Application Of Tumor Markers In The Prediction And Prewarning System Of Lung Cancer Based On Data Mining Technique

ObjectiveLung cancer is the leading cause of cancer-related mortality in the world and its morbidity and mortality are increasing year by year in many countries. Detection of early molecular events in the carcinogenic process has been proposed as a potentially useful method to detect lung cancer, but it remains unsatisfactory for clinical application. The data mining technology embodies its predominance in the process of medical data. In this study, the p16, RASSF1A and FHIT promoter methylation status and telomere length were determined by qMSP in peripheral blood DNA, and the emphasis is to evaluate the data mining technology which are based on those biomarkers and clinical charaters.Materials and methods1. A group of200nonsurgical patients were recruited consecutively from the First Affiliated Hospital of Zhengzhou University, from Jan2009to May2010. All cases had no previous history of other cancers or cancer-related treatments. There was no restriction on age, sex, or disease stage for case recruitment. A group of200healthy individuals without previous history of cancer were recruited from the individuals during the same period who went to the hospital for physical examinations. A detailed questionnaire that included information on sex, age and smoking was completed for each participant by a trained interviewer.2. The p16, RASSF1A and FHIT genes promoter methylation status in peripheral blood DNA were evaluated using SYBR Green-based quantitative methylation-specific PCR (qMSP).3. Relative telomere length in peripheral blood genomic DNA was measured by Quantitative PCR.4. The lung cancer patiets and healthy controls were randomly divided into training set (150cases of lung cancer,150cases of healthy controls) and test set (50cases of lung cancer,50cases of healthy controls) at the proportion of3:1. Fisher discrimination, decision tree(DT), support vector machines(SVM) and artificial neural networks(ANN) algorithm were used to establish classification model through training data, then test set data were classified by the models and compared with all the models. Fisher discrimination, CART and SVM were performed on the platform of SPSS Clementine12.0, and ANN algorithm was performed on the platform of Matlab7.1(M-ANN) and SPSS Clementine12.0(C-ANN).5. All statistical analyses were performed using the SPSS12.0statistical package for Windows. Methods of representation and examination were based on the distribution of quantitative data. All statistical tests were two-sided, and the level of statistical significance was set at (?)=0.05.Results1. There were statistically significant differences in the methylation status of p16, RASSF1A and FHIT between the lung cancer cases and controls (p16:P=0.008, RASSF1A:P=0.038, FHIT:P=0.002). When the subjects were categorized into quartiles based on the genes methylation status, the risk of lung cancer was found to increase as methylation status increased (p16:Ptrend=0.002. RASSF1A:Ptrend=0.014. FHIT:Ptrent=0.001). When the median of methylation status was used as the cutoff between high and low methylation status, individuals with high methylation status were at a significantly higher risk of lung cancer than those with low methylation status (p16:adjusted odds ratio=1.597, P=0.028; RASSF1A:adjusted odds ratio=1.551, P=0.039; FHIT:adjusted odds ratio=1.763. P=0.008). In addition, there were no significant correlations between p16, RASSF1A or FHIT methylation status and gender (P>0.05), age (P>0.05), smoking history (P>0.05), histological type (P>0.05) or clinical stage (P>0.05).2. The telomere length in the lung cancer patients was significantly shorter than that in the controls (P<0.001). When the subjects were categorized into quartiles based on the telomere length of lung cancer patients and healthy controls, the risk of lung cancer was found to increase as the telomere length shortened (P<0.001). Compared with the individuals who had long telomere length, individuals who had short telomere length had a significantly increased risk of lung cancer (adjusted odds ratio was3.258,95%confidence interval was from2.118to5.011). Furthermore, the telomere length in the controls was significantly shorter in association with aging (P =0.005).3. The sensitivity, specificity, accuracy, positive prognostic value, negative prognostic value and AUC of Fisher discrimination analysis were80.0%,54.0%,67.0%,63.5%,37.0%and0.670, respectively. The sensitivity, specificity, accuracy, positive prognostic value, negative prognostic value and AUC of C-ANN were78.0%,74.0%,76.0%,75.0%,77.8%and0.760, and among21cases who had clinical stage Ⅰ and Ⅱ, the accuracy was71.4%. The sensitivity, specificity, accuracy, positive prognostic value, negative prognostic value and AUC of M-ANN were80.0%,76.0%、78.0%、76.9%、79.2%and0.780, and among21cases who had clinical stage Ⅰ and Ⅱ. the accuracy was71.43%. The sensitivity, specificity, accuracy, positive prognostic value, negative prognostic value and AUC of CART were80.0%、82.0%、81.0%、81.6%、80.4% and0.810, and among21cases who had clinical stage Ⅰ and Ⅱ, the accuracy was76.19%. The sensitivity, specificity, accuracy, positive prognostic value, negative prognostic value and AUC of SVM were82.0%,80.0%,81.0%,80.4%,81.6% and0.810, and among21cases who had clinical stage Ⅰ and Ⅱ, the accuracy was76.19%. The AUC of ANN, CART and SVM were higher than the Fisher discrimination analysis. There was no apparent difference in AUC among ANN, CART and SVM, but the results of ANN proved inconsistent every time, and the results obtained by CART and SVM consistent and are easier to understand.Conclusion1. The detection of p1, RASSF1A and FHIT genes promoter methylation status in peripheral blood DNA may be employed as useful biomarkers for the diagnosis of lung cancer or a panel of epigenetic biomarkers for early warning sign of lung cancer. Shorter telomere length may be associated with higher risk of lung cancer and can be used as an early marker for susceptibility to lung cancer.2. The diagnosis and distinguish of lung cancer by data mining models combined with four tumor markers and clinical characters were better than those by Fisher discrimination analysis.