A database for hepatocellular carcinoma (HCC) patients who had received hepatic resection was used to develop prediction models for 1-, 3- and 5-year disease-free survival based on a set of clinical parameters for this patient group.
The three prediction models included an artificial neural network (ANN) model, a logistic regression (LR) model, and a decision tree (DT) model. Data for 427, 354 and 297 HCC patients with histories of 1-, 3- and 5-year disease-free survival after hepatic resection, respectively, were extracted from the HCC patient database. From each of the three groups, 80% of the cases (342, 283 and 238 cases of 1-, 3- and 5-year disease-free survival, respectively) were selected to provide training data for the prediction models. The remaining 20% of cases in each group (85, 71 and 59 cases in the three respective groups) were assigned to validation groups for performance comparisons of the three models. Area under receiver operating characteristics curve (AUROC) was used as the performance index for evaluating the three models.
The ANN model outperformed the LR and DT models in terms of prediction accuracy. This study demonstrated the feasibility of using ANNs in medical decision support systems for predicting disease-free survival based on clinical databases in HCC patients who have received hepatic resection.
Citation: Ho W-H, Lee K-T, Chen H-Y, Ho T-W, Chiu H-C (2012) Disease-Free Survival after Hepatic Resection in Hepatocellular Carcinoma Patients: A Prediction Approach Using Artificial Neural Network. PLoS ONE 7(1): e29179. https://doi.org/10.1371/journal.pone.0029179
Editor: Hana Algül, Technische Universität München, Germany
Received: July 25, 2011; Accepted: November 22, 2011; Published: January 3, 2012
Copyright: © 2012 Ho et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This work was in part supported by the National Science Council, Taiwan, Republic of China, under grant numbers NSC 99-2320-B-037-026-MY2 and NSC 95-2314-B-037-079-MY3. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. No additional external funding received for this study.
Competing interests: The authors have declared that no competing interests exist.
Globally, hepatocellular carcinoma (HCC) is among the most prevalent malignant tumors . Of all cancers, HCC has had the highest and second highest mortality rates in males and in females, respectively, since the early 1980s . In Taiwan, the incidence rates of HCC have steadily increased in the past two decades: the respective age-standardized incidence rates for men and women increased from 55.8 and 22.3 per 100,000 in 2002 to 62.1 and 25.6 per 100,000 in 2007 . In 2009, HCC also comprised 38.0% and 14.9% of all cancer-related deaths in men and women in Taiwan, respectively . Hepatic resection is the most common treatment modality for HCC and is among the most effective interventions – for achieving long-term survival. However, even after undergoing hepatic resection, patients with HCC may still have very poor prognoses because of the low survival and high recurrence rates associated with this procedure . Therefore, the aim of this study was to construct an accurate and effective model for predicting disease-free survival in HCC patients who have received hepatic resection. An improved model would enable further development of computerized medical decision support systems for aiding surgeons and healthcare institutions in constructing guidelines for interpreting clinical outcomes. Although previous studies ,  have examined disease-free survival rates at various endpoints, none have evaluated the accuracy of models for predicting disease-free survival after hepatic resection in HCC patients at different endpoints (i.e., 1, 3, and 5 years after resection).
Recently, machine-learning and statistical methods have been applied to develop prediction models for clinical diagnosis and treatment, e.g., artificial neural networks (ANNs), logistic regression (LR) and decision tree (DT) (see, e.g., – and the references therein). Clinical application of these prediction models can potentially improve diagnostic accuracy, treatment decisions, and efficiency in using limited health care resources .
Artificial neural networks have proven particularly effective for nonlinear mapping based on human knowledge and are attracting interest for use in solving complex classification problems , . A multilayer ANN containing layers of simple computing nodes is analogous to brain neural networks that can accurately approximate nonlinear continuous functions and reveal previously unknown relationships between given input and output variables , . Because of their unique structure, ANNs can learn by using algorithms such as backpropagation algorithm and evolutionary algorithm , . Potential medical applications of ANNs include problems in which the relationship between independent variables and clinical outcome are poorly understood . Because ANNs are capable of self training with minimal human intervention, many studies of large epidemiology databases have, in addition to traditional statistical methods, used ANNs for further insight into the interrelationships among variables. However, since few studies have compared performance between ANNs and other modeling techniques such as LR and DT, these interrelationships are still unclear . Our objective was to fill a gap in the current literature by comparing the predictive performance of three modeling techniques so that improved models for predicting 1-, 3- and 5-year disease-free survival can be implemented in knowledge-based computer programs and in medical decision support systems.
This study therefore constructed a database of HCC patients who had received hepatic resection between 2000 and 2007 at either of two hospitals in Kaohsiung, Taiwan: Kaohsiung Medical University Hospital and Yuan's Hospital. The database included demographic, clinical, surgical and outcome data. An ANN model, an LR model, and a DT model were constructed to predict 1-, 3- and 5-year disease-free survival. The three models were based on data for 80% of the cases, which were randomly selected. The remaining 20% of the cases were then used for performance tests of the three models. Predictive accuracy was compared by areas under receiver operating characteristics curve (AUROC) analyses.
Data collection and variable selection
The study population included 482 patients who had received liver resection for HCC and were currently disease-free. The exclusion criteria were any history of the following: (i) liver resection; (ii) treatment with radiofrequency ablation or microwave ablation; (iii) histopathological evidence of benign tumor and/or non-primary liver cancer; (iv) unavailable and/or incomplete medical history; (v) death within thirty days after surgery; (vi) tumor remaining after resection; (vii) incomplete data for key explained variables; and (viii) follow-up data for less than 1 year. Therefore, 427, 354 and 297 patients were classified into the 1-, 3- and 5-year disease-free survival groups, respectively. In each patient, medical records were reviewed by the attending physician. Data collection included demographic data, clinical features, and surgical process and outcome. Ethical approval was provided by Institutional Review Board of the Kaohsiung Medical University Chung-Ho Memorial Hospital (KMUH-IRB-990166). Patients provided written informed consent.
Patients were classified as disease-free hepatic resection survivors if no death or recurrence occurred during the 1-, 3-, or 5-year periods considered in the three survival models. In other words, survival (no event) was defined as disease-free survival after 1, 3, or 5 years. Therefore, presence of an event (death or recurrence) was coded as 1, and absence of an event (disease-free survival) was coded as 0.
First, continuous explanatory variables were transformed into categorical variables to minimize the effects of extreme values and to enhance the computing efficiency of the ANN model. The cut-off points for these variables were based on those used in previous clinical studies , , –. Low and high risk were coded as 0 and 1, respectively. The variables included BUN AST, α-fetoprotein, ALT, total bilirubin, and others. Other recoded items included TNM stage, a common prognostic index of cancer risk or severity, and ASA, a risk score for surgical procedures. The TNM stage ranges from 1 to 6, and ASA score ranges from 1 to 4. Two variables were recoded as 0 for low risk, 1 for medium risk, and 2 for high risk (Table 1). High risk was assumed to increase the probability of recurrence (event). Second, to enhance the calculation efficiency and prediction performance of the ANN models, univariate Cox proportional hazard model was used to test relationships among potential variables. Variables with statistically significant associations (log-rank test, P<0.05) with disease-free survival were retained to construct the ANN models (Table 1). Finally, of the 31 input variables, the 15 statistically significant variables used to construct the ANN models were liver cirrhosis, chronic hepatitis, AST, ALT, total bilirubin, albumin, creatinine, ASA classification, Child-Pugh classification, TNM stage, tumor number, portal vein invasion, biliary invasion, surgical procedure, and post-operative complication. Age and gender were also included as control variables.
Training and validation data sets
From each of the three survival groups, 80% of the cases were assigned to training groups for developing the ANN, LR and DT models, and the remaining 20% were assigned to validation groups for performance tests of the models for predicting 1-, 3-, and 5-year disease-free survival. That is, of the 427 1-year cases, 342 were used for training, and 85 were used for validation; of the 354 3-year cases, 283 were used for training, and 71 were used for validation; of the 297 5-year cases, 238 were used for training, and 59 were used for validation (Table 2). Table 2 shows that (i) the specific data contained in each clinical case were summarized with their descriptive characteristics for 1-, 3-, and 5-year disease-free survival. For example, 245 (71.6%) patients were aged older than 65 years and 97 (28.4%) patients were aged 65 years or younger. In the 1-year training group, 252 (73.7%) patients were male, and 90 (26.3%) patients were female; (ii) at 1-, 3-, and 5 years after the resection procedure, post-resection events (i.e., recurrence or death) had occurred in 155 (36.3%), 226 (63.8%) and 247 (83.2%) patients; and (iii) in all three survival models, the effects of input variables did not significantly differ between training and validation (P>0.05), which confirmed the reliability of the data selection.
The training group data were used to construct an ANN model, an LR model and a DT model. The ANN model included input, hidden, and output layers. Figure 1 shows the three independent ANN models for 1-, 3- and 5-year disease-free survival. The input layer in each of the three models contained 17 neurons: age, gender, liver cirrhosis, chronic hepatitis, AST, ALT, total bilirubin, albumin, creatinine, ASA classification, Child-Pugh classification, TNM stage, tumor number, portal vein invasion, biliary invasion, surgical procedure, and post-operative complication. In the hidden layers, the numbers of neurons were optimized using training and validation data in a trial-and-error process to maximize predictive accuracy , which resulted in 30, 17 and 7 neurons in the 1-, 3- and 5-year models, respectively. The output layer in each of the three models had only one neuron representing the disease-free survival of HCC patients after hepatic resection.
The input layer in each of the three models contained 17 neurons. In the hidden layers, the numbers of neurons were 30, 17 and 7 the 1-, 3- and 5-year models, respectively. The output layer in each of the three models had only one neuron representing the disease-free survival of HCC patients after hepatic resection.
The LR model generates the coefficients for the following formula used for logit transformation of the probability of a patient having a characteristic of interest: . The formula used for calculating the probability of the characteristic of interest in this study, where 1 = disease-free survival status and 0 = non-disease-free survival status.
Because of its easily interpreted decision rules, the DT model with C4.5  was used for classification and regression. In this model, each object in the input dataset belongs to a class. Each object is characterized by a set of attributes (variables or predictors) that may have numerical and categorical (non-numerical) values. The goal of DT is to use a training dataset with known attribute-class combinations for generating a tree structure with a rule set for correctly classifying and predicting a similar test dataset. In addition to its root and internal (non-terminal) decision nodes, a DT has a set of terminal nodes (leaves), each of which represents a class. The rules associated with the DT, from the root to each terminal node (leaf), are easily interpretable for predicting a class. The steps of the learning process are (i) using an impurity function to select the most discriminative variable for data partitioning, (ii) repeating the partitioning until the nodes are sufficiently pure for use as terminal nodes, and (iii) pruning the completed tree to avoid over-fitting .
The software used to construct the ANN and DT models was Waikato Environment for Knowledge Analysis (WEKA) version 3.6.0 . The LR model was constructed using SPSS for Windows version 6.1.
For the training and validation groups, Figs. 2 and 3, respectively, show the receiver operating characteristics (ROC) curves for the 1-, 3- and 5-year disease-free survival models constructed using ANN, LR and DT. Tables 3 and 4 show the respective AUROC curves constructed using the data shown in Figs. 2 and 3. For example, the AUROCs for 1-year models constructed by ANN, LR and DT were 0.977, 0.771 and 0.734, respectively. For the training data and validation data, Tables 3 and 4 show the respective AUROC values, sensitivities and specificities for the 1-, 3- and 5-year disease-free survival models obtained by ANN, LR and DT. In the 1-year model for the training group, for instance, sensitivity and specificity were 0.962 and 0.916 when using ANN, 0.848 and 0.466 when using LR, and 0.948 and 0.458 when using DT, respectively. Notably, in all training groups and in most validation groups sensitivity and specificity for the 1-, 3- and 5-year models constructed using ANN were not only within acceptable limits, but were actually superior to those for models constructed using LR and DT.
The AUROC values for 1-year (A), 3-year (B) and 5-year (C) disease-free survival were 0.977, 0.989 and 0.963 for ANN models, 0.771, 0.751 and 0.769 for LR models, and 0.734, 0.825 and 0.760 for DT models, respectively. In all disease-free survival models for training groups, AUROC values obtained by ANN were superior to those obtained by LR and DT.
The AUROC values for 1-year (A), 3-year (B) and 5-year (C) disease-free survival were 0.777, 0.774 and 0.864 for ANN models, 0.772, 0.725 and 0.736 for LR models, and 0.718, 0.561 and 0.627 for DT models, respectively. In all disease-free survival models for validation groups, AUROC values obtained by ANN were superior to those obtained by LR and DT.
Model sensitivity and specificity are important when testing whether a model can accurately recognize positive and negative outcomes. Sensitivity and specificity must also be measured to determine the proportion of false negatives or false positives produced by a model . Comparing false positive and false negative rates reveals the tendency of a model to misclassify positive patients as negative patients and vice versa . The ideal model has both high sensitivity and high specificity . In the current study, comparisons of predictive performance showed that the LR and DT models had poor sensitivity (<40%) but high specificity (>80%) for predicting 5-year disease-free survival in the training groups (Table 3); the DT model had poor specificity (<40%) but high sensitivity (>80%) for predicting 1-year disease-free survival in the validation groups (Table 4), and the LR and DT models had poor sensitivity (<40%) but high specificity (>80%) for predicting 5-year disease-free survival in the validation groups (Table 4). Specifically, Table 4 shows that the sensitivity values for predictions of 5-year disease-free survival with LR and DT models in the validation groups were zero. The explanation is the occurrence of false positives (i.e., type I error) . That is, the LR and DT models, which had very low sensitivity, could be not used to screen for disease-free survival in HCC patients who had received hepatic resection since they lacked sufficient specificity for identifying true positives. However, sensitivity and specificity remained high in all ANN models (Tables 3 and 4). Since AUROC provides a superior performance index in addition to superior accuracy, AUROC was used to evaluate the predictive accuracy of classifiers . The AUROC of a classifier can be defined as the probability of the classifier ranking a randomly chosen positive example higher than a randomly chosen negative example . Therefore, the higher the AUROC, the higher the predictive accuracy . This study also used AUROC values for performance comparisons of different prediction models. For the training groups, Table 3 shows that the AUROC values for 1-, 3- and 5-year disease-free survival were 0.977, 0.989 and 0.963 for ANN models, 0.771, 0.751 and 0.769 for LR models, and 0.734, 0.825 and 0.760 for DT models, respectively. In the validation groups (Table 4), the respective values were 0.777, 0.774 and 0.864 for ANN models, 0.772, 0.725 and 0.736 for LR models and 0.718, 0.561 and 0.627 for DT models. In all disease-free survival models, AUROC values obtained by ANN were superior to those obtained by LR and DT. Thus, the ANN models outperformed the LR and DT models in terms of predictive accuracy. The ROC curves in Figures 2 and 3 further show that the ANN was consistently more accurate in predicting 1-, 3- and 5-year disease-free survival compared to the LR and DT models, both of which demonstrated inconsistent results. The above comparisons thus confirm that ANN outperforms both LR and DT in predicting disease-free survival in HCC patients who have received hepatic resection.
Even when only seventeen easily obtainable parameters were used, the ANN models developed in this study demonstrated acceptable accuracy. Variables that were not significantly associated with disease-free survival were intentionally omitted when constructing the ANN models. The dependent variable indicates a decision by the lead surgeon in each case to perform a surgical intervention. In predictive mode, however, it can be considered a reliable estimation of confidence in the decision to operate on a specific patient since the ANN models were trained by a large patient database from teaching hospitals with highly qualified surgeons. Moreover, omitting this variable expanded the potential applications of the resultant model to circumstances in which advanced diagnostic.
Yeh et al.  used multiple logistic regression to predict associations between clinicopathologic factors and >5-year survival without recurrence in HCC patients treated with hepatectomy. Ercolani et al.  also evaluated prognostic factors affecting 5-year disease-free survival after liver resection in HCC patients with cirrhosis. However, the above studies ,  focused on survival rates and predictors and did not compare the predictive accuracy of different statistical models. The current study, however, compared different statistical models in terms of accuracy in predicting 1-, 3- and 5-year disease-free survival after hepatic resection in HCC patients. The comparisons revealed that predictive accuracy significantly differed among ANNs, LRs and DTs. To our knowledge, very few studies have compared predictive performance in these three methods. The model comparisons showed that the ANN models of disease-free survival obtained superior AUROC values and have potential applications in decision support systems used to assess the need for hepatic resection in HCC patients.
In conclusion, comparison of prediction models for 1-, 3- and 5-year disease-free survival in HCC patients who have received hepatic resection revealed that the prediction models obtained by ANN machine learning method were superior to those obtained by conventional LR and DT. The AUROC values in the ANN models were generally higher than those in LR and DT models. That is, The ANN model had superior predictive accuracy. Therefore, this study demonstrated the feasibility of applying ANN in medical decision support systems that use clinical databases to predict disease-free survival in HCC patients who have received hepatic resection. Physicians may also consider machine-learning methods as a supplemental tool for clinical decision-making and prognostic evaluation.
The authors thank Professors Jyh-Horng Chou and Hon-Yi Shi for their careful review of the statistical and artificial neural network methods, respectively.
Conceived and designed the experiments: W-HH H-CC. Performed the experiments: T-WH. Analyzed the data: W-HH H-CC. Contributed reagents/materials/analysis tools: K-TL H-YC. Wrote the paper: W-HH H-CC.
- 1. Bosch FX, Ribes J, Borras J (1999) Epidemiology of primary liver cancer. Seminars in Liver Disease 19: 271–285.
- 2. Kao JH, Chen DS (2002) Global control of hepatitis B virus infection. The Lancet Infectious Diseases 2: 395–403.
- 3. Taiwan Cancer Registry website. Available: http://crs.cph.ntu.edu.tw. Accessed 2010.
- 4. Department of Health (Taiwan) website. Available: http://www.doh.gov.tw Accessed 2010.
- 5. Hanazaki K, Kajikawa S, Shimozawa N, Mihara M, Shimada K, et al. (2000) Survival and recurrence after hepatic resection of 386 consecutive patients with hepatocellular carcinoma. J of the American College of Surgeons 191: 381–388.
- 6. Lin XD, Lin LW (2006) Local injection therapy for hepatocellular carcinoma. Hepatobiliary and Pancreatic Diseases International 5: 16–21.
- 7. Lee KT, Lu YW, Wang SN, Chen HY, Chuang SC, et al. (2009) The effect of preoperative transarterial chemoembolization of resectable hepatocellular carcinoma on clinical and economic outcomes. J of Surgical Oncology 99: 343–350.
- 8. Zhang Z, Liu Q, He J, Yang J, Yang G, et al. (2000) The effect of preoperative transcatheter hepatic arterial chemoembolization on disease-free survival after hepatectomy for hepatocellular carcinoma. Cancer 89: 2606–2612.
- 9. Ercolani G, Grazi GL, Ravaioli M, Del Gaudio M, Gardini A, et al. (2003) Liver resection for hepatocellular carcinoma on cirrhosis univariate and multivariate analysis of risk factors for intrahepatic recurrence. Annals of Surgery 237: 536–543.
- 10. Yeh CN, Lee WC, Chen MF, Tsay PK (2003) Predictors of long-term disease-free survival after resection of hepatocellular carcinoma: two decades of experience at Chang Gung memorial hospital. Annals of Surgical Oncology 10: 916–921.
- 11. Li YC, Liu L, Chiu WT, Jian WS (2000) Neural network modeling for surgical decisions on traumatic brain injury patients. Int J of Medical Informatics 57: 1–9.
- 12. MacDowell M, Somoza E, Rothe K, Fry R, Brady K, et al. (2001) Understanding birthing mode decision making using artificial neural networks. Medical Decision Making 21: 433–443.
- 13. Matsui Y, Egawa S, Tsukayama C, Terai A, Kuwao S, et al. (2002) Artificial neural network analysis for predicting pathological stage of clinically localized prostate cancer in the Japanese population. Japanese J of Clinical Oncology 32: 530–535.
- 14. Hanai T, Yatabe Y, Nakayama Y, Takahashi T, Honda H, et al. (2003) Prognostic models in patients with non-small-cell lung cancer using artificial neural networks in comparison with logistic regression. Cancer Science 94: 473–477.
- 15. Lee SM, Kang JO, Suh YM (2004) Comparison of hospital charge prediction models for colorectal cancer patients: neural network vs. decision tree models. J of Korean Medical Science 19: 677–681.
- 16. Chiu JS, Chong CF, Lin YF, Wu CC, Wang YF, et al. (2005) Applying an artificial neural network to predict total body water in hemodialysis patients. American J of Nephrology 25: 507–513.
- 17. Viazzi F, Leoncini G, Sacchi G, Parodi D, Ratto E, et al. (2006) Predicting cardiovascular risk using creatinine clearance and an artificial neural network in primary hypertension. J of Hypertension 24: 1281–1286.
- 18. Heckerling PS, Canaris GJ, Flach SD, Tape TG, Wigton RS, et al. (2007) Predictors of urinary tract infection based on artificial neural networks and genetic algorithms. Int J of Medical Informatics 76: 289–296.
- 19. Luk JM, Lam BY, Lee NPY, Ho DW, Sham PC, et al. (2007) Artificial neural networks and decision tree model analysis of liver cancer proteomes. Biochemical and Biophysical Research Communications 361: 68–73.
- 20. Ghoshal UC, Das A (2008) Models for prediction of mortality from cirrhosis with special reference to artificial neural network: a critical review. Hepatology International 2: 31–38.
- 21. Das A, Ben-Menachem T, Farooq FT, Cooper GS, Chak A, et al. (2008) Artificial neural network as a predictive instrument in patients with acute nonvariceal upper gastrointestinal hemorrhage. Gastroenterology 134: 65–74.
- 22. Samanta B, Bird GL, Kuijpers M, Zimmerman RA, Jarvik GP, et al. (2009) Prediction of periventricular leukomalacia. Part I: Selection of hemodynamic features using logistic regression and decision tree algorithms. Artificial Intelligence in Medicine 46: 201–215.
- 23. Cucchetti A, Piscaglia F, Grigioni AD, Ravaioli M, Cescon M, et al. (2010) Preoperative prediction of hepatocellular carcinoma tumour grade and micro-vascular invasion by means of artificial neural network: a pilot study. J of Hepatology 52: 880–888.
- 24. Trtica-Majnaric L, Zekic-Susac M, Sarlija N, Vitale B (2010) Prediction of influenza vaccination outcome by neural networks and logistic regression. J of Biomedical Informatics 43: 774–781.
- 25. Nguyen T, Malley R, Inkelis SH, Kuppermann N (2002) Comparison of prediction models for adverse outcome in pediatric meningococcal disease using artificial neural network and logistic regression analyses. Journal of Clinical Epidemiology 55: 687–695.
- 26. Eftekhar B, Mohammad K, Ardebili HE, Ghodsi M, Ketabchi E (2005) Comparison of artificial neural network and logistic regression models for prediction of mortality in head trauma based on initial clinical data. BMC Medical Informatics and Decision Making 5: 3.
- 27. Liew PL, Lee YC, Lin YC, Lee TS, Lee WJ, et al. (2007) Comparison of artificial neural networks with logistic regression in prediction of gallbladder disease among obese patients. Digestive and Liver Disease 39: 356–362.
- 28. Tsai JT, Chou JH, Liu TK (2006) Tuning the structure and parameters of a neural network by using hybrid Taguchi-genetic algorithm. IEEE Trans on Neural Networks 17: 69–80.
- 29. Ho WH, Tsai JT, Hsu GM, Chou JH (2010) Process parameters optimization: a design study for TiO2 thin film of vacuum sputtering process. IEEE Trans on Automation Science and Engineering 7: 143–146.
- 30. Ho WH, Chang CS (2011) Genetic-algorithm-based artificial neural network modeling for platelet transfusion requirements on acute myeloblastic leukemia patients. Expert Systems With Applications 38: 6319–6323.
- 31. Ho WH, Chen JX, Lee IN, Su HC (2011) An ANFIS-based model for predicting adequacy of vancomycin regimen using improved genetic algorithm. Expert Systems With Applications 38: 13050–13056.
- 32. Tsai JT, Liu TK, Chou JH (2004) Hybrid Taguchi-genetic algorithm for global numerical optimization. IEEE Trans on Evolutionary Computation 8: 365–377.
- 33. Ho WH, Chou JH, Guo CY (2010) Parameter identification of chaotic systems using improved differential evolution algorithm. Nonlinear Dynamics 61: 29–41.
- 34. Robinson CJ, Swift S, Johnson DD, Almeida JS (2008) Prediction of pelvic organ prolapse using an artificial neural network. American J of Obstetrics and Gynecology 199: 193.e1–193.e6.
- 35. Terrina N, Schmida CH, Griffitha JL, D'Agostino RB, Selkera HP (2003) External validity of predictive models: a comparison of logistic regression, classification trees, and neural networks. Journal of Clinical Epidemiology 56: 721–729.
- 36. Yeh CN, Chen MF, Lee WC, Jeng LB (2002) Prognostic factors of hepatic resection for hepatocellular carcinoma with cirrhosis: univariate and multivariate analysis. J of Surgical Oncology 81: 195–202.
- 37. Ercolani G, Grazi GL, Ravaioli M, Del Gaudio M, Gardini A, et al. (2003) Liver resection for hepatocellular carcinoma on cirrhosis: univariate and multivariate analysis of risk factors for intrahepatic recurrence. Annals of Surgery 237: 536–543.
- 38. Shimozawa N, Hanazaki K (2004) Longterm prognosis after hepatic resection for small hepatocellular carcinoma. J of the American College of Surgeons 198: 356–365.
- 39. Liau KH, Ruo L, Shia J, Padela A, Gonen M, et al. (2005) Outcome of partial hepatectomy for large (>10 cm) hepatocellular carcinoma. Cancer 104: 1948–1955.
- 40. Sasaki A, Iwashita Y, Shibata K, Ohta M, Kitano S, et al. (2006) Preoperative transcatheter arterial chemoembolization reduces long-term survival rate after hepatic resection for resectable hepatocellular carcinoma. European J of Surgical Oncology 32: 773–779.
- 41. Colombet I, Ruelland A, Chatellier G, Gueyffier F, Degoulet P, et al. (2000) Models to predict cardiovascular risk: comparison of CART, multilayer perceptron and logistic regression. Proc AMIA Symp 156–160.
- 42. Witten IH, Frank E (2005) Data Mining: Practical Machine Learning Tools and Techniques. San Francisco, , CA, USA: Morgan Kaufmann Publishers.
- 43. Simon D, Boring JR III (1990) Clinical Methods: The History, Physical, and Laboratory Examinations. Boston, USA: Butterworth Publishers.
- 44. Fawcett T (2006) An introduction to ROC analysis. Pattern Recognition Letters 27: 861–874.
- 45. Ke WS, Hwang Y, Lin E (2010) Pharmacogenomics of drug efficacy in the interferon treatment of chronic hepatitis C using classification algorithms. Advances and Applications in Bioinformatics and Chemistry 3: 39–44.