【作者】 温变珍；

【导师】 李佩珍；

【作者基本信息】 山西医科大学 ， 流行病与卫生统计学， 2010， 硕士

【摘要】 目的探讨大肠癌患者根治术后生存的影响因素；利用BP神经网络预测大肠癌患者根治术后的生存时间；比较BP神经网络和Logistic回归模型在大肠癌根治术后生存中的性能。方法采取查阅病案的方法收集资料；基因表达产物的检测采用SP免疫组化法；Cox比例风险回归模型筛选大肠癌根治术后生存的影响因素；通过比较大肠癌患者根治术后的实际生存时间和BP神经网络预测生存时间,判断BP神经网络预测生存时间的效果；通过ROC曲线下面积的比较,判断BP神经网络和Logistic回归模型在大肠癌患者根治术后生存中的预测性能。结果Cox多因素分析结果显示：影响大肠癌患者根治术后生存的因素：Dukes’分期(β=1.197,P<0.001,RR=3.309)、P16(β=-0.805,P<0.001,RR=0.447)、MMP-9(β=0.459,P=0.017,RR=1.582)、P53(β=1.799,P<0.001,RR=6.042)、nm23-H1(β=-0.740,P<0.001,RR=0.477)；BP神经网络预测大肠癌患者根治术后的生存时间和实际生存时间差别没有统计学意义(t=0.996,P=0.327,R2=0.663)；BP神经网络在模型拟合和预测方面性能均优于Logistic回归模型,模型拟合方面,两者ROC曲线下面积比较：Z≈1.75,P=0.04006,按(?)=0.05的检验水准,两者判别效果的差别有统计学意义,BP神经网络的模型拟合效果优于Logistic回归模型,预测性能比较：BP神经网络预测的正确判断率92.5%(37／40),Logistic回归模型的正确判断率82.5%(33／40)。结论Dukes分期、P16、MMP-9、P53、nm23-H,是大肠癌患者根治术后生存的重要影响因素,它们可以为临床医生正确判断患者预后、选择合适的治疗方案提供一定的参考依据。BP神经网络可以预测大肠癌患者根治术后的生存时间,BP神经网络为生存时间的预测提供了一个新思路；在大肠癌根治术后的预后研究中,BP神经网络的性能优于Logistic回归模型,医学现象错综复杂,自变量与因变量间可能存在着复杂的非线性关系以及变量之间可能存在交互作用等,传统的统计学方法对资料有一定的限制,BP神经网络对资料限制少,且具有良好的非线性处理能力等优点,在医学领域可以有进一步的推广和应用。更多还原

【Abstract】 Objectives:To explore prognostic influence factors of colorectal cancer; BP neural network was used to predict survival time of patients with colorectal cancer; To cpmpare the forecasting performance between BP neutal network and Logistic Regression in prognostic analysis of colorectal cancer.Method:We adopted consulting disease-cases method to collect data; SP immunohistochemical method was used to detect gene expression products; The Cox proportional hazards regression model was used to detect prognostic influence factors of colorectal cancer; We could evaluate the forecasting performance of BP neural network by comparing the actual survival time and the predicting survival time; Through the comparison of area under the ROC curve between BP neural network and Logistic Regression, we could judge the forecasting performance between the both methods in prognostic analysis of colorectal cancer.Results:Using Cox multivariate analysis,the main prognostic influence factors of colorectal cancer are Dukes stage (β=1.197 P<0.001,RR=3.309), P16(β=-0.805, P<0.001,RR=0.447), MMP-9(β=0.459, P=0.017,RR=1.582), P53(β=1.799, P<0.001,RR=6.042) nm23-H1(β=-0.740, P<0.001,RR=0.477); The diffence between the actual survival time and the predicting survival time by BP neural network is not statistically significant (t= 0.996, P= 0.327, R2= 0.663); BP nueral network is superior to Logistic regression in the both model-fit and prediction. In the model-fit, between the comparison of the area under the ROC curve, Z-1.75, P= 0.04006, the difference is statistically significant according to (?)=0.05. In the prediction, the accuracy rate by BP nueral network is 92.5%(37/40),but the correct rate of Logistic regression is 82.5%(33/40).Conclusions:Dukes stage、P16、MMP-9、P53、nm23-H1 are the main prognostic influence factors of colorectal cancer, They can help clinicians to judge the prognosis of the patients and choose the right treatment options. The BP neural network can predict survival time of patients with colorectal cancer. BP neural network offers a new way to predict the survival time,besides, BP neural network is superior to Logistic regression in the both model-fit and prediction of colorectal cancer. The Medical phenomenon is very complex. Besides,the relation between dependent variable and independent variable may be non-linear or the interaction between variables may exist,and ao on. At the same time, the traditional statistical methods have certain limitations. BP neural network has fewer limits to data and has a good non-linear processing. So it can be applied and promoted in the medical field.更多还原