J Med Syst DOI 10.1007/s10916-008-9241-x
ORIGINAL PAPER
Tuberculosis Disease Diagnosis Using Artificial Neural Networks Orhan Er & Feyzullah Temurtas & A. Çetin Tanrıkulu
Received: 5 November 2008 / Accepted: 8 December 2008 # Springer Science + Business Media, LLC 2008
Abstract Tuberculosis is an infectious disease, caused in most cases by microorganisms called Mycobacterium tuberculosis. Tuberculosis is a great problem in most low income countries; it is the single most frequent cause of death in individuals aged fifteen to forty-nine years. Tuberculosis is important health problem in Turkey also. In this study, a study on tuberculosis diagnosis was realized by using multilayer neural networks (MLNN). For this purpose, two different MLNN structures were used. One of the structures was the MLNN with one hidden layer and the other was the MLNN with two hidden layers. A general regression neural network (GRNN) was also performed to realize tuberculosis diagnosis for the comparison. Levenberg-Marquardt algorithms were used for the training of the multilayer neural networks. The results of the study were compared with the results of the pervious similar studies reported focusing on tuberculosis diseases diagnosis. The tuberculosis dataset were taken from a state hospital’s database using patient’s epicrisis reports. Keywords Tuberculosis disease diagnosis . Multilayer neural network . General regression neural network O. Er Department of Electrical and Electronics Engineering, Sakarya University, 54187 Adapazari, Turkey F. Temurtas (*) Department of Electrical and Electronics Engineering, Bozok University, 66200 Yozgat, Turkey e-mail:
[email protected] A. Ç. Tanrıkulu Department of Chest Diseases, Sutcu Imam University, 46100 Kahramanmaras, Turkey
Introduction Tuberculosis is an infectious disease, caused in most cases by microorganisms called Mycobacterium tuberculosis. The microorganisms usually enter the body by inhalation through the lungs. They spread from the initial location in the lungs to other parts of the body via the blood stream, the lymphatic system, via the airways or by direct extension to other organs. When infectious people cough, sneeze, talk or spit, they propel tuberculosis bacterium into the air. A person needs only to inhale a small number of these to be infected. The risk of becoming infected depends principally on how long and how intense the exposure to the bacterium is. The risk is greatest in those with prolonged, close household exposure to a person with infectious tuberculosis. [1–3]. Typical symptoms of pulmonary tuberculosis include chronic cough, weight loss, intermittent fever, night sweats and coughing blood. Tuberculosis develops in the human body in two stages. The first stage occurs when an individual who is exposed to micro-organisms from an infectious case of tuberculosis becomes infected (tuberculosis infection), and the second is when the infected individual develops the disease (tuberculosis) [1, 2]. Tuberculosis is a major cause of illness and death worldwide, especially in Asia and Africa. It is a great problem in most low income countries; it is the single most frequent cause of death in individuals aged fifteen to fortynine years. Globally, 9.2 million new cases and 1.7 million deaths from tuberculosis occurred in 2006, of which 0.7 million cases and 0.2 million deaths were in HIV-positive people. [1, 3]. According to the government health ministry’s statistic report in 2006, approximate 23.875 patients have tuberculosis and every year 3.448 patients die because of tuberculosis in our Country [3].
J Med Syst
The multilayer neural networks (MLNNs) have been successfully used in replacing conventional pattern recognition methods for the disease diagnosis systems [4–7]. The back-propagation (BP) algorithm [8] is widely recognized as a powerful tool for training of the MLNNs. But, since it applies the steepest descent method to update the weights, it suffers from a slow convergence rate and often yields suboptimal solutions [9, 10]. A variety of related algorithms have been introduced to address that problem. A number of researchers have carried out comparative studies of MLNN training algorithms [11–13]. Levenberg-Marquardt (LM) algorithm [11] used in this study provides generally faster convergence and better estimation results than other training algorithms [7, 13]. This paper aims to present a comparative study for the realization of the tuberculosis diagnosis using multilayer neural networks (MLNNs). For this purpose, two different MLNN structures were used. One of the structures was the MLNN with one hidden layer and the other was the MLNN with two hidden layers. A general regression neural network (GRNN) was also performed to realize tuberculosis diagnosis for the comparison. Levenberg-Marquardt algorithms were used for the training of the multilayer neural networks. The study aims also to provide machine learning based decision support system for contributing to the doctors in their diagnosis decisions. The results were also compared with the results of the pervious study reported [14, 15] focusing on tuberculosis disease diagnosis and using different methods and database. The tuberculosis dataset were taken from a state hospital’s database using patient’s epicrisis reports.
Data source In order to perform the research reported in this manuscript, the patient’s epicrisis reports taken from Diyarbakir Chest Diseases Hospital from southeast of Turkey was used. The dataset were prepared using these epicrisis reports the dataset which consists of the tuberculosis disease measurements contains two classes and 150 samples. The class distribution is: & &
transferase (AST), bilirubin (total+ direct), CK/ creatine kinase total, CK-MB, iron (SERUM), gamma-glutamil transferase (GGT), glukoz, HDL cholesterol, calcium (CA), blood urea nitrogen (BUN), chlorine (CL), cholesterol, creatinin, lactic dehydrogenase(LDH), potassium(K), sodium (NA), total protein, triglesid, uric acid.
Previous studies There have been several studies reported focusing on tuberculosis disease diagnosis problem [14, 15]. El-Solh, et al., used a general regression neural network (GRNN) using clinical and radiographic information to predict active pulmonary tuberculosis at the time of presentation at a health-care facility that is superior to physicians’ opinion [14]. The input patterns were formed by 21 distinct parameters which were divided into three groups: demographic variables, constitutional symptoms, and radiographic findings. The output of the GRNN provided an estimate of the likelihood of active pulmonary tuberculosis. The authors utilized a 10-fold cross-validation procedure to train the neural networks. The authors reported approximately 92.3 % diagnosis accuracy [14]. Santos, et al., used a prediction model for diagnosis of smear negative pulmonary tuberculosis (SNPT) [15]. They used symptoms and physical signs for constructing the neural network (NN) modelling. They reported approximately 77 % diagnosis accuracy. They used a MLNN structure with one hidden layer
Diagnosis of the tuberculosis disease using neural networks In the first stage of the study, the multilayer neural network structures with one and two hidden layers were used for the tuberculosis disease diagnosis. One of the structures was the MLNN with one hidden layer and the other was the MLNN with two hidden layers. The MLNN with two hidden layers is shown in Fig. 1. The hidden layer neurons
Class 1: Tuberculosis (50) Class 2: Normal (100)
All samples have thirty eight features. These features are (Laboratory examination): complaint of cough, body temperature, ache on chest, weakness, dyspnea on exertion, rattle in chest, pressure on chest, sputum, sound on respiratory tract, habit of cigarette, leucocyte (WBC), erythrocyte (RBC), trombosit (PLT), hematocrit (HCT), hemoglobin (HGB), albumin2, alkalen phosphatase 2°L, alanin aminotransferase (ALT), amylase, aspartat amino-
Fig. 1 Implementation of multilayer neural network for the tuberculosis disease diagnosis
J Med Syst
(50 neurons for each hidden layer) and the output layer neurons use nonlinear sigmoid activation functions. In this system, thirty eight inputs were features, and three outputs are index of three classes (normal and tuberculosis). Detailed computational issues about the implementation of multilayer neural network structures for the disease diagnosis can be found in our previous studies [6, 7]. For the training of the MLNNs, back-propagation with momentum (BPwM) and Levenberg-Marquardt (LM) [11] algorithms were used. Detailed computational issues about the application of the training algorithms to MLNN structures can be found in references [13, 16]. A general regression neural network (GRNN) was also performed to realize tuberculosis diagnosis for the comparison. Computational issues about the GRNN structure can be found in references [14, 16]. If a neural network learns the training set of a problem, it makes generalization to that problem. So, this type trained neural network gives similar result for untrained test sets also. But, if a neural network starts to memorize the training set, its generalization starts to decrease and it’s performance may not be improved for untrained test sets [6, 13]. The kfold cross-validation method shows how good generalization can be made using neural network structures [17]. In this study, 3-fold cross-validation approaches were used to estimate the performance of the used neural networks. Detailed computational issues about the k-fold crossvalidation method can be found in references [5, 6]. As performance measures, we used the classification accuracies n this study [6, 18] : jN j P
classification accuracyðN Þ ¼
assessðni Þ
i¼1
jN j
;
ð12Þ
ni 2 N ( assessðnÞ ¼
1 0
if classifyðnÞ ¼ nc otherwise:
ð13Þ
Table 1 Classification accuracies for tuberculosis disease dataset problem Study
Method
Classification accuracy (%)
[14] [15] This study
GRNN (one hidden layer) MLNN with BP (one hidden layer) GRNN (one hidden layer) MLNN with BPwM (one hidden layer) MLNN with LM (one hidden layer) MLNN with BPwM (two hidden layers) MLNN with LM (two hidden layers)
92.30 77.00 93.18 93.04 93.42 93.93 95.08
reported in same table. This small accuracy difference between this study and reference [14] can be because of that the different input features used in this study. Approximately 77 % accuracy ratio was reported by the reference [15]. But, with the similar structure and same training algorithm, our result was much better (95.08 % accuracy) in this study. This better result can be because of that the input features used in this study represents tuberculosis disease diagnosing much better than the input features used in reference [15]. From the Table 1, it can be seen also that the results obtained using MLNN with two hidden layers were better than the results obtained using MLNN with one hidden layer and the best results for the classification accuracy were obtained from MLNN with two hidden layers trained by LM training algorithm in this study. So we can easily say that MLNN with two hidden layers is better than MLNN with one hidden layer and Levenberg-Marquardt (LM) training algorithm converges better than BP with momentum training algorithm for tuberculosis disease diagnosing. As the conclusion, the following results can be summarised; & &
where N is the set of data items to be classified (the test set), n 2 N , nc is the class of the item n, and classify (n) returns the classification of n by neural networks.
&
Results and conclusions & The classification accuracies obtained by this and other studies for tuberculosis disease dataset were presented in Table 1. According to Table 1, accuracy ratio obtained by the reference [14] was 92.30 %. On the other hand, with the same GRNN structure, we obtained a bit better accuracy as
% % % % % % %
&
Generally, the classification accuracies obtained by this study were better than those obtained in the pervious study reported by other authors [14, 15]. Levenberg-Marquardt (LM) training algorithm converges better than BP with momentum training algorithm for the tuberculosis disease diagnoses problem. The results obtained using MLNN with two hidden layers were better than the results obtained using MLNN with one hidden layer for tuberculosis disease diagnosing. The best results for the classification accuracy were obtained from the MLNN structure with two hidden layers trained by LM training algorithm for the tuberculosis disease diagnosing. The GRNN was also a good choice for the tuberculosis disease diagnosing.
J Med Syst
&
And, it was obtained that neural network structures could be successfully used to help diagnosis of tuberculosis disease. So, these structures can be helpful as learning based decision support system for contributing to the doctors in their diagnosis decisions.
References 1. Enarson, D. A., Rieder, H. L., Arnadottir, T., Trébucq, A., “A Guide For Low Income Countries”, International Union Against Tuberculosis and Lung Disease, 68 boulevard Saint-Michel, 75006 Paris, France, ISBN: 2-914365-00-4, 2000. 2. Royal College of Physicians of London: “Tuberculosis”, 11 St. Andrews Place, London NW1 4LE, ISBN : 1 86016 277 0, 2006. 3. World Health Organization: “Global Tuberculosis Control 2008 Surveillance Planning Financing”, ISBN : 978 92 4 156354 3, Geneva, 2008. 4. Kayaer, K., Yıldırım, T., Medical Diagnosis on Pima Indian Diabetes Using General Regression Neural Networks. In Proc. of International Conference on Artificial Neural Networks and Neural Information Processing (ICANN/ICONIP): Istanbul, (pp. 181–184), 2003. 5. Delen, D., Walker, G., and Kadam, A., Predicting breast cancer survivability: A comparison of three data mining methods. Artif. Intell. Med. 34:2113–127, 2005. 6. Temurtas, F., A comparative study on thyroid disease diagnosis using neural networks. Expert Syst. Appl. 36:944–949, 2009. doi:10.1016/j.eswa.2007.10.010. 7. Er, O., and Temurtas, F., A Study on Chronic Obstructive Pulmonary Disease Diagnosis Using Multilayer Neural Networks. J. Med. Syst. 32:5429–432, 2008. doi:10.1007/s10916-008-9148-6.
8. Rumelhart, D. E., Hinton, G. E., and Williams, R. J., Learning internal representations by error propagation. In: Rumelhart, D. E., and McClelland, J. L. (Eds.), Parallel Distributed Processing: Explorations in the Microstructure of CognitionMIT Press, Vol. 1. Cambridge, MA, pp. 318–362, 1986. 9. Brent, R. P., Fast training algorithms for multi-layer neural nets. IEEE Trans. Neural Netw. 2:346–354, 1991. doi:10.1109/72.97911. 10. Gori, M., and Tesi, A., On the problem of local minima in backpropagation. IEEE Trans. Pattern Anal. Mach. Intell. 14:76– 85, 1992. doi:10.1109/34.107014. 11. Hagan, M. T., and Menhaj, M., Training feed forward networks with the Marquardt algorithm. IEEE Trans. Neural Netw. 5:989– 993, 1994. doi:10.1109/72.329697. 12. Hagan, M. T., Demuth, H. B., and Beale, M. H., Neural Network Design. PWS Publishing, Boston, MA, 1996. 13. Gulbag, A., and Temurtas, F., A study on quantitative classification of binary gas mixture using neural networks and adaptive neuro fuzzy inference systems. Sens. Actuators B Chem. 115:252– 262, 2006. doi:10.1016/j.snb.2005.09.009. 14. El-Solh, A. A., Hsiao, C.- B., Goodnough, S., Serghani, J., and Grant, B. J. B., Predicting active pulmonary tuberculosis using an artificial neural network. Chest. 116:968–973, 1999. doi:10.1378/ chest.116.4.968. 15. dos Santos, A. M., Pereira, B. B., de Seixas, J. M., “Neural Networks: An Application for Predicting Smear Negative Pulmonary Tuberculosis”, Proceedings of the Statistics in the Health Sciences, March, 2004. 16. Matlab® Documentation., Version 7.0, Release 14, The MathWorks, Inc, 2004. 17. Ozyılmaz, L., and Yıldırım, T., Diagnosis of thyroid disease using artificial neural network methods In Proc. of ICONIP’02 9th international conference on neural information processing. Orchid Country Club, Singapore, pp. 2033–2036, 2002. 18. Watkins, A., “AIRS: A resource limited artificial immune classifier”. Master Thesis, Mississippi State University, 2001.
