number	CCR %	Overall Sensitivity %	Overall Specificity %
Chest Model A
Normal	60	0	0	0	60	100	100	100
Bronchial breathing	0	21	0	0	21	100
Crepitation	0	0	33	4	37	89.1
Wheezes	0	0	2	83	85	97.6
Chest Model B
Normal	60	0	0	0	60	100	100	100
Bronchial breathing	0	21	0	0	21	100
Crepitation	0	0	36	1	37	97.3
Wheezes	0	0	2	83	85	97.6

Notes: Sensitivity = true positive/(true positive + false negative). Specificity = true negative/(true negative + false positive). Number: total number of records for each sound type.

Abbreviations: MFCCs, mel frequency cepstral coefficients; CCR, correct classification rate; ms, millisecond.

Figure 2 Building of chest sound recognition models A & B and recognition of chest sounds wave shape.

Notes: (A) Model A building. (B) Model B building. (C) Wave shape of (a) wave of wheezes sound, (b) wave of bronchial breathing sound, (c) wave of crepitation sound, (d) wave of normal chest sound, (e) wave of underlying pathology pneumonia sound, (f) wave of underlying pathology bronchopneumonia sound.

Abbreviations: PPG, phonopulmonography; s, second.

Table 3 Comparison of Chest Sound Recognition Ability of Previous ACA Models and Current Models A & B

Reference	Analyzed Sounds	Method	Results
Gogus et al, 2015^Citation9	Asthmatic breath sound and normal sound (11 records)	WPT, DWT ANN	DWT accuracy = 67.67-91.67% WPT accuracy = 80.80%-90.0%
Chen et al, 2015^Citation12	Wheezes, ronchi, crackles (29 records)	MFCC K-NN	System A average identification rate= 77.1%. System B average identification rate= 92.3%
Rizal et al, 2006^Citation13	Crackles, grunting, ronchi and pleural rub (324 online records)	ART2 NN	Overall accuracy up to 86%
Riella et al, 2009^Citation14	Wheezes, normal vesicular sounds (112 records)	Spectrogram image	Accuracy= 84.82%-92.86%
Syafria et al, 2014	Normal and abnormal*	VQ, BPNN	VQ accuracy= 86.8%, BPNN accuracy= 93.1%
Mondal et al, 2014^Citation17	Normal and abnormal*	SVM	Sensitivity= 86.3% Accuracy= 92.8%
Sengupta et al, 2015^Citation19	Normal, crackle and wheeze (72 records)	ANN, Cepstral feature	Accuracy= 92–97%
Hashemi et al, 2012^Citation20	Wheezes*	ANN, MFCCs	Accuracy = 92.8%.
Chien et al, 2007^Citation21	Wheezes and normal*	Gaussian Mixture Models, MFCC	Accuracy = 90%.
Lin et al, 2015^Citation22	26 Normal and 32 wheezes	BPNN	Sensitivity = 94.6% Specificity = 100%
^#Chest machine learning model A	Normal, wheezes, bronchial breathing and crepitation (464 records)	HMM, MFCC	Sensitivity = 100%, CCR = 96.7% Specificity = 100%.
^#Chest machine learning model B	Normal, wheezes, bronchial breathing and crepitation (464 records)	HMM, MFCC, Wave shape	Sensitivity = 100%, CCR = 98.7% Specificity = 100%.

Notes: Accuracy measurements: measures amount of uncertainty with respect to an absolute standard. *Number of sound records was not mentioned. ^#Model described in current study.

Abbreviations: ANN, artificial neural network; ART2, adaptive resonance theory 2; HMM, hidden markov model; MFCC, mel frequency cepstral coefficient; DWT, discrete wavelet transform; WPT, wavelet packet transform; VQ, vector quantization; BPNN, back propagation neural network; SVM, support vector machine.

Figure 3 Sensitivity and specificity of models A and B.

Notes: (A) Automatic chest auscultation model CCR, sensitivity and specificity using different number of MFCC elements combined with frame duration length 40 ms, and (B) according to different frame duration length combined with 13 MFCCs coefficients. (C) Comparison between first and second model according CCR, sensitivity, and specificity.

Abbreviation: ms, millisecond.

Göğüş FZ, Bekir Karlık B, Harman G. Classification of asthmatic breath sounds by using wavelet transforms and neural networks. Int J Signal Process Syst. 2015;3:106–111.

 Google Scholar

Chen CH, Huang WT, Tan TH, Chang CC, Chang YJ. Using K-nearest neighbor classification to diagnose abnormal lung sounds. Sensors. 2015;15:13132–13158. doi:10.3390/s150613132

 Web of Science ®Google Scholar

Rizal A, Mengko TLR, Suksmono AB. Lung sound recognition using wavelet packet decomposition and ART2 (adaptive resonance theory 2) neural network. Biomedical Engineering Day 2006; 2006: 168–172. Available from: http://achmadrizal.staff.telkomuniversity.ac.id/wpcontent/uploads/sites/11/2013/12/PR04-BMEDAY06.pdf. Accessed December 15, 2017.

 Google Scholar

Riella RJ, Nohama P, Maia JM. Method for automatic detection of wheezing in lung sounds. Braz J Med Biol Res. 2009;42:674–684. doi:10.1590/S0100-879X2009000700013

 Web of Science ®Google Scholar

Kiyokawa H, Greenberg M, Shirota K, Pasterkamp H. Auditory detection of simulated crackles in breath sounds. Chest. 2001;119(6):1886–1892. doi:10.1378/chest.119.6.1886

 PubMed Web of Science ®Google Scholar

Hashemi A, Arabalibeik H, Agin K. Classification of wheeze sounds using cepstral analysis and neural networks. Stud Health Technol Inform. 2012;173:161–165.

 PubMedGoogle Scholar

Chien JC, Wu HD, Chong FC, et al. Wheeze detection using cepstral analysis in Gaussian mixture models. Conf Proc IEEE Eng Med Biol Soc. 2007;3168–3171.

 Google Scholar

Lin BS, Wu HD, Chen SJ. Automatic wheezing detection based on signal processing of spectrogram and back-propagation neural network. J Healthc Eng. 2015;6:649–672. doi:10.1260/2040-2295.6.4.649

 Web of Science ®Google Scholar

Ramsay CF. An evaluative study of breath sounds. Thorax. 2011;66:A144. doi:10.1136/thoraxjnl-2011-201054c.185

 Web of Science ®Google Scholar