Comprehensive Prognostic Modeling for Esophageal Cancer: A Bayesian Approach

描述(由申请人提供)：我们的长期目标是利用复杂的数据源开发准确和可推广的预后模型，并为改善患者预后做出贡献。本研究的动机是开发准确的食道癌预后模型的临床需要和改进复杂数据源的数据分析的方法学需要。食道癌带来了一系列独特的挑战，表明了这一努力的重要性。首先，其相对较低的发病率使大型随机实验难以进行。大型观测数据库成为研究的一个特别重要的来源。其次，它的发病率上升和高死亡率迫切需要准确地确定可以进行有效干预的预后因素。然而，由于数据来源或统计方法的限制，疾病的组织学类型、治疗的差异(即手术的广泛性)和护理质量(以提供者的数量作为替代)的预后作用及其临床和卫生政策含义一直备受争议。第三，尽管有先进的统计方法，如混合效应模型，来处理来自大型数据库的数据复杂性的一个关键方面--集群造成的异质性--但很少有预测模型建立在这些方法的基础上。这在一定程度上是因为缺乏关于如何对集群数据结构进行最佳建模的指导。关于这种先进的数据结构建模是否会提高预测模型的准确性，提高多少，以及在什么情况下进行，这些问题仍有待回答。本研究的目的是提高我们对食道癌预后的关键变量的理解，并提高我们对先进的统计方法在使用复杂数据源进行预后建模中的有效性和重要性的认识。它将通过以下具体目标实现：(1)利用SEER-Medicare数据库，为食道癌治疗后的多种结果(包括致命和非致命的短期并发症、总体和特定疾病的长期生存)建立综合预后模型。将使用贝叶斯混合效应模型来解释集群数据结构。这些模型将在内部和外部进行验证。(2)使用统计模拟对聚类生存数据的方法进行评估。通过模拟模拟真实研究的复杂结构的数据，研究几种常用的对预测建模有重要意义的脆弱模型(包括个体水平预测效应估计、集群水平预测效应估计、异质性评估和结果预测)的操作特性。将比较脆弱模型方法和其他技术，包括边际方法和分层方法在预测者效果估计和结果预测方面的准确性。本研究结果将对老年食道癌患者的临床和卫生决策具有直接的指导意义。它们还将对使用注册、管理和观察性数据库的未来预后建模产生积极影响。公共卫生相关性：拟议的研究旨在使用SEER Medicare关联数据库开发食道癌的综合预后模型，并使用统计模拟比较聚集生存数据的方法。新的预后模型将有助于更准确地预测具有特定疾病特征并由特定机构的医生进行特定治疗的食道癌患者的结果。模拟研究的结果将促进我们对相关先进统计方法的操作特征的了解，并为未来使用登记、管理和观察数据库的预后建模工作提供指导。

DESCRIPTION (provided by applicant): Our long-term objective is to develop accurate and generalizable prognostic models using complex data sources and contribute to improve patient outcome. This study is motivated by the clinical need of developing accurate prognostic models for esophageal cancer and the methodological need of improvement in data analysis with complex data source. Esophageal cancer presents a unique set of challenges that signifies the importance of this effort. First, its relatively low incidence rate makes large randomized experiments difficult to carry out. Large observational database becomes a particularly important source for research. Second, its rising incidence and high mortality rate present an urgent need to accurately identify prognostic factors where effective interventions could be directed. However, due to limitations from either data sources or statistical approaches, the prognostic roles of histology type of the disease, variation in treatments (i.e., the extensiveness of the surgery) and quality of care (using provider volume as a surrogate) and their clinical and health policy implications are hotly debated. Third, despite the availability of advanced statistical methods, such as mixed effects models, to handle a key aspect of data complexity from large databases - the heterogeneity due to clustering - few prognostic models are built upon these methods. This is partly due to the lack of guidance on how to best model the clustered data structure. Questions regarding whether, by how much, and under which setting such advanced data structure modeling would improve the accuracy of the prognostic model remain to be answered. The goal of this study is to improve our understanding regarding the key variables in esophageal cancer prognosis and to advance our knowledge regarding the validity and importance of the advanced statistical methods in prognostic modeling using complex data sources. It will be achieved through the following specific aims: (1) To develop comprehensive prognostic models for multiple outcomes (including fatal and no-fatal short-term complications, overall and disease-specific long-term survival) following esophageal cancer treatment using SEER-Medicare database. Bayesian mixed-effects models will be used to account for the clustered data structure. The models will be validated internally and externally. (2) To evaluate methods for clustered survival data using statistical simulations. By simulating data of complex structure that mimic real studies, the operational characteristics of several commonly used frailty models for issues of importance to prognostic modeling (including the individuallevel predictor effect estimation, cluster-level predictor effect estimation, heterogeneity assessment and outcome prediction) will be investigated. The performance of the frailty model approach and other techniques including the marginal and stratified approaches will be compared with regard to accuracy in predictors' effects estimation and outcome prediction. Results from this study will have direct applicability in aiding clinical and health policy decision making for the care of elder patients with esophageal cancer. They will also impact positively on future prognostic modeling using registry, administrative, and observational databases. PUBLIC HEALTH RELEVANCE: The proposed studies aim to develop comprehensive prognostic models for esophageal cancer using SEER Medicare linked database and to compare methods for clustered survival data using statistical simulations. The new prognostic models will help to more accurately predict outcomes of esophageal cancer patients with a particular set of disease characteristics and treated with a specific treatment by physicians at a particular institution of certain quality of care. Results from the simulation study will advance our knowledge on the operational characteristics of the related advanced statistical methods and provide guidance for future prognostic modeling efforts using registry, administrative, and observational databases.