肿瘤生存数据中比例风险假定失效时的统计分析策略

发布时间：2018-07-25 12:24

【摘要】：背景子宫内膜癌、宫颈癌、卵巢癌是严重危害妇女生命健康的三大妇科肿瘤,即使在医学技术发达的今天,其病死率和死亡率仍然高居不下。子宫内膜癌是女性生殖系统的三大恶性肿瘤之首,占女性生殖系统肿瘤的20%-30%,在欧美国家,子宫内膜癌发病率已占据妇科恶性肿瘤第一位,2016年美国新增的子宫内膜癌病例即超过了宫颈癌和卵巢癌的总和,而近些年来发展中国家的发病率也明显上升。剧WHO统计,2012年全球新发宫颈癌患者为527,624(占女性癌症7.9%),死人数为265,672(占所有女性癌症7.5%),位居世界妇女发病率的第四位,虽然近年来,在发达国家的发病率正在下降,但在一些发展中国家位居首位,在2015年,300,000例死者中,约90%的病例来自于中低等收入国家,因此宫颈癌的预后也是我们不容忽视的问题。卵巢癌在妇科恶性肿瘤中的发病率位居第二,病死率位居第一,全球每年约有19万新发病例,流行病学研究显示,妇女一生患卵巢癌的风险为1.4%,由于卵巢癌深居盆腔,缺乏早期症状及有效的筛查手段,被确诊时多数已达晚期,总的5年生存率仅有45%,在妇科恶性肿瘤中最难早期诊断,最难治愈,预后最差,所以构建恰当的预后模型探索其影响因素和预测患者的生存率将变得尤为重要。Cox比例风险模型是肿瘤数据分析中最常见的回归模型。然而,当比例风险假定失效时,Cox比例风险模型违背其前提条件,这种情况下使用Cox比例风险模型得到的结果不可靠。而加速失效模型中的Buckley-James模型应用线性回归思想处理生存时间与影响因素之间的关系不需要满足该假定。Trinquart等人提倡用限制性平均生存时间(RMST)作为另一个概括型统计量来评价组间效应,然而,Buckley-James模型和RMST模型得到的指标均为概括型指标,不能展现不同时间点的变化的趋势,Cox提出可以用时间函数和时间相依协变量的交互项来构造扩展Cox模型探索不同时间点的相对风险比。在实际的临床治疗中,病人可能更为关心的是其自身在不同治疗期间的生存率,动态预测中的比例基线界标超级模型(PBLS模型)是一个条件模型,可以探索不同时间点的相对风险比,更可以预测w年的动态生存率。目的本研究选取来自美国监测、流行病和最终结果数据库中在2004年1月1日到2013年12月31日10年间三大妇科恶性肿瘤患者的生存数据,采用Cox比例风险模型、半参数加速失效模型(AFT模型)、以RMST为指标的广义线性模型(RMST模型)、Cox时间相依模型(扩展Cox模型)和动态预测分析中的PBLS模型探索宫颈癌、子宫内膜癌、卵巢癌患者预后的影响因素,并进行不同时间点的5年生存率的预测分析,为三大妇科恶性肿瘤患者的预后提供基本临床资料依据,帮助临床研究者针对不同患者而制定最佳治疗方案。方法本研究将子宫内膜癌患者、宫颈癌患者、卵巢癌患者的死亡原因(或终点事件)为全因死亡,将患者失访或存活等作为删失。单因素分析采用Kaplan-Meier方法来估计不同癌症病人(宫颈癌、子宫内膜癌、卵巢癌)各个协变量的生存率,并用Log-rank检验生存曲线之间的差异是否有统计学意义。用Cox比例风险模型探索协因素的相对风险比,用AFT模型探索各个因素的加速失效因子、用RMST模型探索协变量对限制性平均生存时间的影响,用扩展Cox模型探索各个因素在不同时间点对相对风险比的影响,用PBLS模型预测不同时间点的5年生存率。评价模型的指数采用C-index、AIC、AUC。分析使用R软件(3.3.4版本)进行,检验均为双侧检验,检验水准α = 0.05。结果三大妇科恶性肿瘤主要以已婚为主,白人为主,子宫内膜癌和卵巢癌的诊断年龄较大、宫颈癌的诊断年龄较小,不同诊断年份没有差异,子宫内膜癌和宫颈癌的FIGO以一期为主、卵巢癌以三期为主,发生淋巴结转移较少,子宫内膜癌接受放疗人数较少、宫颈癌人数较多、卵巢癌本研究中没有纳入放疗的患者,子宫内膜癌和卵巢癌的手术率高达90%,而宫颈癌低于70%,分化程度由高到低为,恶性度由低到高为:子宫内膜癌、宫颈癌、卵巢癌,子宫内膜癌、卵巢癌主要以腺癌为主、宫颈癌主要以鳞癌为主,注册地点东西部相当,宫颈癌好发于子宫颈,卵巢癌好发于双边。对于婚姻状态,已婚分离(离异、分居、丧偶)相对于已婚的死亡风险高,生存率低,未婚女性较为复杂,子宫内膜癌的未婚女性与已婚女性的生存率没有统计学差异,在宫颈癌中,未婚女性的生存率显著高于已婚女性,在卵巢癌中,相对风险比随着时间发生变化;诊断年龄越大,生存率越低,在宫颈癌中,年龄间的相对风险具有时间效应;不同种族的子宫内膜癌患者的生存率不同,宫颈癌也是,但是白人的卵巢癌患者和其他人种的生存率没有统计学差异;FIGO分期越高,生存率越低,其中子宫内膜癌的FIGO分期的相对风险比呈下降趋势;淋巴结转移的病人的生存率均低于没有淋巴结转移的病人,其相对风险比在子宫内膜癌先增大后减少,宫颈癌呈下降趋势、而卵巢癌不变;手术对于三大妇科恶性肿瘤是一个保护因素。应用动态预测分析发现,PBLS模型能体现不同时间点的5年生存率,而Cox比例风险模型不能体现不同时间点的变化过程。在三大妇科恶性肿瘤的5个模型分析中,无论从C-index还是从AIC,都是扩展Cox模型表现最好,同样30次的重抽样结果也显示扩展Cox模型最好,在子宫内膜癌和卵巢癌中,AFT模型的C-index较大,而在宫颈癌中,RMST模型的C-index较大,发现在AUC值和Slope指数中,PBLS模型显著高于Cox比例风险模型,动态预测不但能探索癌症患者的预后因素,最重要的是预测不同时间点的w年生存率。结论婚姻状态、诊断年龄、种族、FIGO分期、淋巴结转移、放疗等都是影响女性生殖器三大恶性肿瘤的影响因素,且部分因素相对风险并不是永恒不变的。首次使用动态预测分析中的PBLS模型预测美国女性的妇科三大女恶性肿瘤的不同时间点的5年生存率,临床研究者制定患者的个体治疗方案,指导病人持续治疗、增加依从性、最终提高生存率。
[Abstract]:Background endometrial cancer, cervical cancer, and ovarian cancer are the three major gynecologic tumors that seriously harm the life and health of women. Even in the advanced medical technology, the mortality and mortality rate still remain high. Endometrial cancer is the first of the three major female reproductive system tumors, which accounts for the 20%-30% of female reproductive system tumors, in European and American countries. The incidence of endometrial cancer has taken the first place in gynecologic malignancies. In 2016, the new cases of endometrial cancer in the United States were more than the total of cervical and ovarian cancer. In recent years, the incidence of the developing countries was also significantly increased. The WHO statistics showed that in 2012, the number of new cervical cancer patients in the world was 527624 (7.9% of female cancer) and the number of dead people was 265,67 2 (7.5% of all women's cancer), ranking fourth in the world's incidence of women, although the incidence of the disease in the developed countries is declining in recent years, but in some developing countries, in 2015, about 90% of the 300000 cases of the deceased are from the middle and lower income countries, so the prognosis of cervical cancer is also a problem we can not ignore. The incidence of ovarian cancer in gynecologic malignancies is second, the mortality rate is the first, and there are about 19 million new cases in the world every year. The epidemiological study shows that the risk of women's ovarian cancer is 1.4%. Because of the deep pelvic cavity, the lack of early symptoms and effective screening methods, most of the ovarian cancer has reached the late period and the total of 5 years. The survival rate is only 45%. It is the most difficult to diagnose in the malignant tumor of Gynecology, the most difficult to cure, and the worst prognosis. Therefore, it is particularly important to construct an appropriate prognostic model to explore its influencing factors and predict the patient's survival rate. The.Cox proportion risk model is the most common regression model in the analysis of tumor data. However, when the proportional risk is assumed to be invalid, Cox The proportional risk model is contrary to its precondition, and the results obtained by using the Cox proportional hazard model in this case are not reliable. While the Buckley-James model in the accelerated failure model applies the linear regression idea to deal with the relationship between the survival time and the influencing factors, it does not need to satisfy the assumption that.Trinquart and others advocate using the restrictive average survival time. RMST is another generalized statistic to evaluate inter group effects. However, the Buckley-James model and the RMST model are all generalized indexes, which can not show the trend of change at different time points. Cox proposes to use the time function and time dependent covariate to construct extended Cox model to explore different time points. Relative risk ratio. In practical clinical treatment, patients may be more concerned about their own survival rate during different treatments. The proportional baseline supermodel (PBLS model) in dynamic prediction is a conditional model, which can explore the relative risk ratio at different time points, and can predict the dynamic survival rate of W years. The survival data of three major gynecologic cancer patients from January 1, 2004 to December 31, 2013 were selected from the US monitoring, epidemic and final result database, using the Cox proportional hazard model, the semi parametric accelerated failure model (AFT model), the generalized linear model (RMST model) with RMST as the index, and the Cox time dependent model (expansion). The Cox model) and the PBLS model in dynamic prediction analysis are used to explore the factors affecting the prognosis of cervical cancer, endometrial cancer and ovarian cancer, and to predict the 5 year survival rate at different time points, providing basic clinical data for the prognosis of three major gynecologic malignancies, and helping clinical researchers to make the most for different patients. Methods the cause of death (or endpoint event) of endometrial cancer patients, cervical cancer patients and ovarian cancer patients was all caused by death, and the patients were lost or survived. The Kaplan-Meier method was used to estimate the survival of different cancer patients (cervical, endometrial and ovarian cancer). We use the Cox proportional hazard model to explore the relative risk ratio of the co factors, explore the accelerated failure factors of each factor with the AFT model, explore the influence of the covariate on the restrictive average survival time with the AFT model, and explore the factors with the extended Cox model, and explore the different factors by using the extended Cox model. The impact of time point on relative risk ratio, PBLS model was used to predict the 5 year survival rate at different time points. The index of the evaluation model was C-index, AIC, and AUC. analysis used R software (3.3.4 version). The test was both bilateral test. The test level was alpha = 0.05. results and the three major gynecologic malignancies were mainly married, white predominantly, endometrium. The diagnosis of cancer and ovarian cancer is older, the age of the diagnosis of cervical cancer is smaller, there is no difference in different diagnostic years, the FIGO of endometrial and cervical cancer is mainly in the first stage, the ovarian cancer is mainly in the three stage, the lymph node metastases are less, the number of endometrium cancer is less, the number of cervical cancer is more, the ovarian cancer is not included in this study. The operation rate of endometrial and ovarian cancer in patients with radiotherapy is up to 90%, while cervical cancer is less than 70%, the degree of differentiation is from high to low, and the degree of malignancy is from low to high: endometrial cancer, cervical cancer, ovarian cancer, endometrial cancer, ovarian cancer mainly adenocarcinoma, cervix cancer mainly based on squamous cell carcinoma, the location of registered location is equal, cervical cancer is good hair. In the sub cervix, ovarian cancer is more bilateral. For marital status, marriage separation (divorce, separation, widowhood) is higher than married death, low survival rate, unmarried women more complex, the survival rate of unmarried women with endometrial cancer and married women is not statistically different. In cervical cancer, the survival rate of unmarried women is significantly higher than that of the already married women. Married women, in ovarian cancer, the relative risk is changed over time; the greater the diagnostic age, the lower the survival rate, the relative risk of age in the cervical cancer, the time effect; the survival rate of the endometrium cancer patients of different races, the cervical cancer is also, but the survival rate of the white ovarian cancer patients and other ethnic groups is not unified. The higher the FIGO stage, the lower the survival rate, the relative risk ratio of the FIGO staging of endometrial carcinoma decreased, the survival rate of the patients with lymph node metastasis was lower than that of the patients without lymph node metastasis, and the relative risk was decreased after the endometrial cancer, and the cervical cancer was declining, and the ovarian cancer was unchanged; the operation was not constant. The three major gynecologic malignant tumor is a protective factor. The application of dynamic prediction analysis shows that the PBLS model can reflect the 5 year survival rate at different time points, while the Cox proportional hazard model can not reflect the change process at different time points. In the 5 model analysis of three major gynecologic malignant tumors, both from C-index or from AIC, the Cox model is extended. Best performance, the same 30 resampling results also showed that the extended Cox model was best. In endometrial and ovarian cancer, the C-index of the AFT model was larger. In the cervical cancer, the C-index of the RMST model was larger. It was found that the PBLS model was significantly higher than the Cox ratio risk model in the AUC value and the Slope index, and the dynamic prediction could not only explore the cancer patients. The most important factor in prognosis is to predict the w year survival rate at different time points. Conclusion marriage status, age, race, FIGO stage, lymph node metastasis, radiotherapy are all influencing factors of three major female genital cancers, and the relative risk of some factors is not permanent. For the first time, the PBLS model in dynamic prediction analysis is used. The model predicts the 5 year survival rate at different time points of three major female gynecologic malignancies in American women. The clinical researchers formulate individual treatment programs for patients, guide patients to continue treatment, increase compliance, and ultimately improve survival.
【学位授予单位】：南方医科大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R73-31

【参考文献】