肿瘤标志对肺癌诊断价值PLS-DA和ANN-MPL模型评估  

PLS-DA and ANN-MPL models to estimate the diagnostic value of four serum tumor markers for lung cancer

作　　者：田刚[1] 周明术[1] 杭永伦[1] 王开正[1] 刘靳波[1] 

机构地区：[1]泸州医学院附属医院检验科,四川泸州646000

出　　处：《中华肿瘤防治杂志》2014年第5期380-383,392,共5页Chinese Journal of Cancer Prevention and Treatment

基　　金：四川省卫生厅资助(100258);泸州医学院自然科学基金(12043);泸州医学院附属医院人才基金(12277)

摘　　要：目的:探讨联合检测血清癌胚抗原(carcinoembryonic,CEA)、神经元特异性烯醇化酶(neuron specific enolase,NSE)、细胞角蛋白19片段(cyfra 21-1)和CA125对肺癌的诊断价值。方法:应用电化学发光免疫分析法,测定2010-01-19-2012-12-24泸州医学院附属医院病理或细胞学确诊70例肺癌和54例肺部其他良性疾病患者血清中CEA、NSE、cyfra21-1和CA125,结合主成分分析(principal component analysis,PCA)法、受试者工作特征曲线(receive operating characteristic curve,ROC)、偏最小二乘判别分析(partial-least-squares discriminant analysis,PLS-DA)线性模型和人工神经网络多层感知(artificial neural network-multiplayer perceptron,ANN-MPL)非线性模型对CEA、NSE、cyfra21-1和CA125进行建模分析。结果:肺癌组血清CEA、NSE、cyfra21-1和CA125水平均高于对照组。肺癌组血清中CEA、NSE和cyfra21-1水平显著升高,差异有统计学意义,P<0.05。PCA模型中肺癌组个体空间分布较分散,而对照组则能较好地聚类,2组个体有分离趋势。PLS-DA模型能够很好地鉴别肺癌组和对照组,P<0.01;具有100%的灵敏度、96.3%的特异性、98.4%的准确性和97.7%的预测能力。ROC曲线分析显示,CEA、cyfra21-1、NSE和CA125的ROC曲线下面积(area under the curve,AUC)分别为0.787(P<0.05)、0.784(P<0.05)、0.563(P<0.05)和0.503(P=0.59)。在ANNMPL-ROC曲线模型中,联合检测CEA和cyfra21-1的AUC为0.849,均优于单一的肿瘤标志,具有74.3%的灵敏度、76.0%的特异性,75.9%准确性和75.0%的预测能力,P<0.05。联合CEA、NSE、cyfra21-1和CA125的AUC为0.880,灵敏度为84.3%,特异性为76.0%,具有80.6%准确性和82.5%的预测能力。结论:ANN-MPL的ROC模型中联合检测血清CEA、NSE、cyfra21-1和CA125与联合检测CEA和cyfra21-1相比灵敏度略有升高而特异性不变,2种模型的诊断效能相当。从经济和临床应用价值考虑,可选择联合检测CEA和cyfra21-1用于肺癌的早期诊断。OBJECTIVE：To evaluate the diagnostic significance of four tumor markers carcinoembryonic（CEA） ,neu- ron specific enolase （NSE）,cyfra 21-1 and CA125 for patients with lung cancer. METHODS： The serum concentration of CEA,NSE,cyfra 21-1 and CA125 was measured with electrochemilurrlinescence immunoassay in 70 patients with lung cancer and 54 patients with benign pulmonary lesions. Data sets were analyzed by the line model of partial-least-squares discriminant analysis（PLS-DA） and the nonlinear model of artificial neural network- multiplayer perceptron（ANN-MPL）. RESULTS： Serum levels of CEA,NSE,eyfra21-1 and CA125 in lung cancer group were higher than that in pulmonary be- nign diseases group,and serum levels of CEA, NSE and cyfra21-1 in lung cancer group were significantly elevated than that in pulmonary benign diseases group（P〈0.05）. In PCA model,each member in lung cancer group was scattered and hard to get together while there was a tendency to get together in control group. The AUC under the ROC of four serum tumor markers was CEA（0. 787,P〈0.05）,cyfra21-1（0. 784,P〈0.05）,NSE（0. 563,P〈0.05） and CA125（0. 503,P〈 0.59）,separately. Further study found that PLS-DA model showed a clear separation between patients with lung cancer and patients with benign pulmonary lesions（P〈0.05）. This model provided 100% of sensitivity,96.3% of specificity and 98.4% of accuracy for the diagnosis of lung cancer and 97.7% power for prediction. In ANN-MPL model,the AUC of joint detection of CEA and cyfra21-1 was 0. 849, which was better than single marker, with 74.3 % sensitivity, 76.0 % spe- cificity, 75.9% accuracy and 75.0% prediction. The AUC under the ROC based on the joint detection of four tumor markers waslarger than that of single tumor marker, providing more powerful diagnostic importance for lung cancer（AUC= 0. 880, P〈0.05）. ANN-MPL model provided 84.3% of sensitivity,76.0% of specificity and 80.6% of accuracy for the diagnosis of lung cancer and 82. 5% power for pr

关 键 词：肺癌 诊断 偏最小二乘判别分析 人工神经网络 受试者工作特征曲线 

分 类 号：R734.2[医药卫生—肿瘤]