Professional Documents
Culture Documents
DOI 10.1007/s10916-010-9498-8
ORIGINAL PAPER
analysis technique [6, 8–13]. Yule-Walker, Burg, covari- bility. (ii) Many parameters such as the number of hidden
ance, least squares, and maximum likelihood estimation are layers for training of the MLP NN, number of neurons, type
examples to the various estimation methods for the of the activation functions, a set of parameters for learning
parameters of AR method [1, 2]. Burg AR method stands function and a probable pruning step must be selected by
as a powerful calculation method in terms of computational the user so that the optimum result can be taken. These
efficiency, and resulting in consistent estimates, is frequent- parameters also make the MLP NN algorithm complicated
ly used in PSD estimates of internal carotid artery Doppler and make it difficult to coding. Existence of many
signals [8, 19, 20]. additional parameters will extend the training phase. The
In the literature, there are various studies on classifica- adverse check of MLP NN exhibits further difficulty for
tion of the carotid artery Doppler signals. Özşen et al. requiring the classification margin notion of a target
extracted the features with AR method and classified these presented recently [38].
features using a new Artificial Immune Systems classifier LVQ NN is a nearest-neighbor pattern classifier based on
[11]. Ceylan et al. used Complex Valued Artificial Neural competitive learning [32]. A LVQ NN has a competitive
Network (CVANN) structure to classify carotid artery (Kohonen) layer and a linear output layer. The linear layer
Doppler signals using Principal Component Analysis and transforms the classes of competitive layer into user-defined
Fuzzy-C-Means Clustering (FCM) as feature extraction classifications [31]. LVQ NN structures have been success-
methods before the CVANN classifier [14]. Özbay & fully used in the tasks of diagnostics of various diseases
Ceylan used Fast Fourier Transform, Hilbert Transform, [33–35], facial recognition [36] and text classification [37].
and Welch Method with different window types. They LVQ NN applications have superiorities in terms of
investigated effects of window types on classification of more qualitative specifications [33]: (I) LVQ networks have
carotid artery Doppler signals [15]. Polat et al. used Support feature reproducibility because the initialization is more
Vector Machine (LSSVM) with a fuzzy-weighted prepro- deterministic. So they produce the same outputs in different
cessing step in classification progress [16]. In Güler & runs, (II) LVQ networks are fast during training; they just
Übeyli’s studies, attributes attained from the carotid artery need a few adjusted parameters for training, and can be
Doppler signals by the help of the various spectral analysis retrained with the templates of the local individuals. Also,
methods were classified according to MLP NN-based since the algorithms are quite simple, application of the
classification systems [8, 17, 18]. LVQ NN into the software is simple as such (III) Outlier
Artificial Neural Network (ANN) is a calculation system, detection of LVQ NN, which is based on classifying by
having been developed through inspiration from the structure, distance, can easily be configured to reject patterns that are
as well as from the learning characteristics of neural cells. It is not within a certain distance of the prototype patterns, (IV)
similarly and successfully applied on the functional features of LVQ networks can be trained with the unstable dimensions
human brain, in terms of such aspects as learning, optimiza- of various classes. In this study, LVQ NN was used as
tion, prediction, clustering, generalization, and classification. classifier for such specifications.
Main reasons of the frequent use of ANN in classification This study basically consists of feature extraction and
applications are as follows [21]: (i) ANN’s simple structure for classification stages. At the feature extraction stage, power
easier use in hardware platform; (ii) ANN’s easier mapping of spectral density (PSD) estimates of internal carotid artery
complex class-distributed features; (iii) ANN’s generalization Doppler signals were obtained by using Burg AR spectrum
feature. Its generation of results appropriate for input vectors analysis technique. At classification stage, LVQ NN was
unavailable in training set; (iv) Weights indicating the result used to classify internal carotid artery stenosis and healty
are found via repeated trainings. While it is possible to use all subjects by using obtained features with Burg AR spectrum
ANN models as classifiers, the most widely-used model is analysis technique. The classification success which is
MLP with back-propagation NN in medical field. MLP with observed within the proposed LVQ NN method is compared
back-propagation NNs show a very broad range of applica- with MLP NN, SVM, KNN, NB and C4.5 decision tree
tions like diagnosis of different diseases [22, 26–28], analysis methods’ classification success, based on the sensitivity/
of EEG [23, 24] and ECG patterns [25, 29] and classification specificity and receiver operating characteristic (ROC) curve
of Doppler signals [8, 30] in medical field. techniques. The classification results showed that the LVQ
Although MLP NN is successfully used in a broad NN method was effective for classification of internal carotid
scope, it is not always the best model among the ANN artery Doppler signals.
models. A MLP NN features the following disadvantages. Rest of the paper is organized as follows. In “Materials
(i) When the same data set is used for training and test, and methods”, by the obtainment of internal carotid artery
MLP NN produces different outputs in each run because Doppler signals, background theory of Burg AR and
the MLP NN is run with random initialization of the LVQ NN methods are explained. The efficiency of the
weights. This makes difficult reproducibility and compara- LVQ NN method, intended for classification of carotid
J Med Syst
artery Doppler signals is demonstrated in “Experimental In Burg method, PSD estimation can be defined as
results”. “Conclusions”, the paper is concluded with the Eq. 1.
discussion section.
^ ðf Þ ¼ e^p
2 ð1Þ
1 þ Pp
P xx
k¼1 a^p ðkÞej2pf k
Materials and methods
where a^p ðkÞ is AR parameters to reflection coefficients, e^p
The parts of suggested biomedical system structure are is total least-squares error, p denotes model order, f denotes
explained in the following subsections: frequency.
Selecting the model order makes up a key component in
Raw data obtaining parametric methods. Through various techniques, the
optimal model order can be determined [2, 43, 44]. Akaike
Doppler signals were recorded from the 191 subjects, 136 proposed a better criterion for choosing the model order,
patients as well as of 55 healthy people in Radiology clinics namely the Akaike information criterion (AIC) [44]. In
of Medical Center of Fırat University. Internal carotid artery AIC, the model order is selected through the minimization
examinations were performed with a Doppler unit using a of Eq. 2.
6.2–8.4 MHz linear transducer. Various tools consist of
Doppler unit (Toshiba SSA-770 Aplio 80, Toshiba, Tokyo, AICð pÞ ¼ ln s^ 2 þ 2p=N ð2Þ
Japan), an analog/digital interface board and Personal
Computer was used for raw data obtaining. The analog where s^ 2 is the estimated variance of the linear prediction
Doppler unit works in not only continuous mode but also error. In this study, model order of the AR method was
pulse which wave mode. The probe was most often placed taken as 10 by using Eq. 2.
at an angle of 60° towards the internal carotid artery.
Learning vector quantization
Power spectral density using Burg AR method
LVQ NN is the nearest-neighbor classifier based on
Different spectrum estimation methods fall under either competitive learning [32]. A LVQ NN has two layers, being
the nonparametric or parametric method categories. In the first competitive and the second linear. The neurons in the
Nonparametric methods, it is the signal that is used competitive layer are designated as subclasses. Each subclass
directly in the estimation of PSD. Periodogram is the has a weight vector. When an input is applied to a LVQ NN,
easiest method that can be given as an example to that. the best matching is searched for in the competitive layer. The
On the other hand, in Parametric methods, a model is best matching is designated as the winning neuron. If a
used in the process of estimating the power spectrum. particular neuron in the competitive layer wins, the particular
Burg, Welch, and Yule Walker methods are the most output of the class of the neuron is set high [36]. The
common parametric methods used. competitive layer learns the classification of input vectors.
There are two steps for estimating the spectrum in the The linear layer transforms the classes of competitive layer
parametric methods. Method parameters are estimated into user-defined classifications. Both competitive and linear
according to the data series x(n), 0 ≤ n ≤ N−1. PSD layers have one neuron per class [31].
estimate is computed using the estimations obtained here. During the training process of the LVQ NN, the
As stated above, AR method is preferred as a spectrum Euclidean distance from a training vector, x, to each node’s
analysis technique, due to the simplification of the weight vector, wi, in the competitive layer is computed
estimation of AR parameters (such as Yule-Walker, Burg, according to the Eq. 3.
covariance, least squares, and maximum likelihood esti- ( )1=2
mation). Unlike other AR estimation methods, Burg XN 2
di ¼ kwi xk ¼ wij xj ð3Þ
method doesn’t calculate the autocorrelation function but j¼1
directly estimates the reflection coefficients. Because of
the fact that PSD estimates are close to the true values in The nearest node is declared as the winner, and the
the AR case, to resolve closely spaced sinusoids in signals weight vector is adjusted according to whether the winner
with low noise levels and to estimate short data records node is included in the class of the training vector:
can be stated as the main advantages of the Burg’s method
[42]. Moreover, Burg AR method is efficient giving stable wiþ1 ¼ wi þ aðx wi Þ; if the winner is the correct class
wiþ1 ¼ wi þ g ðx wi Þ; otherwise
estimates. On the grounds of these advantages, in this
study Burg method is used. ð4Þ
J Med Syst
Table 3 The statistical parameters of the classifiers methods mance of the classifier using sensitivity and specificity
Classifier Specificity (%) Sensitivity (%) Accuracy (%) values. The ROC curve which is seen Fig. 4 represents the
LVQ NN performance on the test data set. In addition to
LVQ NN 96.36 98.53 97.91 Fig. 4, the areas under the ROC curves were calculated for
SVM 96.23 97.10 96.86 all classifiers in this study. The areas under the ROC curves
KNN 98.15 98.54 98.43 were found to be 0.974 for LVQ NN; 0.956 for SVM, and
Naive Bayes 83.08 99.21 93.75 MLP NN; 0.967 for C4.5; 0.978 for KNN; and 0.950 Naive
C4.5 92.98 98.51 96.86 Bayes classifiers. According to these results, proposed
MLP NN 96.23 97.10 96.86 system can be used confidently to help experts for decision
making in their diagnosis problems.
of internal carotid artery Doppler signals. Overall LVQ NN 18. Güler, İ., and Übeyli, E. D., Implementing wavelet/probabilistic
neural networks for Doppler ultrasound blood flow signals. Expert
can be used confidently to help experts for decision making
Syst. Appl. 33(1):162–170, 2007.
in their diagnosis problems. 19. Dirgenali, F., and Kara, S., Recognition of early phase of
atherosclerosis using principles component analysis and artificial
Acknowledgements The authors acknowledge the support of this neural networks from carotid artery Doppler signals. Expert Syst.
study provided by Selçuk University Scientific Research Projects. Appl. 31(3):643–651, 2006.
20. Kemaloğlu, S., Erdoğan, N., and Kara, S., Discontinuous doppler
signals simulating respiratory misregistration: Effect on autoregres-
sive frequency spectra. Comput. Biol. Med. 36(5):465–472, 2006.
References 21. Ölmez, T., and Dokur, Z., Classification of heart sound using an
artificial neural network. Pattern Recogn. Lett. 24:617–629, 2003.
1. Stoica, P., and Moses, R., Introduction to spectral analysis. 22. Türe, M., and Kurt, İ., Comparison of four different time series
Prentice-Hall, New Jersey, 1997. methods to forecast hepatitis A virus infection. Expert Syst. Appl.
2. Kay, S. M., Modern spectral estimation: theory and application. 31(1):41–46, 2006.
Prentice-Hall, New Jersey, 1988. 23. Walczak, S., and Nowack, W. J., An artificial neural network
3. Baker, W. H., Diagnosis and treatment of carotid artery disease. approach to diagnosing epilepsy using lateralized bursts of theta
Futura Publishing Company Inc., New York, 1985. EEGs. J. Med. Syst. 25(1):9–20, 2001.
4. Wright, I. A., Gough, N. A. J., Rakebrandt, F., Wahab, M., and 24. Subaşı, A., Application of classical and model-based spectral
Woodcock, J. P., Neural network analysis of Doppler ultrasound methods to describe the state of alertness in EEG. J. Med. Syst. 29
blood flow signals: A pilot study. Ultrasound Med. Biol. 23 (5):473–486, 2005.
(5):683–690, 1997. 25. Sinha, R. K., Aggarwal, Y., and Das, B. N., Backpropagation
5. Evans, D. H., McDicken, W. N., Skidmore, R., and Woodcock, J. artificial neural network detects changes in electro-encephalogram
P., Doppler ultrasound: Physics. Instrumentation and clinical power spectra of syncopic patients. J. Med. Syst. 31:63–68, 2007.
applications. Wiley, Chichester, 1989. 26. Hayashi, Y., and Setiono, Y., Combining neural network pre-
6. Übeyli, E. D., and Güler, İ., Comparison of eigenvector methods with dictions for medical diagnosis. Comput. Biol. Med. 32(4):237–
classical and model-based methods in analysis of internal carotid 246, 2002.
arterial Doppler signals. Comput. Biol. Med. 33(6):473–493, 2003. 27. Er, O., and Temurtaş, F., A study on chronic obstructive
7. Güler, İ., Kıymık, M. K., Akın, M., and Aklan, A., AR Spectral pulmonary disease diagnosis using multilayer neural networks.
analysis off EEG signals by using maximum likelihood estima- J. Med. Syst. 32:429–432, 2008.
tion. Comput. Biol. Med. 31:441–450, 2001. 28. Çınar, M., Engin, M., Engin, E. Z., and Ateşçi, Y. Z., Early
8. Übeyli, E. D., and Güler, İ., Neural network analysis of internal prostate cancer diagnosis by using artificial neural networks and
carotid arterial Doppler signals: Predictions of stenosis and support vector machines. Expert Syst. Appl. 36(3):6357–6361,
occlusion. Expert. Syst. Appl. 25:1–13, 2003. 2009.
9. Übeyli, E. D., and Güler, İ., Spectral analysis of internal carotid 29. Moavenian, M., and Khorrami, H., A qualitative comparison of
arterial Doppler signals using FFT, AR, MA, and ARMA Artificial Neural Networks and Support Vector Machines in ECG
methods. Comput. Biol. Med. 34(4):293–306, 2004. arrhythmias classification. Expert Syst. Appl. 37(4):3088–3093,
10. Übeyli, E. D., and Güler, İ., Selection of optimal AR spectral 2010.
estimation method for internal carotid arterial Doppler signals 30. Yıldırım, H., Altınsoy, H. B., Barışçı, N., Ergün, U., Oğur, E.,
using Cramer-Rao bound. Comput. Electr. Eng. 30(7):491–508, Hardalaç, F., and Güler, İ., Classification of the frequency of
2004. carotid artery stenosis with MLP and RBF neural networks in
11. Özşen, S., Kara, S., Latifoğlu, F., and Güneş, S., A new patients with coroner artery disease. J. Med. Syst. 28(6):591–601,
supervised classification algorithm in artificial immune systems 2004.
with its application to carotid artery Doppler signals to diagnose 31. Demuth, H., and Beale, M., Neural netwoks toolbox for use with
atherosclerosis. Comp. Methods Programs Biomed. 88:246–255, MATLAB. The MathWorks Inc., Massachusetts, USA, 1994.
2007. 32. Bashar, M. K., Ohnishi, N., Matsumoto, T., Takeuchi, Y., Kudo,
12. Kara, S., and Dirgenali, F., A system to diagnose atherosclerosis H., and Agusa, K., Image retrieval by pattern categorization using
via wavelet transforms, principal component analysis and artificial wavelet domain perceptual features with LVQ neural network.
neural networks. Expert Syst. Appl. 32(2):632–640, 2007. Pattern Recogn. Lett. 26:2315–2335, 2005.
13. Übeyli, E. D., Statistics over features for internal carotid arterial 33. Dieterle, F., Müller-Hagedorn, S., Liebich, H. M., and Gauglitz,
disorders detection. Comput. Biol. Med. 38:361–371, 2008. G., Urinary nucleosides as potential tumor markers evaluated by
14. Ceylan, M., Ceylan, R., Dirgenali, F., and Özbay, Y., Classifica- learning vector quantization. Artif. Intell. Med. 28:265–279, 2003.
tion of carotid artery Doppler signals in the early phase of 34. Temurtaş, F., A comparative study on thyroid disease diagnosis
atherosclerosis using complex-valued artificial neural network. using neural networks. Expert Syst. Appl. 36:944–949, 2009.
Comput. Biol. Med. 37(1):28–36, 2007. 35. Kara, S., and Güven, A., Training a learning vector quantization
15. Özbay, Y., and Ceylan, M., Effects of window types on network using the pattern electroretinography signals. Comput.
classification of carotid artery Doppler signals in the early phase Biol. Med. 37:77–82, 2007.
of atherosclerosis using complex-valued artificial neural network. 36. Bashyal, S., and Venayagamoorthy, G. K., Recognition of facial
Comput. Biol. Med. 37(3):287–295, 2007. expressions using Gabor wavelets and learning vector quantiza-
16. Polat, K., Kara, S., Latifoğlu, F., and Güneş, S., Pattern detection tion. Eng. Appl. Artif. Intell. 21:1056–1064, 2008.
of atherosclerosis from carotid artery Doppler signals using fuzzy 37. Martin-Valdivia, M. T., Urena-Lopez, L. A., and Garcia-Vega, M.,
weighted pre-processing and least square support vector machine The learning vector quantization algorithm applied to automatic
(LSSVM). Ann. Biomed. Eng. 35(5):724–732, 2007. text classification tasks. Neural Netw. 20(6):748–756, 2007.
17. Güler, İ., and Übeyli, E. D., A recurrent neural network classifier 38. Allwein, E. L., Schapire, R. E., and Singer, Y., Reducing
for Doppler ultrasound blood flow signals. Pattern Recogn. Lett. multiclass to binary: a unifying approach for magrin classifiers.
27(1):1560–1571, 2006. J. Machine Learn Res. 1:113–41, 2000.
J Med Syst
39. Masouma, S., Bouveresse, D. J. R., Vercauteren, J., Jalali-Heravi, 41. Akın, M., and Kıymık, M. K., Application of periodogram and AR
M., and Rutledgeb, D. N., Discrimination of wines based on 2D spectral analysis to EEG signals. J. Med. Syst. 24(4):247–256, 2000.
NMR spectra using learning vector quantization neural networks 42. Faust, O., Acharya, R. U., Allen, A. R., and Lin, C. M., Analysis
and partial least squares discriminant analysis. Anal. Chim. Acta of EEG signals during epileptic and alcoholic states using AR
558:144–149, 2006. modeling techniques. IRBM 29(1):44–52, 2008.
40. Beksaç, M. S., Başaran, F., Eskiizmirliler, S., Erkmen, A. M., 43. Proakis, J. G., and Manolakis, D. G., Digital signal processing
and Yörükan, S., A computerized diagnostic system for the principles, algorithms, and applications. Prentice-Hall, New
interpretation of umbilical artery blood flow velocity wave- Jersey, 1996.
forms. Eur. J. Obstet. Gynecol. Reprod. Biol. 64(1):37–42, 44. Akaike, H., A new look at the statistical model identification.
1996. IEEE Trans. Autom. Control 19:716–723, 1974.