- Email: [email protected]
Diagnosis of chest diseases using artificial immune system
Expert Systems with Applications 39 (2012) 1862–1868
Contents lists available at SciVerse ScienceDirect
Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Diagnosis of chest diseases using artificial immune system Orhan Er a, Nejat Yumusak b, Feyzullah Temurtas a,⇑ a b
Bozok University, Department of Electrical and Electronics Engineering, 66200 Yozgat, Turkey Sakarya University, Department of Computer Engineering, 54187 Adapazari, Turkey
a r t i c l e
i n f o
Keywords: Chest disease diagnosis Artificial immune system
a b s t r a c t Chest diseases are one of the greatest health problems for people living in the developing world. Millions of people are diagnosed every year with a chest disease in the world. Chronic obstructive pulmonary, pneumonia, asthma, tuberculosis, lung cancer diseases are most important chest diseases and these are very common illnesses in Turkey. In this paper, a study on chest diseases diagnosis was realized by using artificial immune system. We obtained the classification accuracy with artificial immune system 93.84%. The result of the study was compared with the results of the previous similar studies reported focusing on chest diseases diagnosis. The chest diseases dataset were prepared from a chest diseases hospital’s database using patient’s epicrisis reports. Ó 2011 Elsevier Ltd. All rights reserved.
1. Introduction There are a number of other chest diseases which should challenge the ingenuity and determination of progressive-minded physicians. Diseases of the various organs and structures of the chest represent a large enough group and a complexity of problems to completely engage the talent and ability of interested physicians. Competence and efficiency in handling these conditions are more readily acquired when studied as interrelated items. The very nature of the close anatomic and functional connections between the organs potentially involved implies that only through a correlated understanding of each problem is it possible to diagnose and treat diseases of the chest. Chest diseases have some phenomenon such as air pollution various infections, and smoking habit have recently increased its risk factors drastically. Thus, millions of people are diagnosed every year with a chest disease in the world (Baemani, Monadjemi, & Moallem, 2008). Tuberculosis (TB), chronic obstructive pulmonary disease (COPD), pneumonia, asthma, lung cancer diseases are the most important chest diseases which are very common illnesses in the world (MedHelp, http://www.medhelp.org/MedicalDictionary/Terms/2/8964.htm (last accessed: 18 March 2009)). 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 tuberculosis develops in the human body in two stages. The first stage occurs when an ⇑ Corresponding author. Tel.: +90 532 207 56 33. E-mail address: [email protected] (F. Temurtas). 0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.08.064
individual who is exposed to micro-organisms from an infectious case of tuberculosis becomes infected (tuberculous infection), and the second is when the infected individual develops the disease (tuberculosis) (Enarson, Rieder, Arnadottir, & Trébucq, 2000). TB is a major cause of illness and death worldwide and globally, 9.2 million new cases and 1.7 million deaths from tuberculosis occurred in 2006 (Royal College of Physicians of London, 2006; World Health Organization, 2008). COPD is a disease state characterized by airflow limitation that is not fully reversible. The airflow limitation is usually both progressive and associated with an abnormal inflammatory response of the lungs to noxious particles or gases (Celli & MacNee, 2004). Clinically, patients with COPD experience shortness of breath (dyspnea) and cough, productive of an excess of mucus. There may also be wheeze (Jeffery, 1998). According to the World Health Organization (WHO) data is found 600 million patients who have COPD and every year 2.3 million persons die because of COPD in the world (Sönmez & Uzaslan, 2006). Pneumonia is an inflammation or infection of the lungs most commonly caused by a bacteria or virus. Pneumonia can also be caused by inhaling vomit or other foreign substances. In all cases, the lungs’ air sacs fill with pus, mucous, and other liquids and cannot function properly. This means oxygen cannot reach the blood and the cells of the body effectively. According to the World Health Organization (WHO) data, every year approximate 2.4 million persons die because of pneumonia (Global Action Plan for the Prevention, 2007). Asthma is a chronic disease characterized by recurrent attacks of breathlessness and wheezing. During an asthma attack, the lining of the bronchial tubes swell, causing the airways to narrow and reducing the flow of air into and out of the lungs. Recurrent asthma symptoms frequently cause sleeplessness, daytime fatigue,
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868
reduced activity levels and school and work absenteeism. Asthma has a relatively low fatality rate compared to other chronic diseases. WHO estimates that 300 million people currently suffer from asthma. Asthma is the most common chronic disease among children (http://www.who.int/en/ (last accessed: 18 March 2009)). Lung cancer is a disease of uncontrolled cell growth in tissues of the lung. This growth may lead to metastasis, which is the invasion of adjacent tissue and infiltration beyond the lungs. The vast majority of primary lung cancers are carcinomas of the lung, derived from epithelial cells. Lung cancer, the most common cause of cancer-related death in men and the second most common in women, is responsible for 1.3 million deaths worldwide annually (http://www.who.int/en/ (last accessed: 18 March 2009)). It is possible to improve the post-treatment lung function by early diagnosis, better treatment and follow up. The appropriate implementation of these methods, which are so important in the early diagnosis of chest, not only improves the opportunity for treatment of chest disease but also has an effect on decreasing deaths from this disease. Artificial immune system (AIS) is a new artificial intelligence (AI) technique which is beginning to mature through the collaborative effort of many interdisciplinary researchers (Andrews & Timmis, 2005). By modeling some metaphors existing in natural immune system or by inspiring from these metaphors, successful applications have being conducted in AI literature. Classification is among these and there have been some promising studies in this branch of AIS. Considering medical diagnosis as an application domain for AISs, there are several studies like (Castro, Coelho, Caetano, & Von Zuben, 2005; Er, Sertkaya, Temurtas, & Tanrikulu, 2009; Hamaker & Boggess, 2004; Polat, Sahan, & Gunes, 2006; Polat, Sahan, Kodaz, & Gunes, 2005; Sahan, Polat, Kodaz, & Gunes, 2005) in AIS literature. Artificial immune systems (AIS) can be defined as abstract or metaphorical computational systems developed using ideas, theories, and components, extracted from the immune system. Most AIS aim at solving complex computational or engineering problems, such as pattern recognition, classification, elimination, and optimization. The AIS has been formed on the basis of the working principles of the natural immune system found in the human body (Engin & Döyen, 2004). The biological immune system (BIS) is a subject of great research interest because of its powerful information processing capabilities; in particular, understanding the distributed nature of its memory, self-tolerance and decentralized control mechanisms from an informational perspective, and building computational models believed to better solve many science and engineering problems (Dasgupta, 2006). Chest disease diagnosis via proper interpretation of the chest diseases data is an important classification problem. And, AIS can provide an alternative, efficient way for solving chest disease diagnosis problems. In this paper, a comparative study on chest diseases diagnosis was realized by using artificial immune systems. The chest diseases dataset were prepared by using patient’s epicrisis reports from a chest diseases hospital’s database. The study aims also to provide machine learning based decision support system for contributing to the doctors in their diagnosis decisions.
2. Method 2.1. Data source In order to perform the research reported in this manuscript, the patient’s epicrisis taken from Diyarbakir Chest Diseases Hospital from southeast of Turkey was used. The dataset were prepared using these epicrisis reports. The study included 357 patients suffering from variety of respiratory diseases and 38 healthy subjects.
1863
The patients were classified as having tuberculosis, COPD, pneumonia, asthma, lung cancer diseases or not sick. Tuberculosis disease were diagnosed in 50 patients, COPD disease were diagnosed in 71 patients, pneumonia disease were diagnosed in 60 patients, asthma disease were diagnosed in 44 patients and lung cancer disease were diagnosed in 32 patients. In dataset we have 100 people who have healthy features. All samples have 38 features. These features are (Laboratory examination): complaint of cough, body temperature, ache on chest, weakness, dyspnoea 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 aminotransferase (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. Diagnostic tests of each patient were recorded by an attending physician. 2.2. Previous studies There have been several studies reported focusing on chest disease diagnosis problem using artificial neural network and artificial immune system structures as for other clinical diagnosis problems. These studies have applied different structures to the various chest diseases diagnosis problem using their various dataset (Aliferis, Hardin, & Massion, 2002; Ashizawa et al., 2005; Coppini, Miniati, Paterni, Monti, & Ferdeghini, 2007; El-Solh, Hsiao, Goodnough, Serghani, & Grant, 1999; Er & Temurtas, 2008; Er, Temurtas, & Tanrikulu, 2010; Er, Yumusak, & Temurtas, 2010; Er et al., 2009; Hanif, Lan, Daud, & Ahmad, 2009; Paul, Ben, Thomas, & Robert, 2004; Santos, Pereira, & Seixas, 2004). El-Solh et al., used generalized 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 (El-Solh et al., 1999). 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 (El-Solh et al., 1999). Er et al., used multilayer, and generalized regression neural networks for diagnosis of tuberculosis (Er, Temurtas, et al., 2010). They used 38 features for the diagnosis and reported approximately 93.3% diagnosis accuracy for GRNN and 95% diagnosis accuracy for MLNN with LM algorithm and two hidden layer. Aliferis et al., used KNN, Decision Tree Induction, Support Vector Machines and Feed-Forward Neural Networks for classify nonsmall lung cancers. The primary goal of their study was to develop machine learning models that classify non-small lung cancers according to histopathology types and to compare several machine learning methods in this learning task. The best multi-gene model found had a leave-one-out accuracy of 89.2% with Feed-Forward Neural Networks (Aliferis et al., 2002). Ashizawa et al., used the MLNN with one hidden layer and they used BP training algorithm for diagnosis of COPD disease (Ashizawa et al., 2005). They used 26 features for the diagnosis. The authors reported approximately 90% diagnosis accuracy. Coppini et al., used the MLNNs with one and two hidden layers and they used BP with momentum as the training algorithm for diagnosis of COPD disease (Coppini et al., 2007). The authors utilized a 10fold cross-validation procedure to train the neural networks. The
1864
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868
Table 1 The best classification accuracies of the previous studies using neural networks. Database
Study
Disease
Method
Different Database
El-Solh et al. (1999) Ashizawa et al. (2005) Coppini et al. (2007) Paul et al. (2004)
Tuberculosis
Pneumonia
Hanif et al. (2009)
Asthma
Aliferis et al. (2002) Er and Temurtas (2008) Er et al. (2009)
Lung cancers
GRNN with one hidden layer MLNN with one hidden layer MLNN with two hidden layers MLNN with two hidden layers NN – radial basis function NN – feed forward
Same Database
COPD COPD
COPD
MLNN with two hidden layers PNN AIS MLNN with two hidden layers
COPD and pneumonia
Er, Temurtas, et al. (2010)
Tuberculosis
Er, Yumusak, et al. (2010)
Chest disease (tuberculosis, COPD, pneumonia, asthma, lung cancer diseases or not sick)
authors were obtained 90.6% diagnosis accuracy using the MLNN with two hidden layers as the best result. They used radiograph shape features for the diagnosis. Er and Temurtas used multilayer neural networks for diagnosis of COPD (Er & Temurtas, 2008). They used 38 features for the diagnosis and reported approximately 96% diagnosis accuracy for MLNN with LM algorithm and two hidden layer. Er et al., used multilayer, probabilistic, and learning vector quantization neural networks for diagnosis of COPD and pneumonia diseases (Er et al., 2009). They used 38 features for the diagnosis and reported approximately 93.92% diagnosis accuracy for probabilistic neural network as the best result. Paul et al., used the MLNN with one and two hidden layers and they used BP with momentum as the training algorithm for predicting community acquired pneumonia among patients with respiratory complaints (Paul et al., 2004). They performed genetic algorithms to search for optimal hidden layer architectures, connectivity, and training parameters for the neural network. The authors reported an ROC (Heckerling, 2002) accuracy ratio of 82.8% for the pneumonia disease diagnosis. Hanif et al., used three different artificial neural networks to classify different severity of asthma and the suitable control measures to overcome it. These neural networks were feed forward back propagation neural network (multilayer neural network), Elman backpropagation neural network and radial basis function neural network (RBF). The accuracy of the trained architectures was tested by inputting new sets of data to a created Graphical User Interface (GUI). They obtained best accuracy result (90%) using the radial basis function network (Hanif et al., 2009). Er et al., used some neural networks structures (shown Table 1 as: MLNN with BP (one and two hidden layer), MLNN with LM (one and two hidden layer), PNN, LVQ, GRNN, RBF) for diagnosis of chest diseases (Er, Yumusak, et al., 2010). They used 38 features for the diagnosis and reported approximately 92.16% diagnosis accuracy for probabilistic neural network as the best result.
MLNN with one hidden layer MLNN with two hidden layer MLNN with one hidden layer MLNN with two hidden layer PNN LVQ GRNN RBF
Training algorithm
Classification accuracy (%) 92.30
BP
90.00
BP with momentum BP with momentum BP
90.60 82.8 (ROC accuracy ratio) 90.00 89.20
LM
96.08
LM
93.92 94.00 95.08
BPwM
89.08
BPwM
90.76
LM
90.48
LM
91.60 92.16 88.52 88.24 90.20
Er et al., used an AIS structure (shown Table 1) for diagnosis of COPD and pneumonia diseases (Er et al., 2009). They used 38 features for the diagnosis and reported approximately 94.00% diagnosis accuracy for AIS. There have been several studies reported focusing on chest diseases diagnosis using artificial neural network and artificial immune system structures as summarized in Table 1. These studies have applied different neural networks structures to the various chest diseases diagnosis problem and achieved high classification accuracies using their various dataset. On the other hand, direct comparison of the results is impossible. Because, the different data set were used by these studies. So, these neural networks were compared using the same data set which consists of the 38 features. 2.3. Diagnosis of chest disease using artificial immune system The artificial immune system has been formed on the basis of the working principles of the natural immune system found in the human body. Tissues and organs related with the natural immune system in the body are the thymus gland, the bone marrow, the lymph nodes, the spleen and the tonsils. A central organ coordinating the functions of the associated tissue, the organ, the molecule and the cells does not exist in the immune system. The immune system, via its special cells, recognizes the foreign (external) cells filtering through the body and neutralizes those (Trojanowski & Wierzchon, 2002). Two different selection methods are utilized for purposes of reaching a solution in different types of problems as regards to artificial immune systems functioning on the basis of the natural immune system. The negative selection mechanism is used for problems such as pattern recognition, anomaly detection, computer and network security and time series analysis. The clonal selection mechanism, on the other hand, is particularly used for problems such as multi-purpose and combinatory optimization,
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868
Antibody cells
1
2
…
32
33
34
35
36
37
38
0
0
…
1.38
354
5.38
151
6.83
139
2.14
1. existing antibody (pneumonia)
0
1
…
1.44
455
5.11
140
6.46
162
0.59
2. existing antibody (pneumonia)
0
1
…
1.44
455
5.38
140
6.83
162
0.59
new clone (antibody / antigen)
1865
Fig. 1. An example of the generating clones of antibodies for pneumonia class.
Fig. 2. Calculations of the step 3.
computer and network security and error detection (De Castro & Von Zuben, 2000). These problem-solving methods that are used in artificial immune systems thoroughly imitate the mechanisms found in the natural immune system that the human body possesses. At the fourth stage of this study, an artificial immune system model was used for the chest disease diagnosis. The algorithmic steps of AIS model used for this purpose are, Step 1: Create the antibody population and determine the suppression threshold. For this study suppression threshold is 0.5. Step 2: Generate clones (new antibody/antigen) for each antibody. Step 3: Calculate the affinity among antibody cells and kill the antibodies whose affinities are less than the suppression threshold and determine the number of antibodies after suppression. Step 4: If not ensure that memory population is constant, return to step 2. Step 5: Classify the given values.
The antibody values are normal, CPOD, pneumonia, tuberculosis, asthma, lung cancer classes at the algorithm. For generating clones of the antibodies, the antibody cells are mutated. In this study, each antibody has 38 antibody cells. In other words, 38 features were used as 38 antibody cells. An example of the generating clones of antibodies which used in AIS model shown in Fig. 1. For measuring affinity of generated antibody cells, Hamming model was used. Detailed calculations of the step 3 which was used in the artificial immune system based model algorithm can be seen in Fig. 2. Where, N is numbers of training patterns, Abi(n) is existing antibody, Agi is new clone, Ap is previous affinity, Af is final affinity, Yf is estimated class, Yd is desired class, E is suppression threshold (is equal to 0.5 for this study), n is training pattern index. The step 5 of the artificial immune system based model was used for the classification of the test patterns. The calculation details of this step are similar to the calculation details of the step 3. These calculation details can be seen in Fig. 3. Detailed computational issues about the application of the AIS structures can be found in De Castro and Timmis (2002) and De Castro and Von Zuben (2000, 2002).
1866
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868
Fig. 3. Calculations of the step 5.
The performance of AIS was evaluated by the following measures. 2.3.1. Classification accuracy In this study, we used the classification accuracies as performance measures (Temurtas, 2009; Watkins, 2001):
PjNj classification accuracyðNÞ ¼
assessðnÞ ¼
i¼1 assessðni Þ
jNj
1 if classifyðnÞ ¼ nc 0
otherwise
;
ni 2 N
ð1Þ
ð2Þ
where N is the set of data items to be classified (the test set), n e N, nc is the class of the item n, and classify(n) returns the classification of n by AIS. 2.3.2. k-Fold cross-validation In order to minimize the bias associated with the random sampling of the training and holdout data samples in comparing the predictive accuracy of two or more methods, researchers tend to use k-fold cross-validation. In k-fold cross-validation, whole data are randomly divided to k mutually exclusive and approximately equal size subsets. The classification algorithm trained and tested k times. In each case, one of the folds is taken as test data and the remaining folds are added to form training data. Thus k
different test results exist for each training-test configuration (Delen, Walker, & Kadam, 2005; Gulbag & Temurtas, 2006; Temurtas, 2009). The average of these results gives the test accuracy of the algorithm. If an AIS learns the training set of a problem, it makes generalization to that problem. So, this type AIS gives similar result for untrained test sets also. But, if an AIS starts to memorize the training set, its generalization starts to decrease and its performance may not be improved for untrained test sets (Sertkaya, 2009). The k-fold cross-validation method shows how good generalization can be made using AIS structures (Ozyılmaz & Yıldırım, 2002). In this study, while conducting the classification procedure, 10fold cross validation method was used to estimate the performance of the used AIS. For test results to be more valuable, k-fold cross validation (10-fold for our case) is used among the researchers. It minimizes the bias associated with the random sampling of the training (Delen et al., 2005). The whole data was randomly divided to 10 mutually exclusive and approximately equal size subsets. Because the whole dataset contains 257 patient data and 100 healthy data, each fold was to consist of 36 samples. Among these 36 samples, 26 were taken from recordings from patients while the remaining 10 samples were of healthy persons. The classification algorithm trained and tested 10 times. In each case, one of the folds is taken as test data and the remaining folds are added to form training data. Thus 10 different test results exist for each training-test configuration. The average of these 10 results gives the test
1867
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868 Table 2 Average of classification accuracies for each disease. Database
Method
Our previous study
MLNN MLNN MLNN MLNN PNN LVQ GRNN RBF
This study
AIS
with with with with
Classification accuracy (%)
BPwM (one hidden layer) BPwM (two hidden layers) LM (one hidden layer) LM (two hidden layers)
Tuberculosis
COPD
Pneumonia
Asthma
Lung cancer
Normal
Average
84.00 84.00 84.00 90.00 88.00 84.00 86.00 86.00
84.51 87.32 87.32 88.73 88.73 84.51 83.10 87.32
88.33 90.00 91.67 90.00 88.33 86.67 88.33 90.00
88.64 90.91 88.84 90.91 90.91 86.37 86.37 88.64
87.50 87.50 93.75 90.63 93.75 93.75 84.38 90.63
96.00 98.00 95.00 96.00 99.00 94.00 95.00 95.00
89.08 90.76 90.48 91.60 92.16 88.52 88.24 90.20
90.00
92.96
93.33
90.91
93.75
98.00
93.84
accuracy of the algorithm (Delen et al., 2005; Er, Yumusak, et al., 2010). 3. Results The classification accuracy obtained by artificial immune system structure for chest diseases is presented in Table 2. The table shows classification accuracies for each chest diseases and average values by our previous study and this study. From Table 2, it can be seen easily that, the best result for tuberculosis was obtained using MLNN with LM (two hidden layers) and AIS (90%). For COPD (92.96%) and pneumonia (93.33%), the best results were obtained using AIS. For asthma, the best result was obtained using PNN, MLNN with LM (two hidden layers), MLNN with BP (two hidden layers) and AIS (90.91%). For lung cancer, the best result was obtained using PNN, MLNN with LM (two hidden layers), MLNN with BP (two hidden layers) and AIS (93.75%). In this study, the best result for the average classification accuracy was obtained using AIS structure as seen in the same table. The second best result for the classification accuracy was obtained using PNN. The third best result for the classification accuracy was obtained using MLNN with LM (two hidden layers). The classification accuracy performances of MLNN with BP (two hidden layers) and with LM (one hidden layer) were similar and closer to that of RBF. 4. Conclusions The first aim of this study was to develop a diagnostic system leading to more effective usage of the AIS and so to advance the research of chest disease. The second objective was to use AISs in a real world medical classification system and so to show the effectiveness of this Artificial Intelligence field in this problem domain. This paper presents a comparative study for the realization of the chest diseases diagnosis using artificial immune system. There have been several studies reported focusing on chest diseases diagnosis using artificial immune system structures. These studies have applied different structures to the various chest diseases diagnosis problem using their various dataset. Because of the different dataset used by the studies, the direct comparison of the results was impossible. So, this artificial immune system was compared with our previous study on neural networks using the same dataset which consists of the 38 features. As the conclusion, the following results can be summarized; The best results for the average classification accuracy were obtained using AIS for the chest disease diagnosis problem. This classification accuracy is highly reliable for such a problem because only a few samples were misclassified by the system.
This system is capable of conducting the classification process with a good performance to help the expert while deciding the healthy and patient subjects. And, it was obtained that artificial immune system structure could be successfully used to help diagnosis of chest disease. So, this structure can be helpful as learning based decision support system for contributing to the doctors in their diagnosis decisions.
References Aliferis, C. F., Hardin, D., & Massion, P. P. (2002). Machine learning models for lung cancer classification using array comparative genomic hybridization. In Biomedical informatics, Proceedings of the AMIA Symposium 2002:7–11. Nashville, TN, USA: Vanderbilt University. Andrews, P. S., & Timmis, J. (2005). Inspiration for the next generation of artificial immune systems. Lecture Notes in Computer Science, 3627, 126–138. Ashizawa, K., Ishida, T., MacMahon, H., Vyborny, C. J., Katsuragawa, S., & Doi, K. (2005). Artificial neural networks in chest radiography: Application to the differential diagnosis of interstitial lung disease. Academic Radiology, 11(1), 29–37. Baemani, M. J., Monadjemi, A., & Moallem, P. (2008). Detection of respiratory abnormalities using artificial neural networks. Journal of Computer Science, 1549-3636, 4(8), 663–667. Castro, P. D., Coelho, G. P., Caetano, M. F., & Von Zuben, F. J. (2005). Designing ensembles of fuzzy classification systems: An immune-inspired approach. Lecture Notes in Computer Science, 3627, 469–482. Celli, B. R., & MacNee, W. (2004). Standards for the diagnosis and treatment of patients with COPD: A summary of the ATS/ERS position paper. European Respiratory Journal, 23(6), 932–946. Coppini, G., Miniati, M., Paterni, M., Monti, S., & Ferdeghini, E. M. (2007). Computeraided diagnosis of emphysema in COPD patients: Neural-network-based analysis of lung shape in digital chest radiographs. Medical Engineering & Physics, 29, 76–86. Dasgupta, D. (2006). Advances in artificial immune systems. EEE Computational Intelligence Magazine. USA: University of Memphis, November (pp. 40–49). De Castro, L. N., & Von Zuben F. J. (2000). The clonal selection algorithm with engineering applications. In GECCO 2000, Las Vegas, Nevada, USA, July 8. De Castro, L. N., & Timmis, J. (2002). A novel approach to pattern recognition. In L. Alonso, J. Corchado, & C. Fyfe (Eds.), Artificial neural networks in pattern recognition (pp. 67–84). University of Paisley. De Castro, L. N., & Von Zuben, F. J., (2002). Learning and optimization using the clonal selection principle. In the special issue on artificial immune systems of the Journal IEEE Transactions on Evolutionary Computation, June 6(3). Delen, D., Walker, G., & Kadam, A. (2005). Predicting breast cancer survivability: A comparison of three data mining methods. Artificial Intelligence in Medicine Artificial Intelligence in Medicine, 34(2), 1127–1163. El-Solh, A. A., Hsiao, C-B., Goodnough, S., Serghani, J., & Grant, B. J. B. (1999). Predicting active pulmonary tuberculosis using an artificial neural network. Chest, 116, 968–973. Enarson, D. A., Rieder, H. L., Arnadottir, T., & Trébucq, A. (2000). 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. Engin, O., & Döyen, A. (2004). Artificial immune systems and applications in industrial problems. Gazi University Journal of Science, 17(1), 71–84. Er, O., & Temurtas, F. (2008). A study on chronic obstructive pulmonary disease diagnosis using multilayer neural networks. Journal of Medical Systems, 32(5), 429–432. doi:10.1007/s10916-008-9209-x. Er, O., Sertkaya, C., Temurtas, F., & Tanrikulu, A. C. (2009). A comparative study on chronic obstructive pulmonary and pneumonia diseases diagnosis using neural networks and artificial immune system. Journal of Medical Systems, 33(6).
1868
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868
Er, O., Temurtas, F., & Tanrikulu, A. C. (2010). Tuberculosis disease diagnosis using artificial neural networks. Journal of Medical Systems, 34(3). doi:10.1007/ s10916-008-9241-x. Er, O., Yumusak, N., & Temurtas, F. (2010). Chest diseases diagnosis using artificial neural networks. Expert Systems with Applications, 55, 555. doi:10.1016/ j.eswa.2010.04.078. Global Action Plan for the Prevention and Control of Pneumonia (GAPP) (2007). Report of an informal consultation, La Mainaz, Gex, France, 5–7 March 2007. Gulbag, A., & Temurtas, F. (2006). A study on quantitative classification of binary gas mixture using neural networks and adaptive neuro fuzzy inference systems. Sensors and Actuators B: Chemical, 115, 252–262. Hamaker, J. S., & Boggess, L. (2004). Non-euclidean distance measures in AIRS, an artificial immune classification system. In Proceedings of the 2004 congress on evolutionary computing. Hanif, N. H. H. M., Lan, W. H., Daud, H. B., & Ahmad, J. (2009). Classification of Control Measures for Asthma using Artificial Neural Networks. Malaysia: ACTA Press. Heckerling, P. S. (2002). Parametric receiver operating characteristic (ROC) curve analysis using Mathematica. Computer Methods and Programs in Biomedicine, 69, 65–73. Jeffery, P. K. (1998). Structural and inflammatory changes in COPD: A comparison with asthma. Thorax, 53(2), 129–136. MedHelp (2009). Thoracic organs. Last accepted 18.03.09. Ozyılmaz, L., & Yıldırım, T. (2002) Diagnosis of thyroid disease using artificial neural network methods. In Proceeding of the ICONIP’02 9th international conference on neural information processing, Orchid Country Club, Singapore (pp. 2033–2036). Paul, S. H., Ben, S. G., Thomas, G. T., & Robert, S. W. (2004). Use of genetic algorithms for neural networks to predict community-acquired pneumonia. Artificial Intelligence in Medicine, 30, 71–84.
Polat, K., Sahan, S., Kodaz, H., & Gunes, S. (2005). A new classification method for breast cancer diagnosis: Feature selection artificial immune recognition system (FS-AIRS). Lecture Notes in Computer Science, 3611, 830–838. Polat, K., Sahan, S., & Gunes, S. (2006). A new method to medical diagnosis: Artificial immune recognition system (AIRS) with fuzzy weighted pre-processing and application to ECG arrhythmia. Expert Systems with Applications, 31(2), 264–269. Royal College of Physicians of London (2006). Tuberculosis. 11 St. Andrews Place, London NW1 4LE. ISBN: 1 86016 277 0. Sahan, S., Polat, K., Kodaz, H., & Gunes, S. (2005). The medical applications of attribute weighted artificial immune system (AWAIS): Diagnosis of heart and diabetes diseases. Lecture Notes in Computer Science, 3627, 456–468. Santos, A. M. dos, Pereira, B. B., & Seixas, J. M. de, (2004). Neural networks: An application for predicting smear negative pulmonary tuberculosis. In Proceedings of the statistics in the health sciences, March. Sertkaya, C. (2009). Immune base system in computer security. Master Thesis. Sakarya University, Institute of Science and Technology. Sönmez, S., & Uzaslan, E. (2006). Kronik Obstrüktif Akcig˘er Hastalıg˘ı’nın Genetig˘i ve Sitokin Gen Polimorfizmi – Derleme. Archives of Lung, Cilt, 7, 75–78. Temurtas, F. (2009). A comparative study on thyroid disease diagnosis using neural networks. Expert Systems with Applications, 36, 944–949. Trojanowski, K., & Wierzchon, S. T. (2002). Searching for memory in artificial immune system. In The eleventh international symposium on intelligent information systems, June 3–6. Watkins, A. (2001). AIRS: A resource limited artificial immune classifier. Master Thesis, Mississippi State University. World Health Organization (2008). Global Tuberculosis Control 2008 Surveillance Planning Financing, Geneva. ISBN: 978 92 4 156354 3.
Contents lists available at SciVerse ScienceDirect
Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Diagnosis of chest diseases using artificial immune system Orhan Er a, Nejat Yumusak b, Feyzullah Temurtas a,⇑ a b
Bozok University, Department of Electrical and Electronics Engineering, 66200 Yozgat, Turkey Sakarya University, Department of Computer Engineering, 54187 Adapazari, Turkey
a r t i c l e
i n f o
Keywords: Chest disease diagnosis Artificial immune system
a b s t r a c t Chest diseases are one of the greatest health problems for people living in the developing world. Millions of people are diagnosed every year with a chest disease in the world. Chronic obstructive pulmonary, pneumonia, asthma, tuberculosis, lung cancer diseases are most important chest diseases and these are very common illnesses in Turkey. In this paper, a study on chest diseases diagnosis was realized by using artificial immune system. We obtained the classification accuracy with artificial immune system 93.84%. The result of the study was compared with the results of the previous similar studies reported focusing on chest diseases diagnosis. The chest diseases dataset were prepared from a chest diseases hospital’s database using patient’s epicrisis reports. Ó 2011 Elsevier Ltd. All rights reserved.
1. Introduction There are a number of other chest diseases which should challenge the ingenuity and determination of progressive-minded physicians. Diseases of the various organs and structures of the chest represent a large enough group and a complexity of problems to completely engage the talent and ability of interested physicians. Competence and efficiency in handling these conditions are more readily acquired when studied as interrelated items. The very nature of the close anatomic and functional connections between the organs potentially involved implies that only through a correlated understanding of each problem is it possible to diagnose and treat diseases of the chest. Chest diseases have some phenomenon such as air pollution various infections, and smoking habit have recently increased its risk factors drastically. Thus, millions of people are diagnosed every year with a chest disease in the world (Baemani, Monadjemi, & Moallem, 2008). Tuberculosis (TB), chronic obstructive pulmonary disease (COPD), pneumonia, asthma, lung cancer diseases are the most important chest diseases which are very common illnesses in the world (MedHelp, http://www.medhelp.org/MedicalDictionary/Terms/2/8964.htm (last accessed: 18 March 2009)). 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 tuberculosis develops in the human body in two stages. The first stage occurs when an ⇑ Corresponding author. Tel.: +90 532 207 56 33. E-mail address: [email protected] (F. Temurtas). 0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.08.064
individual who is exposed to micro-organisms from an infectious case of tuberculosis becomes infected (tuberculous infection), and the second is when the infected individual develops the disease (tuberculosis) (Enarson, Rieder, Arnadottir, & Trébucq, 2000). TB is a major cause of illness and death worldwide and globally, 9.2 million new cases and 1.7 million deaths from tuberculosis occurred in 2006 (Royal College of Physicians of London, 2006; World Health Organization, 2008). COPD is a disease state characterized by airflow limitation that is not fully reversible. The airflow limitation is usually both progressive and associated with an abnormal inflammatory response of the lungs to noxious particles or gases (Celli & MacNee, 2004). Clinically, patients with COPD experience shortness of breath (dyspnea) and cough, productive of an excess of mucus. There may also be wheeze (Jeffery, 1998). According to the World Health Organization (WHO) data is found 600 million patients who have COPD and every year 2.3 million persons die because of COPD in the world (Sönmez & Uzaslan, 2006). Pneumonia is an inflammation or infection of the lungs most commonly caused by a bacteria or virus. Pneumonia can also be caused by inhaling vomit or other foreign substances. In all cases, the lungs’ air sacs fill with pus, mucous, and other liquids and cannot function properly. This means oxygen cannot reach the blood and the cells of the body effectively. According to the World Health Organization (WHO) data, every year approximate 2.4 million persons die because of pneumonia (Global Action Plan for the Prevention, 2007). Asthma is a chronic disease characterized by recurrent attacks of breathlessness and wheezing. During an asthma attack, the lining of the bronchial tubes swell, causing the airways to narrow and reducing the flow of air into and out of the lungs. Recurrent asthma symptoms frequently cause sleeplessness, daytime fatigue,
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868
reduced activity levels and school and work absenteeism. Asthma has a relatively low fatality rate compared to other chronic diseases. WHO estimates that 300 million people currently suffer from asthma. Asthma is the most common chronic disease among children (http://www.who.int/en/ (last accessed: 18 March 2009)). Lung cancer is a disease of uncontrolled cell growth in tissues of the lung. This growth may lead to metastasis, which is the invasion of adjacent tissue and infiltration beyond the lungs. The vast majority of primary lung cancers are carcinomas of the lung, derived from epithelial cells. Lung cancer, the most common cause of cancer-related death in men and the second most common in women, is responsible for 1.3 million deaths worldwide annually (http://www.who.int/en/ (last accessed: 18 March 2009)). It is possible to improve the post-treatment lung function by early diagnosis, better treatment and follow up. The appropriate implementation of these methods, which are so important in the early diagnosis of chest, not only improves the opportunity for treatment of chest disease but also has an effect on decreasing deaths from this disease. Artificial immune system (AIS) is a new artificial intelligence (AI) technique which is beginning to mature through the collaborative effort of many interdisciplinary researchers (Andrews & Timmis, 2005). By modeling some metaphors existing in natural immune system or by inspiring from these metaphors, successful applications have being conducted in AI literature. Classification is among these and there have been some promising studies in this branch of AIS. Considering medical diagnosis as an application domain for AISs, there are several studies like (Castro, Coelho, Caetano, & Von Zuben, 2005; Er, Sertkaya, Temurtas, & Tanrikulu, 2009; Hamaker & Boggess, 2004; Polat, Sahan, & Gunes, 2006; Polat, Sahan, Kodaz, & Gunes, 2005; Sahan, Polat, Kodaz, & Gunes, 2005) in AIS literature. Artificial immune systems (AIS) can be defined as abstract or metaphorical computational systems developed using ideas, theories, and components, extracted from the immune system. Most AIS aim at solving complex computational or engineering problems, such as pattern recognition, classification, elimination, and optimization. The AIS has been formed on the basis of the working principles of the natural immune system found in the human body (Engin & Döyen, 2004). The biological immune system (BIS) is a subject of great research interest because of its powerful information processing capabilities; in particular, understanding the distributed nature of its memory, self-tolerance and decentralized control mechanisms from an informational perspective, and building computational models believed to better solve many science and engineering problems (Dasgupta, 2006). Chest disease diagnosis via proper interpretation of the chest diseases data is an important classification problem. And, AIS can provide an alternative, efficient way for solving chest disease diagnosis problems. In this paper, a comparative study on chest diseases diagnosis was realized by using artificial immune systems. The chest diseases dataset were prepared by using patient’s epicrisis reports from a chest diseases hospital’s database. The study aims also to provide machine learning based decision support system for contributing to the doctors in their diagnosis decisions.
2. Method 2.1. Data source In order to perform the research reported in this manuscript, the patient’s epicrisis taken from Diyarbakir Chest Diseases Hospital from southeast of Turkey was used. The dataset were prepared using these epicrisis reports. The study included 357 patients suffering from variety of respiratory diseases and 38 healthy subjects.
1863
The patients were classified as having tuberculosis, COPD, pneumonia, asthma, lung cancer diseases or not sick. Tuberculosis disease were diagnosed in 50 patients, COPD disease were diagnosed in 71 patients, pneumonia disease were diagnosed in 60 patients, asthma disease were diagnosed in 44 patients and lung cancer disease were diagnosed in 32 patients. In dataset we have 100 people who have healthy features. All samples have 38 features. These features are (Laboratory examination): complaint of cough, body temperature, ache on chest, weakness, dyspnoea 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 aminotransferase (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. Diagnostic tests of each patient were recorded by an attending physician. 2.2. Previous studies There have been several studies reported focusing on chest disease diagnosis problem using artificial neural network and artificial immune system structures as for other clinical diagnosis problems. These studies have applied different structures to the various chest diseases diagnosis problem using their various dataset (Aliferis, Hardin, & Massion, 2002; Ashizawa et al., 2005; Coppini, Miniati, Paterni, Monti, & Ferdeghini, 2007; El-Solh, Hsiao, Goodnough, Serghani, & Grant, 1999; Er & Temurtas, 2008; Er, Temurtas, & Tanrikulu, 2010; Er, Yumusak, & Temurtas, 2010; Er et al., 2009; Hanif, Lan, Daud, & Ahmad, 2009; Paul, Ben, Thomas, & Robert, 2004; Santos, Pereira, & Seixas, 2004). El-Solh et al., used generalized 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 (El-Solh et al., 1999). 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 (El-Solh et al., 1999). Er et al., used multilayer, and generalized regression neural networks for diagnosis of tuberculosis (Er, Temurtas, et al., 2010). They used 38 features for the diagnosis and reported approximately 93.3% diagnosis accuracy for GRNN and 95% diagnosis accuracy for MLNN with LM algorithm and two hidden layer. Aliferis et al., used KNN, Decision Tree Induction, Support Vector Machines and Feed-Forward Neural Networks for classify nonsmall lung cancers. The primary goal of their study was to develop machine learning models that classify non-small lung cancers according to histopathology types and to compare several machine learning methods in this learning task. The best multi-gene model found had a leave-one-out accuracy of 89.2% with Feed-Forward Neural Networks (Aliferis et al., 2002). Ashizawa et al., used the MLNN with one hidden layer and they used BP training algorithm for diagnosis of COPD disease (Ashizawa et al., 2005). They used 26 features for the diagnosis. The authors reported approximately 90% diagnosis accuracy. Coppini et al., used the MLNNs with one and two hidden layers and they used BP with momentum as the training algorithm for diagnosis of COPD disease (Coppini et al., 2007). The authors utilized a 10fold cross-validation procedure to train the neural networks. The
1864
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868
Table 1 The best classification accuracies of the previous studies using neural networks. Database
Study
Disease
Method
Different Database
El-Solh et al. (1999) Ashizawa et al. (2005) Coppini et al. (2007) Paul et al. (2004)
Tuberculosis
Pneumonia
Hanif et al. (2009)
Asthma
Aliferis et al. (2002) Er and Temurtas (2008) Er et al. (2009)
Lung cancers
GRNN with one hidden layer MLNN with one hidden layer MLNN with two hidden layers MLNN with two hidden layers NN – radial basis function NN – feed forward
Same Database
COPD COPD
COPD
MLNN with two hidden layers PNN AIS MLNN with two hidden layers
COPD and pneumonia
Er, Temurtas, et al. (2010)
Tuberculosis
Er, Yumusak, et al. (2010)
Chest disease (tuberculosis, COPD, pneumonia, asthma, lung cancer diseases or not sick)
authors were obtained 90.6% diagnosis accuracy using the MLNN with two hidden layers as the best result. They used radiograph shape features for the diagnosis. Er and Temurtas used multilayer neural networks for diagnosis of COPD (Er & Temurtas, 2008). They used 38 features for the diagnosis and reported approximately 96% diagnosis accuracy for MLNN with LM algorithm and two hidden layer. Er et al., used multilayer, probabilistic, and learning vector quantization neural networks for diagnosis of COPD and pneumonia diseases (Er et al., 2009). They used 38 features for the diagnosis and reported approximately 93.92% diagnosis accuracy for probabilistic neural network as the best result. Paul et al., used the MLNN with one and two hidden layers and they used BP with momentum as the training algorithm for predicting community acquired pneumonia among patients with respiratory complaints (Paul et al., 2004). They performed genetic algorithms to search for optimal hidden layer architectures, connectivity, and training parameters for the neural network. The authors reported an ROC (Heckerling, 2002) accuracy ratio of 82.8% for the pneumonia disease diagnosis. Hanif et al., used three different artificial neural networks to classify different severity of asthma and the suitable control measures to overcome it. These neural networks were feed forward back propagation neural network (multilayer neural network), Elman backpropagation neural network and radial basis function neural network (RBF). The accuracy of the trained architectures was tested by inputting new sets of data to a created Graphical User Interface (GUI). They obtained best accuracy result (90%) using the radial basis function network (Hanif et al., 2009). Er et al., used some neural networks structures (shown Table 1 as: MLNN with BP (one and two hidden layer), MLNN with LM (one and two hidden layer), PNN, LVQ, GRNN, RBF) for diagnosis of chest diseases (Er, Yumusak, et al., 2010). They used 38 features for the diagnosis and reported approximately 92.16% diagnosis accuracy for probabilistic neural network as the best result.
MLNN with one hidden layer MLNN with two hidden layer MLNN with one hidden layer MLNN with two hidden layer PNN LVQ GRNN RBF
Training algorithm
Classification accuracy (%) 92.30
BP
90.00
BP with momentum BP with momentum BP
90.60 82.8 (ROC accuracy ratio) 90.00 89.20
LM
96.08
LM
93.92 94.00 95.08
BPwM
89.08
BPwM
90.76
LM
90.48
LM
91.60 92.16 88.52 88.24 90.20
Er et al., used an AIS structure (shown Table 1) for diagnosis of COPD and pneumonia diseases (Er et al., 2009). They used 38 features for the diagnosis and reported approximately 94.00% diagnosis accuracy for AIS. There have been several studies reported focusing on chest diseases diagnosis using artificial neural network and artificial immune system structures as summarized in Table 1. These studies have applied different neural networks structures to the various chest diseases diagnosis problem and achieved high classification accuracies using their various dataset. On the other hand, direct comparison of the results is impossible. Because, the different data set were used by these studies. So, these neural networks were compared using the same data set which consists of the 38 features. 2.3. Diagnosis of chest disease using artificial immune system The artificial immune system has been formed on the basis of the working principles of the natural immune system found in the human body. Tissues and organs related with the natural immune system in the body are the thymus gland, the bone marrow, the lymph nodes, the spleen and the tonsils. A central organ coordinating the functions of the associated tissue, the organ, the molecule and the cells does not exist in the immune system. The immune system, via its special cells, recognizes the foreign (external) cells filtering through the body and neutralizes those (Trojanowski & Wierzchon, 2002). Two different selection methods are utilized for purposes of reaching a solution in different types of problems as regards to artificial immune systems functioning on the basis of the natural immune system. The negative selection mechanism is used for problems such as pattern recognition, anomaly detection, computer and network security and time series analysis. The clonal selection mechanism, on the other hand, is particularly used for problems such as multi-purpose and combinatory optimization,
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868
Antibody cells
1
2
…
32
33
34
35
36
37
38
0
0
…
1.38
354
5.38
151
6.83
139
2.14
1. existing antibody (pneumonia)
0
1
…
1.44
455
5.11
140
6.46
162
0.59
2. existing antibody (pneumonia)
0
1
…
1.44
455
5.38
140
6.83
162
0.59
new clone (antibody / antigen)
1865
Fig. 1. An example of the generating clones of antibodies for pneumonia class.
Fig. 2. Calculations of the step 3.
computer and network security and error detection (De Castro & Von Zuben, 2000). These problem-solving methods that are used in artificial immune systems thoroughly imitate the mechanisms found in the natural immune system that the human body possesses. At the fourth stage of this study, an artificial immune system model was used for the chest disease diagnosis. The algorithmic steps of AIS model used for this purpose are, Step 1: Create the antibody population and determine the suppression threshold. For this study suppression threshold is 0.5. Step 2: Generate clones (new antibody/antigen) for each antibody. Step 3: Calculate the affinity among antibody cells and kill the antibodies whose affinities are less than the suppression threshold and determine the number of antibodies after suppression. Step 4: If not ensure that memory population is constant, return to step 2. Step 5: Classify the given values.
The antibody values are normal, CPOD, pneumonia, tuberculosis, asthma, lung cancer classes at the algorithm. For generating clones of the antibodies, the antibody cells are mutated. In this study, each antibody has 38 antibody cells. In other words, 38 features were used as 38 antibody cells. An example of the generating clones of antibodies which used in AIS model shown in Fig. 1. For measuring affinity of generated antibody cells, Hamming model was used. Detailed calculations of the step 3 which was used in the artificial immune system based model algorithm can be seen in Fig. 2. Where, N is numbers of training patterns, Abi(n) is existing antibody, Agi is new clone, Ap is previous affinity, Af is final affinity, Yf is estimated class, Yd is desired class, E is suppression threshold (is equal to 0.5 for this study), n is training pattern index. The step 5 of the artificial immune system based model was used for the classification of the test patterns. The calculation details of this step are similar to the calculation details of the step 3. These calculation details can be seen in Fig. 3. Detailed computational issues about the application of the AIS structures can be found in De Castro and Timmis (2002) and De Castro and Von Zuben (2000, 2002).
1866
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868
Fig. 3. Calculations of the step 5.
The performance of AIS was evaluated by the following measures. 2.3.1. Classification accuracy In this study, we used the classification accuracies as performance measures (Temurtas, 2009; Watkins, 2001):
PjNj classification accuracyðNÞ ¼
assessðnÞ ¼
i¼1 assessðni Þ
jNj
1 if classifyðnÞ ¼ nc 0
otherwise
;
ni 2 N
ð1Þ
ð2Þ
where N is the set of data items to be classified (the test set), n e N, nc is the class of the item n, and classify(n) returns the classification of n by AIS. 2.3.2. k-Fold cross-validation In order to minimize the bias associated with the random sampling of the training and holdout data samples in comparing the predictive accuracy of two or more methods, researchers tend to use k-fold cross-validation. In k-fold cross-validation, whole data are randomly divided to k mutually exclusive and approximately equal size subsets. The classification algorithm trained and tested k times. In each case, one of the folds is taken as test data and the remaining folds are added to form training data. Thus k
different test results exist for each training-test configuration (Delen, Walker, & Kadam, 2005; Gulbag & Temurtas, 2006; Temurtas, 2009). The average of these results gives the test accuracy of the algorithm. If an AIS learns the training set of a problem, it makes generalization to that problem. So, this type AIS gives similar result for untrained test sets also. But, if an AIS starts to memorize the training set, its generalization starts to decrease and its performance may not be improved for untrained test sets (Sertkaya, 2009). The k-fold cross-validation method shows how good generalization can be made using AIS structures (Ozyılmaz & Yıldırım, 2002). In this study, while conducting the classification procedure, 10fold cross validation method was used to estimate the performance of the used AIS. For test results to be more valuable, k-fold cross validation (10-fold for our case) is used among the researchers. It minimizes the bias associated with the random sampling of the training (Delen et al., 2005). The whole data was randomly divided to 10 mutually exclusive and approximately equal size subsets. Because the whole dataset contains 257 patient data and 100 healthy data, each fold was to consist of 36 samples. Among these 36 samples, 26 were taken from recordings from patients while the remaining 10 samples were of healthy persons. The classification algorithm trained and tested 10 times. In each case, one of the folds is taken as test data and the remaining folds are added to form training data. Thus 10 different test results exist for each training-test configuration. The average of these 10 results gives the test
1867
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868 Table 2 Average of classification accuracies for each disease. Database
Method
Our previous study
MLNN MLNN MLNN MLNN PNN LVQ GRNN RBF
This study
AIS
with with with with
Classification accuracy (%)
BPwM (one hidden layer) BPwM (two hidden layers) LM (one hidden layer) LM (two hidden layers)
Tuberculosis
COPD
Pneumonia
Asthma
Lung cancer
Normal
Average
84.00 84.00 84.00 90.00 88.00 84.00 86.00 86.00
84.51 87.32 87.32 88.73 88.73 84.51 83.10 87.32
88.33 90.00 91.67 90.00 88.33 86.67 88.33 90.00
88.64 90.91 88.84 90.91 90.91 86.37 86.37 88.64
87.50 87.50 93.75 90.63 93.75 93.75 84.38 90.63
96.00 98.00 95.00 96.00 99.00 94.00 95.00 95.00
89.08 90.76 90.48 91.60 92.16 88.52 88.24 90.20
90.00
92.96
93.33
90.91
93.75
98.00
93.84
accuracy of the algorithm (Delen et al., 2005; Er, Yumusak, et al., 2010). 3. Results The classification accuracy obtained by artificial immune system structure for chest diseases is presented in Table 2. The table shows classification accuracies for each chest diseases and average values by our previous study and this study. From Table 2, it can be seen easily that, the best result for tuberculosis was obtained using MLNN with LM (two hidden layers) and AIS (90%). For COPD (92.96%) and pneumonia (93.33%), the best results were obtained using AIS. For asthma, the best result was obtained using PNN, MLNN with LM (two hidden layers), MLNN with BP (two hidden layers) and AIS (90.91%). For lung cancer, the best result was obtained using PNN, MLNN with LM (two hidden layers), MLNN with BP (two hidden layers) and AIS (93.75%). In this study, the best result for the average classification accuracy was obtained using AIS structure as seen in the same table. The second best result for the classification accuracy was obtained using PNN. The third best result for the classification accuracy was obtained using MLNN with LM (two hidden layers). The classification accuracy performances of MLNN with BP (two hidden layers) and with LM (one hidden layer) were similar and closer to that of RBF. 4. Conclusions The first aim of this study was to develop a diagnostic system leading to more effective usage of the AIS and so to advance the research of chest disease. The second objective was to use AISs in a real world medical classification system and so to show the effectiveness of this Artificial Intelligence field in this problem domain. This paper presents a comparative study for the realization of the chest diseases diagnosis using artificial immune system. There have been several studies reported focusing on chest diseases diagnosis using artificial immune system structures. These studies have applied different structures to the various chest diseases diagnosis problem using their various dataset. Because of the different dataset used by the studies, the direct comparison of the results was impossible. So, this artificial immune system was compared with our previous study on neural networks using the same dataset which consists of the 38 features. As the conclusion, the following results can be summarized; The best results for the average classification accuracy were obtained using AIS for the chest disease diagnosis problem. This classification accuracy is highly reliable for such a problem because only a few samples were misclassified by the system.
This system is capable of conducting the classification process with a good performance to help the expert while deciding the healthy and patient subjects. And, it was obtained that artificial immune system structure could be successfully used to help diagnosis of chest disease. So, this structure can be helpful as learning based decision support system for contributing to the doctors in their diagnosis decisions.
References Aliferis, C. F., Hardin, D., & Massion, P. P. (2002). Machine learning models for lung cancer classification using array comparative genomic hybridization. In Biomedical informatics, Proceedings of the AMIA Symposium 2002:7–11. Nashville, TN, USA: Vanderbilt University. Andrews, P. S., & Timmis, J. (2005). Inspiration for the next generation of artificial immune systems. Lecture Notes in Computer Science, 3627, 126–138. Ashizawa, K., Ishida, T., MacMahon, H., Vyborny, C. J., Katsuragawa, S., & Doi, K. (2005). Artificial neural networks in chest radiography: Application to the differential diagnosis of interstitial lung disease. Academic Radiology, 11(1), 29–37. Baemani, M. J., Monadjemi, A., & Moallem, P. (2008). Detection of respiratory abnormalities using artificial neural networks. Journal of Computer Science, 1549-3636, 4(8), 663–667. Castro, P. D., Coelho, G. P., Caetano, M. F., & Von Zuben, F. J. (2005). Designing ensembles of fuzzy classification systems: An immune-inspired approach. Lecture Notes in Computer Science, 3627, 469–482. Celli, B. R., & MacNee, W. (2004). Standards for the diagnosis and treatment of patients with COPD: A summary of the ATS/ERS position paper. European Respiratory Journal, 23(6), 932–946. Coppini, G., Miniati, M., Paterni, M., Monti, S., & Ferdeghini, E. M. (2007). Computeraided diagnosis of emphysema in COPD patients: Neural-network-based analysis of lung shape in digital chest radiographs. Medical Engineering & Physics, 29, 76–86. Dasgupta, D. (2006). Advances in artificial immune systems. EEE Computational Intelligence Magazine. USA: University of Memphis, November (pp. 40–49). De Castro, L. N., & Von Zuben F. J. (2000). The clonal selection algorithm with engineering applications. In GECCO 2000, Las Vegas, Nevada, USA, July 8. De Castro, L. N., & Timmis, J. (2002). A novel approach to pattern recognition. In L. Alonso, J. Corchado, & C. Fyfe (Eds.), Artificial neural networks in pattern recognition (pp. 67–84). University of Paisley. De Castro, L. N., & Von Zuben, F. J., (2002). Learning and optimization using the clonal selection principle. In the special issue on artificial immune systems of the Journal IEEE Transactions on Evolutionary Computation, June 6(3). Delen, D., Walker, G., & Kadam, A. (2005). Predicting breast cancer survivability: A comparison of three data mining methods. Artificial Intelligence in Medicine Artificial Intelligence in Medicine, 34(2), 1127–1163. El-Solh, A. A., Hsiao, C-B., Goodnough, S., Serghani, J., & Grant, B. J. B. (1999). Predicting active pulmonary tuberculosis using an artificial neural network. Chest, 116, 968–973. Enarson, D. A., Rieder, H. L., Arnadottir, T., & Trébucq, A. (2000). 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. Engin, O., & Döyen, A. (2004). Artificial immune systems and applications in industrial problems. Gazi University Journal of Science, 17(1), 71–84. Er, O., & Temurtas, F. (2008). A study on chronic obstructive pulmonary disease diagnosis using multilayer neural networks. Journal of Medical Systems, 32(5), 429–432. doi:10.1007/s10916-008-9209-x. Er, O., Sertkaya, C., Temurtas, F., & Tanrikulu, A. C. (2009). A comparative study on chronic obstructive pulmonary and pneumonia diseases diagnosis using neural networks and artificial immune system. Journal of Medical Systems, 33(6).
1868
O. Er et al. / Expert Systems with Applications 39 (2012) 1862–1868
Er, O., Temurtas, F., & Tanrikulu, A. C. (2010). Tuberculosis disease diagnosis using artificial neural networks. Journal of Medical Systems, 34(3). doi:10.1007/ s10916-008-9241-x. Er, O., Yumusak, N., & Temurtas, F. (2010). Chest diseases diagnosis using artificial neural networks. Expert Systems with Applications, 55, 555. doi:10.1016/ j.eswa.2010.04.078. Global Action Plan for the Prevention and Control of Pneumonia (GAPP) (2007). Report of an informal consultation, La Mainaz, Gex, France, 5–7 March 2007. Gulbag, A., & Temurtas, F. (2006). A study on quantitative classification of binary gas mixture using neural networks and adaptive neuro fuzzy inference systems. Sensors and Actuators B: Chemical, 115, 252–262. Hamaker, J. S., & Boggess, L. (2004). Non-euclidean distance measures in AIRS, an artificial immune classification system. In Proceedings of the 2004 congress on evolutionary computing. Hanif, N. H. H. M., Lan, W. H., Daud, H. B., & Ahmad, J. (2009). Classification of Control Measures for Asthma using Artificial Neural Networks. Malaysia: ACTA Press. Heckerling, P. S. (2002). Parametric receiver operating characteristic (ROC) curve analysis using Mathematica. Computer Methods and Programs in Biomedicine, 69, 65–73. Jeffery, P. K. (1998). Structural and inflammatory changes in COPD: A comparison with asthma. Thorax, 53(2), 129–136. MedHelp (2009). Thoracic organs.
Polat, K., Sahan, S., Kodaz, H., & Gunes, S. (2005). A new classification method for breast cancer diagnosis: Feature selection artificial immune recognition system (FS-AIRS). Lecture Notes in Computer Science, 3611, 830–838. Polat, K., Sahan, S., & Gunes, S. (2006). A new method to medical diagnosis: Artificial immune recognition system (AIRS) with fuzzy weighted pre-processing and application to ECG arrhythmia. Expert Systems with Applications, 31(2), 264–269. Royal College of Physicians of London (2006). Tuberculosis. 11 St. Andrews Place, London NW1 4LE. ISBN: 1 86016 277 0. Sahan, S., Polat, K., Kodaz, H., & Gunes, S. (2005). The medical applications of attribute weighted artificial immune system (AWAIS): Diagnosis of heart and diabetes diseases. Lecture Notes in Computer Science, 3627, 456–468. Santos, A. M. dos, Pereira, B. B., & Seixas, J. M. de, (2004). Neural networks: An application for predicting smear negative pulmonary tuberculosis. In Proceedings of the statistics in the health sciences, March. Sertkaya, C. (2009). Immune base system in computer security. Master Thesis. Sakarya University, Institute of Science and Technology. Sönmez, S., & Uzaslan, E. (2006). Kronik Obstrüktif Akcig˘er Hastalıg˘ı’nın Genetig˘i ve Sitokin Gen Polimorfizmi – Derleme. Archives of Lung, Cilt, 7, 75–78. Temurtas, F. (2009). A comparative study on thyroid disease diagnosis using neural networks. Expert Systems with Applications, 36, 944–949. Trojanowski, K., & Wierzchon, S. T. (2002). Searching for memory in artificial immune system. In The eleventh international symposium on intelligent information systems, June 3–6. Watkins, A. (2001). AIRS: A resource limited artificial immune classifier. Master Thesis, Mississippi State University. World Health Organization (2008). Global Tuberculosis Control 2008 Surveillance Planning Financing, Geneva. ISBN: 978 92 4 156354 3.