Objective To investigate the potential applications of data mining methods in the diagnosis of digestive tract cancer ( DTC) using several tumor markers( STM) , and to compare the diagnostic performance for DTC with several methods of Logistic regression model, neural network, Bayesian classifier, and clinical diagnosis using a single STM and the combination of STMs. Methods Serum levels of CA19 - 9 , CA242 , CA50 and CEA in 301 patients with DTC and 114 persons with benign digestive disease were used to build diagnostic classifiers based on three data mining methods, including Logistic regression, BP based neural network and Bayesian network. Ten-fold cross validation was employed to test these classifiers. The diagnostic performance was assessed and compared on the basis of sensitivity, specificity and receiver operating characteristic ( ROC ) curve. Results Sensitivity and the area under the ROC curve ( Az) of BP neural network were 92. 0％ and 0. 903, which were greater than the sensitivity of STM parallel diagnosis ( 83. 4％ , P ＜ 0. 001 ) and Az value of CA19 - 9 ( 0. 806 , P ＜ 0. 001 ) , respectively, while the specificity ( 69. 3％ ) was similar with that of STM parallel diagnosis (68. 4％ , P = 1. 00) . Logistic regression model had a higher sensitivity of 91. 4％ than that of STM parallel diagnosis ( P ＜ 0. 001) , a lower specificity of 45. 6％ than that of STM parallel diagnosis (P ＜ 0. 001) , and an similar Az value of 0. 819 with that of STM parallel diagnosis (P = 0. 55 ). The sensitivity of Bayesian classifier was 72. 8％ , which was less than that of STM parallel diagnosis (P ＜ 0. 001) , and the specificity (75.4％ ) and the Az (0. 797) were similar with those of STM parallel diagnosis and CA19 - 9 (P = 0. 13 and P = 0. 61 ) , respectively. Conclusions BP neural network had higher diagnostic accuracy than the parallel diagnosis of the four tumor markers. Logistic regression and Bayesian network had equivalent diagnostic level to the parallel diagnosis of the four tumor markers, and BP neural network has higher diagnostic performance than the other two classifiers.%目的 探讨数据挖掘技术在血清肿瘤标志物(STM)联合检测诊断消化道恶性肿瘤(DTC)中应用的可能性,并比较Logistic回归模型、神经网络和朴素贝叶斯分类器及临床单一及联合STM诊断DTC的性能.方法 对301例DTC和114例消化道良性疾病患者的血清肿瘤标志物CA19-9、CA242、CA50、CEA检测值,分别建立基于统计Logistic回归、反向传播神经网络和朴素贝叶斯方法的诊断分类器,并进行10折交叉验证.利用诊断敏感度、特异度和接受者操作特征(ROC)曲线下面积对三种数据挖掘分类器、CA19-9以及4种STM并联诊断DTC的性能进行评价.结果 神经网络诊断模型的敏感度和ROC曲线下面积(Az)分别为92.0%和0.903,高于STM并联诊断的敏感度83.4%(P＜0.001)和CA19-9诊断的ROC曲线下面积0.806(P＜0.001),特异度69.3%与STM并联诊断的特异度68.4%相当(P=1.00);Logistic回归模型的敏感度91.4%高于STM并联诊断(P＜0.001),特异度45.6%低于STM并联诊断(P＜0.001),Az=0.819与CA19-9诊断相当(P=0.55);贝叶斯分类器的敏感度72.8%低于STM并联诊断(P＜0.001),特异度75.4% 和Az=0.797与STM并联诊断和CA19-9诊断相当(P=0.13和P=0.61).结论 数据挖掘技术的分类方法中,神经网络的分类方法比单一STM及其并联诊断的准确性高,Logistic回归和贝叶斯方法的诊断水平与普通STM并联诊断水平相当;神经网络分类器的诊断性能优于Logistic回归模型和贝叶斯分类器,可进一步应用于计算机辅助诊断中.
展开▼
