目的 通过支持向量机模型(SVM)将临床指标和病理指标进行整合,分析其对结直肠癌(CRC)患者预后的预测价值.方法 选取2002-2011年在上海市第十人民医院胃肠外科行结直肠肿瘤切除术的患者2 951例为研究对象.收集患者的临床指标(性别、年龄、肿瘤大小、肿瘤位置、组织病理类型、T分期、N分期、M分期、G分期)和病理指标〔癌胚抗原(CEA)、增殖细胞核抗原(PCNA)、P53、CD34、S-100、NM23、CerB-b2、P21、Ki-67〕.按照随访信息是否缺失将患者分为两组,第一组为临床数据完整但随访信息缺失,共2 747例;第二组为临床数据及随访信息均完整,共204例.记录患者生存情况.第一组中,若某一患者有超过4个指标缺失,则该患者被剔除;在余下的患者中,若某一指标的缺失率>30.0%,则该指标被剔除;进一步将少发病理类型患者剔除;计算临床指标和病理指标的关联性,随后,将所得关联性在第二组患者中进行验证,若该关联性在第二组中存在,则直接将第二组患者纳入第三组中;若该关联性在第二组中不存在,则采用SVM内部算法剔除部分患者,直至该关联性再次成立,将余下的第二组患者纳入第三组.对第二组中患者的病理指标进行统计,若某一指标的缺失率>50.0%,则该指标被剔除.基于SVM对第三组数据进行处理,建立3种预测模型:SVM1基于临床指标、SVM2基于病理指标、SVM3基于临床指标和病理指标的汇总.结果 第一组中,共834例患者缺失指标<4个,其中性别、年龄、肿瘤位置、组织病理类型、P53、CD34、S-100、CerB-b2、Ki-67共9个指标缺失率<30.0%而被保留,剔除5例少见肿瘤患者后,共剩下829例患者.第一组患者年龄与CerB-b2表达情况存在关联性(P<0.05).第二组剔除105例患者后,余下99例患者,患者年龄与CerB-b2表达情况存在关联性(P<0.05),将这99例患者纳入第三组.第二组中PCNA、P53、CD34、S-100、CerB-b2共5个指标缺失率<50.0%而被保留.第三组患者年龄与S-100、CerB-b2表达情况存在关联性(P<0.05);M分期与PCNA表达情况存在关联性(P<0.05).第三组不同T分期、N分期患者生存曲线比较,差异有统计学意义(P<0.05).SVM1纳入9个临床指标(性别、年龄、肿瘤大小、肿瘤位置、组织病理类型、T分期、N分期、M分期、G分期),准确率为83.4%;SVM2纳入5个病理指标(PCNA、P53、CD34、S-100、CerB-b2),准确率为78.8%;初始的SVM3纳入以上9个临床指标及5个病理指标,准确率为74.8%,通过最小冗余最大相关性(MRMR)法对指标进行进一步筛选,得到最终的SVM3,其纳入4个临床指标(肿瘤位置、组织病理类型、T分期、N分期)和2个病理指标(CD34、CerB-b2),准确率为81.8%.不同风险SVM1、SVM2、SVM3患者生存曲线比较,差异有统计学意义(P<0.05).结论 临床指标如年龄、M分期与病理指标如CerB-b2、S-100和PCNA存在一定的关联性;借助SVM模型将临床指标和病理指标进行整合分析可对CRC患者预后进行有效预测.%Objective To investigate the value of clinical and pathological indexes integrated with support vector machine model(SVM) in the prediction of the prognosis of patients with colorectal cancer(CRC).Methods We enrolled 2 951 patients undergoing resection of colorectal cancer in Department of Gastroenterological Surgery of Shanghai Tenth People′s Hospital from 2002 to 2011,and collected their clinical indexes(sex,age,tumor size,tumor site,histopathological type,T stage,N stage,M stage,G stage) and pathological indexes(CEA,PCNA,P53,CD34,S-100,NM23,CerB-b2,P21,and Ki-67).We divided the patients into 2 groups according to whether they lacked follow-up information.Group 1 had sufficient clinical data but lacked follow-up information,with a total of 2 747 cases.Group 2 had complete clinical data and follow-up information,with a total of 204 cases.The survival situation of the patients was recorded.In group 1,if a patient had more than 4 indicators missing,the patient was eliminated;among the remaining patients,if any index had a loss rate of >30.0%,the index was excluded;further,patients with fewer pathological types were excluded.The association of clinical indicators and pathological indicators was calculated,and then the resulting association was verified in patients in group 2.If the association was present in group 2,the patients of group 2 would be included in group 3 directly;if the association did not exist in group 2,the SVM internal algorithm was used to remove some patients until the association was established and the remaining patients of group 2 were included in group 3.The pathological indexes of group 2 were statistically analyzed,if the missing rate of an index was >50.0%,the index was eliminated.Based on SVM,data of group 3 was processed,and 3 prediction models were established:SVM1 based on clinical indexes,SVM2 based on pathological indexes,SVM3 based on clinical and pathological indexes.Results In group 1,a total of 834 patients had missing indexes <4.The loss rates of 9 indexes(sex,age,tumor site,histopathological type,P53,CD34,S-100,CerB-b2 and Ki-67) were less than 30.0%,and they were remained.After excluding 5 patients with rare tumors,a total of 829 patients remained in group 1.There was a correlation between age and expression of CerB-b2 in group 1(P<0.05).After eliminating 105 patients,the age of the 99 remaining patients was related to the expression of CerB-b2(P<0.05),and the 99 patients were included in group 2,the loss rates of 5 indexes(PCNA,P53,CD34,S-100,CerB-b2) were less than 50.0%,and they were remained.Age correlated with the expression of S-100 as well as with the expression of CerB-b2 in group 3(P<0.05);there was correlation between the M staging and the expression of PCNA in group 3(P<0.05).The survival curve of the patients in group 3 significantly varied by T stage and N stage(P<0.05).SVM1 included 9 clinical indicators(sex,age,tumor size,tumor location,histopathological type,T stage,N stage,M stage,G stage),the accuracy rate was 83.4%.SVM2 included 5 pathological indexes(PCNA,P53,CD34,S-100,CerB-b2),the accuracy rate was 78.8%.The initial SVM3 included the above 9 clinical indicators and 5 pathological indicators,the accuracy rate was 74.8%.The index was further screened by MRMR method to obtain the final SVM3,which included 4 clinical indexes(tumor location,histopathological type,T stage,N stage) and 2 pathological indexes(CD34,CerB-b2),and the accuracy rate was 81.8%.The survival curves of patients with different risk SVM1,SVM2 and SVM3 were statistically significant(P<0.05).Conclusion There was a certain correlation between pathological indexes such as CerB-b2,S-100 and PCNA and clinical indexes such as age and M stage;SVM model can be used to integrate the pathological and clinical indexes and to effectively predict the prognosis of patients with colorectal cancer.
展开▼
