TREE-STRUCTURED SURVIVAL ANALYSIS--METHODS AND SOFTWARE

生存分析是癌症流行病学的基础。比如说， 乳腺癌研究通常使用终点，如复发时间或 至死然而，很少有数据分析工具可以用来探索这些数据。 因此，统计评估可求助于常规使用考克斯 比例风险模型，尽管体现在强有力的假设， 这种方法。此外，考克斯模型并不擅长识别 患者的预后亚组。 该项目的目的是进一步发展和传播， 用于执行树结构生存分析的方法和软件。 树技术的吸引力在于它们在概念上简单， 可解释的，并且是非参数的，不强加强有力的假设。 此外，它们旨在识别不同的预后亚组。 这些子组的特征在于具有共同的预测值 变量，从而使临床上有意义的分裂。除本 约束一样的预测值，所提取的子组具有 最大程度的不同生存前景。 树结构的生存分析已经应用于一系列的 包括乳腺癌标志物的评估。然而，更广泛的 软件缺陷妨碍了使用。目前，只工作 不便于用户使用的代码可用。目的是利用这一点 核心作为产生综合树结构生存的基础 分析程序。这将涉及编写精炼的输入、输出和帮助 设施算法效率也将得到提高。方法学 包括处理缺失数据、随时间变化的 预测因子和评估统计学显著性也将是 incorporated. 两项独立的乳腺癌生存研究的申请将 追求。第一个问题涉及属于联合国系统保护伞的项目， 授予UCSF的乳腺癌SPORE，其中包括几个标志物 问题研究第二个是关于18年的预后价值的研究， 乳腺液细胞学检查由加州大学旧金山分校 流行病学和生物统计学。

Survival analysis is fundamental to cancer epidemiology. For example, breast cancer studies often use endpoints such as time to relapse or time to death. Yet very few data analytic tools exist for exploring such data. Thus, statistical assessments make recourse to routine use of Cox proportional hazards models, despite the strong assumptions embodied in this approach. Additionally, Cox models are not adept at identifying prognostic subgroups of patients. The purpose of this project is to further develop, and disseminate, methods and software for performing tree structured survival analysis. Tree techniques have appeal in that they are conceptually simple, readily interpretable and, being nonparametric, do not impose strong assumptions. Further, they are designed to identify distinct prognostic subgroups. These sub-groups are characterized by having common values of predictor variables thus making for clinically meaningful divisions. Subject to this constraint of like predictor values, the subgroups extracted possess maximally disparate survival prospects. Tree structured survival analysis has already been applied in a range of settings including evaluation of breast cancer markers. However, wider usage has been hampered by software deficiencies. Presently, only working code that is not user-friendly is available. The intention is to use this kernel as a basis for producing a comprehensive tree structured survival analysis program. This will involve writing refined input, output and help facilities. Algorithm efficiency will also be improved. Methodologic developments including means for handling missing data, time-varying predictors and assessing statistical significance will also be incorporated. Applications to two separate breast cancer survival studies will be pursued. The first involves projects falling under the umbrella of the SPORE in breast cancer awarded to UCSF which includes several marker studies. The second concerns an 18 year study of the prognostic value of breast fluid cytology being conducted by the UCSF Department of Epidemiology and Biostatistics.