基于支持向量机模型预测结直肠癌根治术预后风险

Risk prediction of prognosis in patients with colorectal cancer under radical operation based on support vector machine

目的:基于支持向量机(support vector machine,SVM)预测结直肠癌患者术后风险。方法:回顾性分析2011年1月至2013年12月行根治性手术治疗的199例结直肠癌患者临床资料,按照是否无病生存(disease free survival,DFS)分为预后好、差2组。采用SAS9.2进行基本统计分析、logistic回归分析,运用R3.2.4软件建立SVM预测模型、受试者工作曲线(receiver operating characteristic curve,ROC)及曲线下面积(area under the curve,AUC)评价模型预测效果。结果:差异性分析显示,总蛋白、白蛋白、阳性淋巴率、组织类型、p53、N分期、放化疗等因素在预后组间存在统计学差异(P<0.05);多因素逐步logistic回归结果显示,N2分期(OR=13.5,95%CI=3.85~47.5)、p53阳性(OR=0.314,95%CI=0.125~0.790)可能是影响预后的危险因素。线性核、多项式核、sigmoid核及RBF(径向基函数)核SVM模型的预测准确率分别为69.70%、66.70%、69.70%、74.24%。对所有预测模型性能对比显示,SVM模型预测性能高于logistic回归模型,其中RBF-SVM模型效果最好,预测准确率为74.24%,特异度为0.714,灵敏度为0.765,AUC值为0.708(95%CI=0.551~0.866)。结论:N2分期、p53阳性可能是影响预后结局的危险因素,应当予以重视,减少预后风险。

Objective：To predict the postoperative risk of patients with colorectal cancer based on support vector machine（SVM）.Methods：Clinical data of 199 patients with colorectal cancer were collected retrospectively from the hospital from January2011 to December 2013,and patients were divided into two groups according to disease free survival or not. Basic statistics analysis and logistic regression were conducted by SAS9.2. Parameters optimization and establishment of SVM model wereanalyzed by R3.2.4. Receiver operating characteristic curve（ROC）and area under the curve（AUC）were used to evaluate their performance. Results：Differences analysis showed that total protein,albumin,positive lymph node number,tissue type,N,p53,radiotherapy and chemo-therapy were significantly different between prognosis groups（P〈0.05）. Multiple logistic regression showed that N2（OR=13.5,95%CI=3.85 to 47.5）,p53 positive（OR=0.314,95%CI=0.125 to 0.790）maybe the risk factors of prognosis. The prediction accuracy of linear-kernel,polynomial-kernel,sigmoid-kernel and RBF（radial basis function）-kernel was 69.70%,66.70%,69.70%,74.24%,respectively. The results of performance comparison showed that SVM models performed better than logistic regression model;RBF-SVM was the best among all the models,with the prediction accuracy of 74.24%;sensitivity,specificity and AUC were 0.714,0.765,0.708（95%CI=0.551 to 0.866）,respectively. Conclusion：N2 staging and p53 positive maybe the risk factors for prognosis and more attention should be paid to decrease the risk of prognosis.