Comparison of artificial neural networks with logistic regression in prediction of gallbladder disease among obese patients

Digestive and Liver Disease 39 (2007) 356–362

Liver, Pancreas and Biliary Tract

Comparison of artificial neural networks with logistic regression in prediction of gallbladder disease among obese patients P.-L. Liew a,∗ , Y.-C. Lee b,e , Y.-C. Lin c , T.-S. Lee d , W.-J. Lee e , W. Wang f , C.-W. Chien g b

a Department of Pathology, En-Chu Kong Hospital, Taipei Hsien, Taiwan Graduate Institute of Business Administration, Fu-Jen Catholic University, Hsin-Chuang, Taipei Hsien, Taiwan c Department of Business Administration, Soochow University, Taipei City, Taiwan d Graduate Institute of Management, Fu-Jen Catholic University, Hsin-Chuang, Taipei Hsien, Taiwan e Department of Surgery, Min-Sheng General Hospital, Taoyuan Hsien, Taiwan f Department of Surgery, Taipei Medical University Hospital, Taipei City, Taiwan g Institute of Hospital and Health Care Administration, National Yang-Ming University, Taipei City, Taiwan

Received 1 October 2006; accepted 10 January 2007 Available online 20 February 2007

Abstract Background. Obesity is a risk factor for gallbladder disease. The authors retrospectively analyse the prevalence and risk factors of gallbladder disease using logistic regression and artificial neural networks among obese patients in Taiwan. Methods. Artificial neural networks is a popular technique, which can detect complex patterns within data. They have not been applied to risk of gallbladder disease in obese population. We studied the risk factors associated with gallstones in 117 obese patients who were undergoing bariatric surgery between February 1999 and October 2005. Artificial neural networks, constructed with three-layered backpropagation algorithm, were trained to predict the risk of gallbladder disease. Thirty input variables including clinical data (gender, age, body mass index and associated diseases), laboratory evaluation and histopathologic findings of gallbladder were obtained from the patient records. The result was compared with a logistic regression model developed from the same database. Results. Artificial neural networks demonstrated better average classification rate and lower Type II errors than those of logistic regression. The risk factors from both data mining techniques were diastolic blood pressure, inflammatory condition, abnormal glucose metabolism and cholesterolosis. The biological significance of inflammatory condition in obese population requires further investigation. Conclusion. Artificial neural networks might be a useful tool to predict the risk factors and prevalence of gallbladder disease and gallstone development in obese patients on the basis of multiple variables related to laboratory and pathological features. The performance of artificial neural networks was better than traditional modeling techniques. © 2007 Editrice Gastroenterologica Italiana S.r.l. Published by Elsevier Ltd. All rights reserved. Keywords: Gallbladder disease; Logistic regression; Neural networks; Obesity

1. Introduction Obesity is a pan-endemic health problem. Gallstone disease remains an important public health problem because of its extraordinary frequency. Only a minority of these ∗ Corresponding author at: 3 F, Department of Pathology, No. 399, Fuhsing Road, San-shia Town, Taipei Hsien 237, Taiwan. Tel.: +886 2 26723456x7300; fax: +886 2 26723456x7313. E-mail address: lily liew [email protected] (P.-L. Liew).

patients experience symptoms, which often herald complications such as biliary pain, cholecystitis and acute pancreatitis. Gallstone disease is a female-predominant disease and it is associated with age, obesity, family history of gallstone, hypertriglyceridemia, impaired glucose tolerance or type 2 diabetes mellitus, decreasing cholesterolemia, child bearing, smoking and a secondary life style. Prevalence of gallbladder disease in obese populations has been found to range as 60–95% when evaluated by gross and histologic examination after cholecystectomy [1–3]. Studies suggested that chronic

1590-8658/$30 © 2007 Editrice Gastroenterologica Italiana S.r.l. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.dld.2007.01.003

P.-L. Liew et al. / Digestive and Liver Disease 39 (2007) 356–362

inflammatory changes could occur prior to the appearance of stones. Obese patients not only have a high frequency of gallstones, but also a high proportion of abnormal histologic findings in the gallbladder mucosa. The incidence of obesity in Taiwan is accelerating [4]. The prevalence of gallstone for non-obese Chinese people in Taiwan was 4.3%, which associated with age and diabetes mellitus [5]. Data mining is a systemic method to discover new knowledge and to find the trend, underlying patterns and relationships buried between databases. The methodologies include statistical techniques, data visualization, linkage analysis and matching learning [6,7]. Recently, data mining is widely used in medical research, financial forecasting, marketing strategy, process control, decision support and other related fields [8–15]. Logistic regression is a useful differential and analysis method. It is one of the traditional statistical approaches that is usually used to construct classification tasks and predict the outcome of dichotomous outcomes. The original concept of artificial neural networks (ANNs) is derived from neurobiological models. ANNs are massively parallel, computer-intensive and data-driven algorithmic system that is composed of multitude of highly interconnected nodes (neurons). Each elementary node of a neural network is able to receive an input from external sources, according to the relative importance and different weight, which transforms into an output signal to other nodes by different activation function. Considering the interactions of linked nodes, an output obtained from one node can serve as an input for other nodes, and the conversion of inputs into outputs is activated by virtue of certain transforming function that is typically monotone. The specified working function depends on parameters determined for the training set of inputs and outputs. The network architecture is the organization of nodes and the types of connections permitted. The nodes are arranged in a series of layers with connections between nodes in different layers, but not between nodes in the same layer. The ANNs consisted of one input layer, one output layer and one or more hidden layers [16]. Statistically significant laboratory parameters, as determined by means of univariate analysis, were used as input variables for an input layer of ANNs. The number of neurons of the output layer represents the result of the study. The hidden layer, which was connected to the input layer and output layer, encoded a nonlinear sigmoid function that transferred function for each of the neurons. ANNs can be classified into feedforward and feedback networks categories. The nodes in feedforward networks can take inputs only from the previous layer and send outputs to the next layer. During the training process, the connection weights between the neurons were adjusted by using a back-propagation updating algorithm (BPN). BPN is a kind of widely used feedforward networks that is included within the supervised learning network, essentially using a gradient descent-training algorithm. For the gradient descent-training algorithm, the step size, called the learning rate, must be specified first. To decide when to terminate the training process so

357

as to achieve optimum performance and avoid overtraining, a stopping criterion was established. The learning rate is crucial for BPN since smaller learning rates tend to slow down the learning process before convergence, while larger ones may cause network oscillation and are unable to converge. To our knowledge, there are currently no data on the use of artificial neural networks for risk prediction of gallbladder disease and gallstone of obese patients in Asian population. The aim of this study was to retrospectively compare the predictive accuracy of logistic regression and artificial neural networks with respect to the clinicopathologic features of gallbladder disease among obese patients.

2. Materials and methods 2.1. Study design and patient selection The retrospective study was performed with approval of the ethic committee of the En-Chu Kong Hospital. From 1999 to 2005, an extensive preoperative and perioperative data collection was evaluated on 117 obese patients who underwent concomitant cholecystectomy during weight reduction surgery. Written informed consent was obtained from all patients who agreed to undergo weight reduction surgery. Ultrasound of the gallbladder was performed if the patients had symptoms that were suggestive of biliary disease. The preoperative assessment included a clinical and familial assessment, anthropometric measurements and laboratory tests. Laboratory tests included blood count tests, liver function tests, fasting lipid profiles, fasting glucose profile, fasting insulin, uric acid, high-sensitivity C-reactive protein (hsCRP), haemoglobin A1c (HbA1c) and HOMAIR = (insulin × glucose)/22.5. 2.2. Pathological assessment Sections of gallbladder specimens were examined and histologic parameters graded using haematoxylin–eosin stain. Each specimen was blindly interpreted with one pathologist (P-L L). Parameters included (a) degree of acute inflammation (i.e. epithelial and stromal neutrophil infiltraion), (b) chronic inflammation (mononuclear cell infiltration), (c) cholesterosis, (d) presence of cholesterol polyp, (e) gastric metaplasia and (f) eosinophils infiltration. 2.3. ANN background and theory Articifial neural networks are composed of a series of computational nodes structured into several layers. We used three-layered feedforward design in this study. The first layer is an input layer in which each node represents an input variable; the second layer is a “hidden” computational layer, and the third layer is a single output node representing the outcome for a given data record (i.e., presence of gallstone). Two user-dependent parameters that affect training are the learn-

358

P.-L. Liew et al. / Digestive and Liver Disease 39 (2007) 356–362

ing rate and momentum. The learning rate defines the amount of the weights that is changed during each iteration. A larger learning rate leads to a larger weight change. Momentum allows the weight change to be proportional to the previous weight change. We used a sigmoid transfer function and backpropagation learning algorithm in our study. During training, outcome is known for each input data record. When the output node for the artificial neural network is calculated, the value is compared with the known output from the training data. If there is a discrepancy between the calculated and actual output, the error is propagated through the artificial neural network and the weights are adjusted in an iterative manner until the error discrepancy is within the allowed tolerance. Once the training is completed, the neural network can be externally validated on data for which only input values are provided. 2.4. Model development ANNs were constructed by using Qnet97 (Vesta Services Inc., 1998) and conformed to three-layered perception architecture. The calculation method was used in a personal computer (Intel Pentium III, CPU 933 MHz). We studied 117 obese patients with commitment cholecystectomy during bariatric operation. Twelve patients presented with gallstone. We used 70% (training sample)/30% (testing sample) to analyse our database. The training group included 74 patients without gallstone and 8 patients with gallstone. And then we compared performance by using the crossvalidation method. The hidden layer, which was connected to the input layer, consisted of 30 neurons. A nonlinear sigmoid function was used as a transfer function for each of the neurons in the hidden and output layers of the networks. 2.5. Statistical analysis Data were expressed as mean ± S.D. and percentages. The analyses methods were performed using logistic regression and ANN. Logistic regression was used to assess the significant of associations between ordinal or continuous predictors’ variables. A P-value of less than 0.05 was considered statistically significance. The SPSS statistical software (SPSS Inc., Chicago, IL) was used for statistical analysis.

Fig. 1. Chronic cholecystitis with chronic inflammatory infiltrate. (haematoxylin and eosin stain; original magnification ×100).

cell infiltration (Fig. 1). Cholesterolosis (Fig. 2) with foamy macrophages aggregation was present in 56 patients (47.9%), gastric metaplasia was present in 23 patients (19.7%), cholesterol polyp was present in 11 patients (9.4%) and 5 patients (4.3%) had acute inflammation. We used forward stepwise selection procedure to predict the presence of gallstone. According to the 30 input variables (Table 1), only acute inflammation (P = 0.005), fasting blood sugar (P = 0.002) and cholesterolosis (P = 0.022) were independently associated with gallstone. Table 2 showed that four patients with gallstone were misclassified as non-gallstone category. The average correct classification rate was 88.2%. 3.2. ANNs models In our study, 30 laboratory and pathological assessment variables were assessed as input layer. As reported by Cybenko [17] and Hornik et al. [18], one-hidden-layer network was sufficient to model any complex system with any

3. Results 3.1. Demographics, univariate analysis and logistic regression model We studied 117 patients consisting of 11 men and 106 women. The mean age was 35 ± 9 years, and the mean body mass index (BMI) was 35.9 ± 6.3 kg/m2 . Gallstones were detected in 12 cases (10.3%). On pathology examination, the resected gallbladder specimens from all obese patients showed variable degrees of chronic inflammatory

Fig. 2. Cholesterolosis with foamy macrophages aggregation. (haematoxylin and eosin stain; original magnification ×100).

P.-L. Liew et al. / Digestive and Liver Disease 39 (2007) 356–362

359

Table 1 Input variables Gender [F/M] Age (years) BMI (kg/m2 ) Waist circumference (cm) Hip circumference (cm) Waist: hip ratio Systolic blood pressure (mmHg) Diastolic blood pressure (mmHg) Fasting blood sugar (mg/dl) Total cholesterol (mg/dl)

Triglyceride (mg/dl) Uric acid (mg/dl) AST (IU/L) ALT (IU/L) Albumin (g/dl) WBC (103 /ul) Haemoglobin (g/dL) Mean corpuscle volume (fl) Insulin (IU/ml) hsCRP (mg/L)

Total protein HDL-C (mmol/l) HbA1C (%) HOMA – IR (%) Acute inflammation Chronic inflammation Eosinophil Cholesterolosis Cholesterol polyp Gastric metaplasia

HDL-C, high-density lipoprotein cholesterol; HbA1c, haemoglobin A1c; hsCRP, high-sensitivity C-reactive protein; WBC, white blood cell. Table 2 Diagnostic results using logistic regression Actual class

Classified class 0 (no gallstone)

1 (with gallstone)

0 (no gallstone) 1 (with gallstone)

30 (100%) 4 (100%)

0 (0%) 0 (0%)

Average correct classification rate:

88.2%

desired accuracy; we used one hidden layer as the designed network model. The initial number of neurons in hidden layer consisted of 58, 59, 60, 61 and 62 enrolled within the training process. The presence of gallstone represented the output layer. Rumelhart el al. [16] suggested that lower learning rates could provide a better network results and predictive rate. We Table 3 BPN model prediction results Number of hidden nodes

Learning rates

Training RMSE

Testing RMSE

58

0.006 0.007 0.008 0.009 0.010

0.092992 0.089911 0.089606 0.087882 0.087410

0.267698 0.276507 0.276656 0.280357 0.289654

59

0.006 0.007 0.008 0.009 0.010

0.092521 0.091024 0.088354 0.087951 0.087002

0.268198 0.272326 0.278896 0.283216 0.288395

60

0.006 0.007 0.008 0.009 0.010

0.093794 0.090620 0.089269 0.087968 0.083068

0.266941 0.273812 0.277438 0.282413 0.299519

61

0.006 0.007 0.008 0.009 0.010

0.092429 0.087541 0.088751 0.088012 0.087114

0.270074 0.274541 0.277893 0.282327 0.287407

0.006 0.007 0.008 0.009 0.010

0.093104 0.091033 0.088928 0.087600 0.087172

0.268528 0.274763 0.282733 0.285470 0.286154

62

Fig. 3. The RMSE history of (30-60-1) network during training process.

used five sets of learning rate in our study, which included 0.006, 0.007, 0.008, 0.009 and 0.010 during the training process. The convergence criteria used for training are a root mean squared error (RMSE) less than or equal to 0.0001 or a maximum of 4000 iterations. The network topology with the minimum testing RMSE was considered as the optimal network topology. As reported by Vellido et al. [19], BPN was used in more than 75% of application methods for clinical diagnostic model. Table 3 summarized the results of the BPN networks with combinations of different hidden nodes and learning rates. The (30-60-1) topology, represented the number of neurons in the input layer, number of neurons in the hidden layer and number of neurons in the output layer, with a learning rate of 0.010, gave the best result (minimum testing RMSE). To examine the convergence characteristics of the proposed neural networks model, the RMSE during the training process for the (30-60-1) network with learning rate of 0.010 were shown in Fig. 3. The diagnostic results using the BPN model were summarized in Table 4. We could observe that the average correct classification rate was 97.14%, with only one patient misclassified as non-gallstone category. Fig. 4 demonTable 4 Diagnostic results using BPN Actual class

Classified class 0 (No gallstone)

0 (no gallstone) 1 (with gallstone)

31 (100%) 1 (25%)

Average correct classification rate:

1 (with gallstone) 0 (0%) 3 (75%) 97.14%

360

P.-L. Liew et al. / Digestive and Liver Disease 39 (2007) 356–362 Table 5 Type I and Type II errors of ANN and logistic regression models Diagnostic models

BPN Logistic regression

Fig. 4. The contribution percentage of 30 input variables: Input node

Percent contribution

Gender Age BMI Waist circumference Hip circumference Waist: hip ratio5.25 SBP DBP Sugar CHO TG UA AST ALT Alb WBC Haemoglobin MCV Insulin hsCRP Total protein HDL-C HbA1C HOMA Acute inflammation Chronic inflammation Eosinophil Cholesterolosis Cholesterol polyp Gastric metaplasia

3.5 1.5 2 0.25 1.8 5.25 3.25 11.5 1.8 1 2.5 1.7 1.6 1.75 4 2.8 2.5 4.5 1.8 0.5 3.5 2.5 10.5 2.5 5.5 10.25 1.5 3.75 3.45 3.3

strated the predictive value of different variables within the input layer. Diastolic blood pressure, chronic inflammation and HbA1c were the significantly different variables using ANNs model. 3.3. Type I, Type II errors of the constructed models In order to evaluate the overall classification capability of the designed diagnostic models, the misclassification costs have to be taken into account. The costs associated with Type I error (a patient without gallstone was misclassified as a patient with gallstone) and Type II error (a patient with

Performance assessment Type I error (%)

Type II error (%)

0 0

25 100

gallstone was misclassified as a patient without gallstone) is significantly different. Because of the misclassification costs associated with Type II errors are higher than Type I errors, we should pay special attention to Type II errors in order to evaluate the overall diagnostic capability. Table 5 summarized the Type I and Type II errors of BPN and logistic regression in the present study. BPN had the lower Type II error than logistic regression. We concluded that BPN not only has better average classification rate, but also has lower Type II errors and could successfully reduce the possible risks due to the high misclassification costs associated with Type II errors. 4. Discussion In this study, we developed artificial neural network for the prediction of gallstone and to compare its predictive accuracy with conventional logistic regression analysis. Our constructed ANNs model showed high performance in terms of predicting gallstone in obese patients on the basis of its use of multiple variables related to laboratory tests and histopathologic findings. The average correct classification rate of ANNs was higher than that of the traditional logistic regression approach (97.14% versus 88.2%). Besides, ANNs also had a lower Type II error when compared with logistic regression. The risk factors associated with gallstones formation using ANNs model were diastolic blood pressure, chronic inflammation and HbA1c. Obesity is a chronic inflammatory condition and strongly linked to raised levels of pro-inflammatory cytokines [20]. Visceral fat, adipocytes and liver are main regulator sites for the process of inflammation [21], and atherosclerosis is well-known to be an inflammatory response. Mendez-Sanchez et al. [22] concluded that patients with gallstone disease had an increased risk to have coronary heart disease, which frequently involved as part of the metabolic syndrome. The relationship between gallstone disease, inflammation, endothelial dysfunction and increased diastolic blood pressure in obese population is interesting. Further studies are needed to clarify the role of inflammatory cytokines in the pathogenesis of these metabolic abnormalities in obese patients. Another important finding in our study was the association of HbA1c and fasting blood sugar with the risk factor of gallstone development in obese population. A previous study [23] concluded that in patients with type 1 or type 2 diabetes mellitus, the prevalence of gallstone disease was significantly related to age, body mass index and a family history of gallstone disease. Gallbladder motility could be

P.-L. Liew et al. / Digestive and Liver Disease 39 (2007) 356–362

affected by various clinical conditions, such as obesity, diabetes mellitus and coeliac disease [24]. HbA1c, also known as glycated haemoglobin or glycosylated haemoglobin, indicates a patient’s blood sugar control over the last 2–3 months. A1c is formed when glucose in the blood binds irreversibly to haemoglobin to form a stable glycated haemoglobin complex. Most diabetic patients have a higher average HbA1c level than non-diabetic patients. The prevalence and associated factors of gallstone disease may share insulin resistance as a pathogenic mechanism [25]. Insulin resistance could have a major role in the pathogenesis of gallstone favouring the production of cholesterol-supersaturated bile and altering gallbladder function. The association of gallstone with common metabolic abnormalities, including obesity, diabetes, dyslipidemia and hyperinsulinemia, has supported the hypothesis that gallstone is another member of metabolic syndrome. Abnormalities of glucose and lipid metabolism, which include a prothrombotic and a proinflammatory state, are considered to play an importance role in the pathogenesis of atherosclerosis and cardiovascular diseases. The present study was consistent with the previous reports that histological changes of gallbladder could occur prior to the appearance of gallstones [26,27]. Nearly all obese patients in our study presented with variable degrees of chronic mononuclear cell infiltration in the gallbladder mucosa and 47.9% presented with cholesterolosis. It was interesting to find that both acute and chronic inflammation were significantly associated with gallstones formation. The significance of pathological findings of inflammatory cell infiltration and cholesterolosis with respect to the development of gallbladder disease and gallstone is not clear. A previous prospective study in Taiwan [28] suggested that chronic liver disease, particularly liver cirrhosis was a risk factor for cholecystolithiasis. Taiwan is an endemic area for hepatitis B virus (HBV) infection, with 15–20% of its population being hepatitis B surface antigen (HBsAg) carriers. The true prevalence of gallbladder disease and gallstones in obese patients with chronic liver disease is not well-established. We did not enroll the laboratory variable of HbsAg or hepatitis B antibody in our study due to selection bias. Further prospective studies allowing adequate training and independent testing, between the roles of hepatitis B virology factor and the progression of liver disease, to assess the performance of ANNs for prediction of the gallstone formation in obese population are needed. ANNs represent a computer-based method that can identify patterns of variables that predict an outcome in various fields of medicine. They provide a nonlinear approach to data analysis and have been used to model clinical data with results comparable with traditional modeling techniques [29]. ANNs undergo training sessions in which a number of measurements for each example of a training set, prediction value and the desired classification are fed to the network. The networks learn to associate the training examples with the given classification for each case and to predict the outcome. The limitations of classical statistical techniques include

361

variation homogeneity, autocorrelation of residual, linearity and time costing. Comparing with logistic regression, ANNs have higher predictive rate in complex and non-linear relationships among variables and find a relation between completely unrelated input variables; the critical value in clinical medicine is well established. As a result, we used ANNs in this study to show potentially distributing effects in the data [30–32]. Further analysis of accuracy should be done in order to increase the calibration and discrimination of ANNs. In conclusion, the prevalence of gallstone is lower in Asian obese patients and is significantly associated with increased diastolic blood pressure, inflammatory condition, abnormal glucose metabolism and cholesterolosis. The purpose of this study is to propose the risk factors of gallstone in obese patients by comparing artificial neural networks and logistic regression. Our study shows that artificial neural networks are a better modeling technique and the overall predictive accuracy is superior to traditional modeling techniques. Further studies were warranted to collect more important variables that will increase the classification accuracies by using other data mining techniques.

Practice points • ANNs demonstrated better average classification rate and lower Type II errors than logistic regression. • The prevalence of gallstone disease is 10.3% in Taiwanese obese population. Increased diastolic blood pressure, inflammatory condition, abnormal glucose metabolism and cholesterolosis were risk factors of gallstone disease. • Gallstone disease is another member of metabolic syndrome. • The prevalence and associated factors of gallstone disease may share insulin resistance as a common pathogenic mechanism. • The significance of pathological findings of inflammatory cell infiltration and cholesterolosis with respect to the development of gallbladder disease and gallstone is not clear.

Research agenda • Abnormalities of glucose and lipid metabolism are considered to play an important role in the pathogenesis of gallstone disease, atherosclerosis and cardiovascular diseases. The mechanism is not clear.

362

P.-L. Liew et al. / Digestive and Liver Disease 39 (2007) 356–362

• The importance of inflammatory cytokines production in metabolic abnormalities, including gallstone disease, should be highlighted in future studies. • The true prevalence of gallbladder disease and gallstones in obese patients with chronic liver disease, including chronic hepatitis or nonalcoholic fatty liver disease, is not wellestablished. • Further prospective studies allowing adequate treatment, such as prophylactic cholecystectomy, during bariatric surgery in obese population are needed. • Further analysis of accuracy should be done in order to increase the calibration and discrimination of ANNs.

Conflict of interest statement None declared.

List of abbreviations HOMA-IR, homeostatic model assessment methodinsulin resistance; ALT, alanine transaminase; AST, aspartate transaminase; BPN, back-propagation neural network.

References [1] Aidonopoulos AP, Papavramidis ST, Zaraboukas TG, Habib HW, Pothoulakis IG. Gallbladder findings after cholecystectomy in morbidly obese patients. Obes Surg 1994;4:8–12. [2] Fobi M, Lee H, Igwe D, Felahy B, James E, Stanczyk M, et al. Prophylactic cholecystectomy with gastric bypass operation: incidence of gallbladder disease. Obes Surg 2002;12:350–3. [3] Dittrick GW, Thompson JS, Campos D, Bremers D, Sudan D. Gallbladder pathology in morbid obesity. Obes Surg 2005;15:238– 42. [4] Lee WJ, Wang W. Bariatric surgery: Asia-Pacific perspective. Obes Surg 2005;15:751–7. [5] Lu SN, Chang WY, Wang LY, Hsieh MY, Chuang WL, Chen SC, et al. Risk factors for gallstones among Chinese in Taiwan. J Clin Gastroenterol 1990;12:542–6. [6] Curt H. The devil’s in the detail techniques: tools, and applications for database mining and knowledge discovery-Part. Intell Software Strategies 1995:1–15. [7] Pass S. Discovering value in a mountain of data. OR/MS Today 1997:24–8. [8] Chen MS, Han J, Yu PS. Data mining: an overview from a database perspective. IEEE Trans Knowledge Data Eng 1996;8:866– 83. [9] Fayyad U, Gregory PS, Smyth P. The KDD process for extracting useful knowledge from volumes of data. Commun ACM 1996;39:27–34.

[10] Cabena P, Hadjinian P, Stadler R, Verhees J, Zanasi A. Discovering data mining from concept to implementation. Upper Saddle River: PrenticeHall, NJ; 1998. [11] Lee G, Sung TK, Chang N. Dynamics of modeling in data mining: interpretive approach to bankruptcy prediction. J Manage Inf Syst 1999;16:63–85. [12] Ngan PS, Wong ML, Lam W, Leung KS, Cheng JCY. Medical data mining using evolutionary computation. Artif Intell Med 1999;16:73–96. [13] Pendharkar PC, Rodger JA, Yaverbaum GJ, Herman N, Benner M. Associations statistical, mathematical and neural approaches for mining breast cancer patterns. Expert Syst Appl 1999;17:223–32. [14] Chou SM, Lee TS, Shao YE, Chen IF. Mining the breast cancer pattern using artificial neural networks and multivariate adaptive regression splines. Expert Syst Appl 2004;27:133–42. [15] Polak S, Mendyk A. Artificial intelligence technology as a tool for initial GDM screening. Expert Syst Appl 2004;26:455–60. [16] Rumelhart DE, Hinton DE, Williams RJ. Learning internal representations by error propagation in parallel distributed processing, vol. 1. Cambridge, MA: MIT Press; 1986. pp. 318–62. [17] Cybenko G. Approximation by superpositions of a sigmoidal Function. Mathematical Control Signal Syst 1998;2:303–14. [18] Hornik K, Stinchcombe M, White H. Multilayer feedforward networks are universal approximations. Neural Netw 1989;2:336–59. [19] Vellido A, Lisboa P, Vaughan J. Neural networks in business: a survey of applications (1992–1998). Expert Syst Appl 1999;17:51– 70. [20] Dixon JB, O’Brien PE. Obesity and the white blood cell count: changes with sustained weight loss. Obes Surg 2006;16:251–7. [21] Ziccardi P, Nappo F, Giugliano G, Esposito K, Marfella R, Cioffi M, et al. Reduction of inflammatory cytokine concentrations and improvement of endothelial functions in obese women after weight loss over one year. Circulation 2002;105:804–9. [22] Mendez-Sanchez N, Bahena-Aponte J, Chavez-Tapia NC, MotolaKuba D, Sanchez-Lara K, Ponciano-Radriguez G, et al. Strong association between gallstones and cardiovascular disease. Am J Gastroenterol 2005;100:827–30. [23] Pagliarulo M, Fornari F, Fraquelli M, Zoli M, Giangregorio F, Grigolon A, et al. Gallstone disease and related risk factors in a large cohort of diabetic patients. Dig Liver Dis 2004;36:130–4. [24] Fraquelli M, Pagliarulo M, Colucci A, Paggi S, Conte D. Gallblader motility in obesity, diabetes mellitus and coeliac disease. Dig Liver Dis 2004;35:S12–6. [25] Nervi F, Miquel JF, Alvarez M, Ferreccio C, Garcia-Zattera MJ, Gonzalez R, et al. Gallbladder disease is associated with insulin resistance in a high risk hispanic population. J Hepatology 2006;45:299–305. [26] Csendes A, Burdiles P, Smok G, Csendes P, Burgos A, Recio M. Histologic findings of gallbladder mucosa in 87 patients with morbid obesity without gallstones compared to 87 control subjects. J Gastrointest Surg 2003;7:547–51. [27] Csendes A, Smok G, Burdiles P, Diaz JC, Maluenda F, Kornet O. Histological findings of gallbladder mucosa in 95 control subjects and 80 patients with asymptomatic gallstones. Digestive Dis Sci 1998;43:931–4. [28] Sheen I-S, Liaw Y-F. The prevalence and incidence of cholecystolithiasis in patients with chronic liver diseases: a prospective study. Hepatology 1989;9:538–40. [29] Freeman RV, Eagle KA, Bates ER, Werns SW, Kline RE, Karavite D, et al. Comparison of artificial neural networks with logistic regression in prediction of in-hospital death after percutaneous transluminal coronary angioplasty. Am Heart J 2002;140:511–20. [30] Zhang G, Patuwo BE, Hu MY. Forecasting with artificial neural networks: the state of the art. Int J Forecasting 1998;14:35–62. [31] Dayhoff JE, Deleo JM. Artificial neural network—opening the black box. Cancer 2001;91:1615–35. [32] Paul RR, Mukherjee A, Dutta PK, Banerjee S, Pal M, Chatterjee J, et al. A novel wavelet neural network based pathological stage detection technique for an oral precancerous condition. J Clin Pathol 2005;58:932–8.