原发性结直肠恶性肿瘤根治术后结局的分类预测模型研究

发布时间：2018-04-28 06:11

本文选题：结直肠恶性肿瘤 + 风险预测　；参考：《重庆医科大学》2017年硕士论文

【摘要】：结直肠癌(CRC)是一种常见的消化道恶性肿瘤,发病率与死亡率均呈逐年上升趋势,已然成为影响人类健康的重要疾病。目前CRC的主要治疗手段为外科手术,但术后仍有一定比例的复发转移风险,且复发转移将严重影响患者预后,故是否能准确预测患者术后结局,采取适当措施就显得尤为重要。国内外已有研究对CRC术后转移风险及结局进行预测,主要采用多因素Logistic回归和COX回归,但这2种方法对样本量要求较高,不宜分析小样本高维数据。近年兴起的机器学习算法,如基于统计学习理论的支持向量机学习算法(SVM)、随机森林算法(RF)则适用于小样本、高维数据的分类问题,可以得到具有较好泛化能力的预测模型。预测模型结合特征选择算法,在降低空间冗余度的同时,还可以降低训练成本。其中Boruta算法通过衡量特征对结局的重要性进行选择,可避免变量之间的相关性,对医学数据较为适用,但目前尚未应用于恶性肿瘤数据的预测。本研究分为两部分,第一部分利用UCI标准数据集进行模拟预测,运用SPSS 22.0软件进行单因素差异性分析,R 3.30进行Boruta特征选择,随后建立SVM模型和RF模型,运用Stata14.0对模型预测效果进行比较,尝试筛选出较优的特征选择方式,第二部分针对抽样采集的CRC预后相关数据,对第一部分筛选得到的模型进行验证,并对CRC术后不良结局进行预测。研究结果显示:(1)针对UCI标准数据集的分类预测结果显示,基于全数据集的RF预测模型(AUC=0.717)效果较好;基于差异性分析的预测效果较好的是Polynomial-SVM模型(AUC=0.756);基于Boruta特征选择预测效果较好的为RF模型(AUC=0.905)。ROC曲线分析显示,不同预处理方法的最优预测模型间存在差异(2x=7.27,P=0.026)。(2)结直肠癌单因素分析结果显示,肿瘤部位、CA-199、CEA、浸润深度、神经浸润、脉管浸润、T分期、N分期、Dukes分期、术后化疗、阳性淋巴数在预后分组间存在差异(P0.05)。Boruta方法筛选结果显示,CA-199、阳性淋巴数、神经浸润、手术时间、是否化疗、化疗疗程数为影响预后结局的重要因素。(3)对于结直肠癌术后结局预测模型的比较显示,基于全数据集的预测效果最好的为Polynomial-SVM模型(AUC=0.907),基于差异性分析预测效果最好的为Polynomial-SVM模型(AUC=0.911),基于Boruta特征选择的模型其预测效果以RF为好(AUC=0.982),且三种最优模型之间存在差异(2x=7.74,P=0.021)。(4)COX比例风险模型发现,CA-199高(RR=2.002,95%CI:1.143~3.505)、阳性淋巴数(RR=1.244,95%CI:1.141~1.357)、神经浸润(RR=2.206,95%CI:1.130~4.308)、术中放疗(RR=2.098,95%CI:1.191~3.696)为结直肠癌术后结局的可能危险因素。综上所述,基于Boruta特征选择的模型预测效果较基于传统差异性分析的模型预测效果好,Boruta特征选择算法可作为临床数据预测模型前的数据降维方式,以便降低模型复杂度,同时提高预测效果。基于Boruta特征选择的RF模型能在一定程度上预测CRC根治术后结局,对指导临床医生术前干预有所帮助。
[Abstract]:Colorectal cancer (CRC) is a common malignant tumor of the digestive tract. The incidence and mortality are increasing year by year. It has become an important disease affecting human health. At present, the main treatment method of CRC is surgery, but there is still a certain proportion of the risk of recurrence and metastasis after the operation, and the recurrence and metastasis will seriously affect the prognosis of the patients. It is very important to predict the outcome of the patients accurately and take appropriate measures. Research on the risk and outcome of postoperative transfer of CRC has been predicted at home and abroad, mainly using multiple factor Logistic regression and COX regression. However, these 2 methods are not suitable for the analysis of small sample high dimensional data. Based on the statistical learning theory, the support vector machine learning algorithm (SVM), the random forest algorithm (RF) applies to the small sample, the classification of the high dimensional data, the prediction model with good generalization ability can be obtained. The prediction model combined with the feature selection algorithm can reduce the redundancy of space while reducing the training cost. Among them, the Boruta calculation is also reduced. By selecting the importance of the outcome, the method can avoid the correlation between variables and is more suitable for medical data, but it has not been applied to the prediction of malignant tumor data. This study is divided into two parts. The first part is divided into two parts. The first part uses the standard data set of UCI to carry out simulation pretest and SPSS 22 software for single factor difference. Analysis, R 3.30 carries out Boruta feature selection, then establishes SVM model and RF model. Using Stata14.0 to compare the prediction effect of the model, try to screen out the better feature selection method. The second part is to verify the model of the first part, and the bad outcome after the CRC operation. The research results show that: (1) according to the classification prediction results of the UCI standard data set, the RF prediction model based on the full data set (AUC=0.717) has better effect, the better prediction effect based on the difference analysis is the Polynomial-SVM model (AUC=0.756), and the RF model (AUC=0.905).ROC curve based on the Boruta characteristic selection prediction is better. The analysis showed that there were differences between the optimal prediction models of different pretreatment methods (2x=7.27, P=0.026). (2) the results of colorectal cancer single factor analysis showed that the tumor site, CA-199, CEA, infiltration depth, nerve infiltration, vascular infiltration, T staging, N staging, Dukes staging, postoperative chemotherapy, and positive lymphatic numbers were different (P0.05).Boruta square between the prognosis groups. The results of screening showed that CA-199, positive lymph number, nerve infiltration, operation time, chemotherapy and chemotherapy were important factors affecting the prognosis. (3) the comparison of prognosis model of postoperative outcome of colorectal cancer showed that the best prediction effect based on full data set was Polynomial-SVM model (AUC=0.907), based on differential analysis The best test result is Polynomial-SVM model (AUC=0.911), the model based on Boruta feature selection is good for RF (AUC=0.982), and there is a difference between the three optimal models (2x=7.74, P=0.021). (4) COX proportional hazard model found that CA-199 high (RR= 2.002,95%CI:1.143~3.505), positive lymph number (RR=1.244,95%CI:1.141~1.357), God Through infiltration (RR=2.206,95%CI:1.130~4.308), intraoperative radiotherapy (RR=2.098,95%CI:1.191~3.696) is a possible risk factor for the outcome of colorectal cancer surgery. To sum up, the predictive effect of the model based on Boruta feature selection is better than the model based on the traditional difference analysis model, and the Boruta feature selection algorithm can be used as a clinical data prediction model. The previous data reduction method can reduce the model complexity and improve the prediction effect. The RF model based on Boruta feature selection can predict the outcome of CRC after radical operation to a certain extent, which is helpful to the preoperative intervention of clinicians.

【学位授予单位】：重庆医科大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R735.34

【相似文献】