Risk prediction methods

我们继续开发、完善和评估国家癌症研究所乳腺癌风险评估工具（BCRAT）。我们与西雅图华盛顿大学的暑期研究员Mateo Banegas一起量化了BCRAT在妇女健康倡议中为拉丁妇女提供的服务。我们发现BCRAT在一定程度上低估了风险，但对最近的SEER率进行重新校准改善了BCRAT的预测。Banegas博士最近获得了博士学位，并正在进行研究，以开发一个新的模型，为拉丁美洲妇女的绝对乳腺癌风险。乳腺癌、子宫内膜癌和卵巢癌有着共同的激素病因学和流行病学风险因素。使用来自前列腺、肺、结直肠和卵巢（PLCO）癌症筛查试验和AARP-NIH饮食与健康研究的50岁以上白色非西班牙裔女性的数据，以及来自监测、流行病学和最终结果（SEER）登记处的发病率和死亡率的d RR，我们开发了模型来估计女性在特定时间段内发生乳腺癌、子宫内膜癌或卵巢癌的绝对风险。完善和验证这些模型的工作正在进行。我们开发了模型来预测原发性儿童癌症治疗后甲状腺癌的绝对风险。这些模型具有良好的判别准确性（AUC），在存在强风险因素（如甲状腺结节和颈部辐射）的情况下，20年的风险可高达7%。模型在独立队列数据中得到验证。有兴趣确定是否增加单核苷酸多态性（SNP）的信息可以增加风险模型筛选的判别准确性和有用性。我们研究了可预见的全基因组关联研究是否会发现足够多的SNP，这些SNP具有足够强的关联性，从而在风险预测方面做出实质性的进一步改善。我们评估了一系列癌症的几个标准，并得出结论，未来SNP发现的贡献可能不大。使用来自北方加州（KPNC）的140万名接受HPV检测和巴氏涂片检查的妇女的数据，我们计算了多次筛查访视中HPV和巴氏涂片检测结果的所有可能组合的癌前病变和癌症的绝对风险。这些风险对于2012年9月举行的共识宫颈筛查和管理指南委员会会议至关重要。我们正在构建宫颈癌风险和HPV感染清除机会的双变量模型。对于每个可能的筛查间隔，该模型提供了癌症风险以及HPV在无需干预的情况下自然清除的机会。这些资料对制订子宫颈普查指引应该是有用的。我们正在估计PLCO中基于肺癌风险的分层中的绝对死亡风险，以评估某些基于风险的吸烟者子集是否可能从胸部X线肺癌筛查中获得死亡率获益。我们开发了方法和软件，允许人们拟合和验证风险预测模型，其中仅在队列的一个子集上测量关键暴露，并应用于开发肺癌风险的新模型，该模型包括仅在队列的一个子集上测量的血清炎症标志物的组合。我们提出了一个新的二项回归模型，用于估计绝对风险（和风险差异），该模型允许包括具有逻辑或线性效应的暴露。我们将该模型扩展到基于人群的病例对照研究中估计绝对风险，并应用于估计女性与男性吸烟者肺癌的绝对风险。我们提出并公布了两个标准，以评估预测疾病发病风险的模型在筛查和预防中的有用性，或者评估预后模型在疾病诊断后管理中的有用性。第一个标准，遵循PCF（q）的病例比例，是将发生疾病的个体在最高风险人群中的个体比例q中所占的比例。第二个标准是需要随访的比例，即PNF（p），即需要随访的最高风险的一般人群的比例，以便对注定成为病例的比例p进行随访。我们最近扩展了这些标准整合PCF（q）和PNF（p）的范围内的q和p。我们还开发了估计PCF（q）和PNF（p）和他们的综合形式的方法时，风险模型被假定为校准良好，并根据健康结果的经验数据。后一种方法即使在风险模型没有很好校准的情况下也是有效的，但它们产生的估计值不太精确。我们开发了基于代表性调查数据（如国家健康和营养检查调查（NHANES））估计和推断绝对风险的方法。使用影响函数，我们推导出方差估计，是有效的加权和整群抽样调查。我们还提出了一个标准，以估计每个竞争原因的绝对风险计算的重要性。我们开发了多重插补方法，以估计已知死亡但死因不明的绝对死亡风险。在单独的工作中，我们展示了如何使用这些方法与SEER数据，其中只有一小部分这样的死亡原因信息缺失，以估计死于偶发癌症的绝对风险。

We continued to develop, refine and evaluate the National Cancer Institutes Breast Cancer Risk Assessment Tool (BCRAT). With a summer Fellow, Mateo Banegas, from University of Washington in Seattle, we quantified the performance of the BCRAT for Latina women in the Women's Health Initiative. We found that BCRAT underestimated risk somewhat, but that recalibration to more recent SEER rates improved the predictions of BCRAT. Dr. Banegas recently received his doctorate and is pursuing research to develop a new model for absolute breast cancer risk for Latina women. Breast, endometrial and ovarian cancers share a hormonal etiology and epidemiologic risk factors. Using data on white, non-Hispanic women over age 50 years from the Prostate, Lung, Colorectal, and Ovary (PLCO) Cancer Screening Trial and the AARP-NIH Diet and Health Study and d RRs with incidence and mortality rates from the Surveillance, Epidemiology and End Results (SEER) registries, we developed models to estimate a womans absolute risk of developing breast, endometrial or ovarian cancer over specific intervals. Work to refine and validate the models is ongoing. We developed models to predict the absolute risk of thyroid cancer following treatment of primary childhood cancers. These models have good discriminatory accuracy (AUC) and can yield 20-year risks of up to 7% in the presence of strong risk factors, such as thyroid nodules and neck radiation. The models were validated in independent cohort data. There is interest in determining whether adding information from single nucleotide polymorphisms (SNPs) can increase the discriminatory accuracy and usefulness for screening of risk models. We examined whether foreseeable genome-wide association studies will discover enough SNPs with sufficiently strong associations to make substantial further improvements in risk prediction. We evaluated several criteria for a range of cancers and concluded that the contributions from future SNP discoveries are likely to be modest. Using data from 1.4 million women undergoing HPV testing and Pap smears in Kaiser Permanente Northern California (KPNC), we calculated absolute risks of precancer and cancer for all possible combinations of HPV and Pap test results over multiple screening visits. These risks are critical for the consensus cervical screening and management guidelines committee meeting in September 2012. We are constructing a bivariate model of the risk of cervical cancer and the chance of clearance of an HPV infection. For each possible screening interval, the model provides the risk of cancer and the chance that HPV will naturally clear without need for intervention. This information should be useful in developing cervical screening guidelines. We are estimating absolute mortality risks within lung cancer risk-based strata in the PLCO to assess whether some risk-based subsets of smokers might have a mortality benefit from chest x-ray lung cancer screening. We developed methods and software to allow one to fit and validate a risk prediction model where key exposures are measured on only a subset of a cohort, with application to developing a new model of lung cancer risk that includes combinations of serum inflammatory markers measured on only a subset of the cohort. We proposed a new binomial regression model for estimating absolute risk (and risk differences) that allows one to include exposures that have logistic or linear effects. We extended this model to estimate absolute risk from population-based case-control studies, with application to estimating absolute risk of lung cancer in female vs. male smokers. We proposed and published two criteria to assess the usefulness of models that predict risk of disease incidence for screening and prevention, or the usefulness of prognostic models for management following disease diagnosis. The first criterion, the proportion of cases followed PCF(q), is the proportion of individuals who will develop disease who are included in the proportion q of individuals in the population at highest risk. The second criterion is the proportion needed to follow-up, PNF(p), namely the proportion of the general population at highest risk that one needs to follow in order that a proportion p of those destined to become cases will be followed. We recently extended these criteria by integrating PCF(q) and PNF(p) over ranges of q and p. We also developed methods of estimating PCF(q) and PNF(p) and their integrated forms both when the risk model was assumed to be well calibrated, and on the basis of empirical data on health outcomes. The latter methods are valid even when the risk models are not well calibrated, but they yield less precise estimates. We developed approaches for estimating and performing inference on absolute risk based on representative survey data, such as the National Health and Nutrition Examination Survey (NHANES). Using influence functions, we derived variance estimates that are valid for surveys with weighting and cluster sampling. We also proposed a criterion to estimate the importance of each competing cause on the calculation of the absolute risk of a particular cause. We developed multiple imputation methods to estimate absolute risk of death when some individuals were known to have died but their causes of death were unknown. In separate work, we showed how these methods could be used with SEER data, which have only a small proportion of such missing cause of death information, to estimate the absolute risk of dying from an incident cancer.