中文摘要：

目的比较人工神经网络（ANN）和逻辑回归模型对肝细胞癌患者的诊断价值,以提高肝细胞癌血清诊断效率。方法收集2012年4—12月第二军医大学第三附属医院收治的630例肝病患者的临床资料,其中肝细胞癌患者428例、肝硬化患者139例、肝炎患者63例,根据疾病类型将患者分为肝细胞癌组和非肝细胞癌组。采用单因素分析初步筛选两组中的差异性指标,再进行多因素逻辑回归分析确定最终的差异性指标。指标分别纳入ANN和逻辑回归模型,比较两种模型的诊断效率及与病理诊断结果的一致性。结果经研究筛选确认,最终将甲胎蛋白、总胆汁酸、活化部分凝血活酶时间、碱性磷酸酶和血小板因素分别纳入ANN和逻辑回归模型。两种诊断模型比较,其受试者工作曲线下面积分别为0.92（0.89-0.94）和0.87（0.84-0.90）;两组受试者工作曲线下面积比较,Delong检验Z=3.882,P〈0.001。两种诊断模型截断值分别为0.593和0.647,敏感度分别为82.9%和79.9%,特异度分别为84.7%和81.7%;两种模型与病理结果的一致性指数（Kappa系数）分别为0.68和0.58,准确度分别为83.4%和80.4%。结论在选择相同指标建立肝细胞癌诊断模型时,ANN的诊断效率优于逻辑回归模型,故ANN为肝癌辅助诊断更有效力的模型。

英文摘要：

Objective To compare the diagnostic value between artificial neural network （ANN） and Logistic regression model to predict hepatocellular carcinoma, further optimizing the diagnostic efficiency of hepatocellular carcinoma. Methods All the samples were collected from 630 liver disease patients in the Third Affiliated Hospital of Second Military medical university from Apr. 2012 to Dec. 2012, including 428 hepatocellular carcinoma patients, 139 cirrhosis of the liver patients and 63 hepatitis patients. According to disease type, all the patients were divided into hepatocellular carcinoma group and non-hepatocellular carcinoma group. Firstly, univariate analysis was adopted to preliminarily screen differential markers between hepatocellular carcinoma group and non-hepatocellular carcinoma group, then multivariate Logistic regression analysis was taken to finally confirm differential markers between two groups. Markers were taken into ANN model and Logistic regression model respectively comparing diagnostic efficiency between two models, contrasting diagnostic accuracy of two models to corresponding pathologic diagnosis. Results Confirmed by screening, alpha-fetoprotein, total bile acid, activated partial thrombin time, alkaline phosphatase and platelet factor was finally adopted into ANN model and Logistic regression model. The area under the curve of ANN model Logistic regression model was 0.92 （0.89-0.94） and 0.87 （ 0.84- 0. 90） respectively; The value of Z-test was 3. 882, which meant difference existing between the area under the curve of two models. The cut-off value of two models was O. 593 and 0. 647 respectively, sensitivity of which being 82.9% and 79.9% , specificity of which being 84.7% and 81.7%. The index of conformity of ANN model and Logistic regression model to pathologic diagnosis was 0.68 and 0.58, accuracy was 83.4% and 80.4% , respectively. Conclusion After choosing same markers to establish two different diagnostic models, ANN model revealed higher diagnostic efficiency than that of