Deep Learning for COPD Analysis Using Lung Sounds

Abstract

In this study, Hilbert-Huang Transform (HHT) was applied to the lung sounds from RespiratoryDatabase@TR and the statistical features were calculated from the different modulations of the HHT. The statistical features were fed into the DBN to classify the lung sounds from Chronic Obstructive Pulmonary Disease (COPD) and healthy subjects.

Full text

DEEP LEARNING FOR COPD ANALYSIS USING LUNG SOUNDS
GOKHAN ALTAN1, YAKUP KUTLU2, NOVRUZ ALLAHVERDI3
1,2Iskenderun Technical University, Computer Engineering Dept., Iskenderun, Hatay, Turkey
3KTO Karatay University, Computer Engineering Dept., Karatay, Konya, Turkey
e-mail: gokhan altan@hotmail.com, yakup.kutlu@iste.edu.tr, novruz.allahverdi@karatay.edu.tr
Keywords: Extreme Learning Machines, ELM, Hessenberg Decomposition, ELM Autoencoder,
RespiratoryDatabase@TR, COPD.
1. Introduction
Deep Learning (DL) algorithms have become popular with the detailed analyzing capabilities
with many hidden layers in recent years. The size of hidden layer in the classifier models is
completely correlated with the analyzing capability of the proposed models. Multiple hidden
layers and neuron size in the hidden layers enhance the analyzing capability of the models,
whereas increasing the training time [2]. When using lots of hidden layers provides enhancing
analyzing capabilities by assessing different presentations of the input, on the other hand it
costs much training time. The idea of reducing the training time for the DL algorithms is the
main focus point of recent researches. Although the DL is a neural network structure which
has many hidden layers, they differ in consequence of performing variant back-propagation
procedures during the training and the definition of the classification parameters (including
weight and biases) in pre-training. While the input weights and the hidden node parameters
are randomly defined for neural network model, the DL algorithm pre-defines the weights and
biases using unsupervised learning models including Restricted Boltzmann Machines (RBM)[9],
Sparse autoencoders [8], and Extreme Learning Machines (ELM) Autoencoders [6].
The DL algorithms differ with the unsupervised learning phase and the feature learning
models in the training process. The Deep Belief Networks (DBN) is frequently used for training
considering the capabilities of accessing the global minimum and high classification performances
with fast greedy layer-wise pre-training of the layers [8]. The DBN has two stage classification.
The input weights and the hidden layer biases are defined using RBM in the first stage, the
pre-defined parameters are optimized unfolding them into neural network model with the same
structure at the second stage [9, 10, 8]. The main point of the DBN is fast training speed causing
pre-defining the parameters before optimization and enabling the global minimum with small
number of iterations for optimization [10].
The DBN was applied to classify various biomedical signals for asthma disease diagnosis
models [7], diagnosis of the coronary artery disease [2], arryhthmia classification [3] and brain
activity detection [4]. In this study, Hilbert-Huang Transform (HHT) was applied to the lung
sounds from RespiratoryDatabase@TR and the statistical features were calculated from the
different modulations of the HHT. The statistical features were fed into the DBN to classify the
lung sounds from Chronic Obstructive Pulmonary Disease (COPD) and healthy subjects.
1

2 GOKHAN ALTAN, YAKUP KUTLU, NOVRUZ ALLAHVERDI
2. Materials and Method
2.1. Classifier. The DBN algorithm performs RBM to pre-define the classification parameters
using unsupervised ways to address the deficiency of training time on deep models with multiple
hidden layers. The DBN performs layer-by-layer top-down directed learning operations and
defines generative weights. The generative weights represents the relationship between adjacent
layers, how the parameters in a layer rely on the parameters in the adjacent layer above. Upper
layers of the DBN provide to represent more abstract features where as the lower layers of the
DBN learn simple features. Each RBM in the DBN model generates different presentation of
the input data [10]. Energy function (1) and probability function (2) of the DBN model are :
E(v, h) = X
i
fi(v, h) = −bv −ch −W v h (1)
P(v, h) = 1
Ze−E(v,h)(2)
vis input layer vector, his hidden layer vector. band crepresent for biases of the DBN model
for visible and hidden layer, respectively. Zis normalization constant for the RBM distribution,
Wis the weights for pre-training phase of the model.
2.2. Database. RespiratoryDatabase@TR is a unique multimedia respiratory database which
has 12-channel lung sounds, chest X-rays, 4-channel heart sounds, spirometry metrics from
subjects with the COPD and healthy subjects [5]. It generates a wide analysing potentiality
for the COPD and asthma diseases using computerized signal analysis and machine learning
approaches. 12-channel lung sounds with 15s duration from 30 subjects (15 COPD+15 Healthy)
were utilized in the analysis.
2.3. Hilbert-Huang Transform. HHT is an adaptable and efficient transformation method to
overcome the non-linearity and non-stationarity signal problems. The HHT enables extracting
time-frequency-energy characteristics of the signal [11].
The HHT is a two step transformation including Empirical Mode Decomposition(EMD) and
Hilbert Transform (HT), consecutively. The EMD extracts Intrinsic Mode Functions (IMFs)
which are the ortogonal basis frequency modulations of the signals without leaving the time
domain.The formulation of the EMD process is:
X(t) =
n
X
j=1
IMFj+rn(3)
rnis the residual signal, Xrepresent the input signal and nis the number of the sifted
IMFs. The HT is applied to the sifted IMF modulations for counting instantaneous frequency
characteristics[1, 11]. Analytical function of the HT for an x(t) is formulated as follows:
x(t) = <(n
X
i=1
ai(t)ejWi(t)dt )(4)
3. Experimental Results
The HHT is applied to the 12-channel lung sounds. The HHT-based statistical features
including standard deviation, mean, median, maximum, minimum, variance, mode, correlation
coefficient, kurtosis, moment, cumulant, and energy for each IMFs were calculated as dataset. It
was fed into the DBN model with 2 hidden layers (340,580 neurons). The DBN was iterated for
50 epochs. The learning rate was selected as 2 and the sigmoid activation function was utilized
as the output function for the DBN model. The statistical metrics such as accuracy, sensitivity,
and selectivity were calculated from the contingency table of classification to estimate differences
in distribution of lung sounds from the COPD patients and healthy subjects using 6-fold cross

G. ALTAN, Y. KUTLU, N. ALLAHVERDI: DEEP LEARNING FOR COPD ANALYSIS USING LUNG SOUNDS 3
Table 1. The classification performances (%) of the DBN
Accuracy Sensitivity Specificity
IMF1 33.61 28.89 38.33
IMF2 62.78 66.67 58.89
IMF3 50.83 53.33 48.33
IMF4 47.50 54.44 40.56
IMF5 38.06 36.67 39.44
All 70.28 67.22 73.33
SFFS 90.83 94.44 87.22
validation technique. The achieved results are presented in Table 1 considering the IMF-based
feature sets and the entire feature set.
The proposed DBN model has achieved high results for each IMF feature set and entire
feature set expect IMF1. IMF1 is the first modulation which has still noise, that is why it is
the lowest responsible feature for the classifier. The DBN has separated the lung sounds from
the COPD and healthy lung sounds with classification performance rates of 70.28%, 67.22%,
and 73.33% for accuracy, sensitivity, and specificity, respectively. The sequential forward feature
selection (SFFS) algorithm is performed on the Deep ELM classifier model and has increased
the classification performance rates to 90.83%, 94.44%, and 87.22% for accuracy, sensitivity and
specificity.
4. Acknowledgements
This study is supported by Scientific and Technological Research Council of Turkish (TUBITAK-
116E190). The authors express their thanks to TUBITAK for providing fully support.
References
[1] Altan G., Kutlu Y., Allahverdi N., A new approach to early diagnosis of congestive heart failure disease by
using Hilbert Huang transform, Computer Methods and Programs in Biomedicine, Vol.137, December 2016,
p.23-34, ISSN 0169-2607, http://dx.doi.org/10.1016/j.cmpb.2016.09.003
[2] Altan G., Allahverdi N., Kutlu Y., Diagnosis of Coronary Artery Disease Using Deep Belief Networks,
European Journal of Engineering and Natural Sciences, 2017, Vol:2 Issue(1),ISSN:2458-8156, p.29-36
[3] Altan G., Kutlu Y., Allahverdi N., A Multistage Deep Belief Networks Application on Arrhythmia Classifi-
cation, International Journal of Intelligent Systems and Applications in Engineering, Volume 4, Special Issue
1, December 2016, Pages 23-34, ISSN 2147-6799, DOI: 10.18201/ijisae.270367
[4] Altan G., Kutlu Y., Allahverdi N., Deep Belief Networks Based Brain Activity Classification Using EEG from
Slow Cortical Potentials in Stroke, International Journal of Applied Mathematics, Electronics and Computers,
Vol.4, Issue.I,2016, p.205-210, http://dx.doi.org/10.18100/ijamec.270307
[5] Altan G., Kutlu Y., Garbi Y., Pekmezci A.O., Nural S., Multimedia Respiratory Database (Respiratory-
Database@TR): Auscultation Sounds And Chest X-rays, Natural and Engineering Sciences, 2017, ISSN:2458-
8989, Vol.2, Issue:3, pp.5972, DOI: 10.28978/nesciences.349282
[6] Altan G., Kutlu Y., Pekmezci A.O., Yayik A., Diagnosis of Chronic Obstructive Pulmonary Disease using
Deep Extreme Learning Machines with LU Autoencoder Kernel, 7th International Conference on Advanced
Technologies, 2018, (Accepted)
[7] Altan G., Kutlu Y., Pekmeci A.O., Nural S., Asthma Analysis using Deep Learning, Full Text, 7th Interna-
tional Conference on Advanced Technologies (ICAT18) (Accepted)
[8] Bengio Y., Lamblin P., Popovici D., Larochelle H., Greedy layer-wise training of deep networks, NIPS’06
Proceedings of the 19th International Conference on Neural Information Processing Systems, 2006, p.153-160
[9] Hinton G.E, Osindero S., Teh Y.W., A fast learning algorithm for deep belief nets. Neural Computation,
2006, Vol.18 Issue:7, p.1527-1554, DOI=http://dx.doi.org/10.1162/neco.2006.18.7.1527.
[10] Hinton G., Deep belief networks, Scholarpedia, 2009, Vol.4 Issue:5, p.5947. doi:10.4249/scholarpedia.5947.
[11] Huang N.E., Shen Z., Long S.R. et al., The empirical mode decomposition and the Hilbert spectrum for
nonlinear and non-stationary time series analysis. Proceedings of the Royal Society of London, 1998, A(454),
p.903-995.