Risk prediction methods

我们继续发展、完善和评估国家癌症研究所乳腺癌风险评估工具（BCRAT）。使用来自亚裔美国人乳腺癌研究的数据，我们获得了亚裔美国女性的种族特异性相对风险和归因风险，并将这些与SEER中年龄和种族特异性乳腺癌发病率相结合，得出绝对风险。在发表和纳入BCRAT之前，我们正处于评估该模型的最后阶段。我们与一位夏季研究员Mateo Banegas（他是西雅图华盛顿大学的博士生）一起，在妇女健康倡议中量化了拉丁裔妇女的BCRAT表现。我们发现BCRAT在一定程度上低估了风险，但重新校准到最近的SEER率改善了BCRAT的预测。BCRAT（也称为Gail模型2）使用1983-1987年SEER的年龄特异性乳腺癌发病率和竞争死亡率来预测浸润性乳腺癌的绝对风险。受20世纪90年代乳腺癌发病率变化的影响，我们对来自NIH-AARP队列（1995-2003）和PLCO筛选试验（1993-2006）的高加索绝经后妇女的模型校准进行了评估。我们通过比较Gail模型中预期的乳腺癌数(E)和观察到的乳腺癌数(O)来评估校准。然后，我们使用一个更新的模型来评估校准，该模型将Gail模型相对风险与我们的队列对应时期（1995-2003）的SEER浸润性乳腺癌发病率相结合。总体而言，Gail模型显著低估了浸润性乳腺癌的数量，在NIH-AARP中低估了13%，E/O=0.87 (95% CI: 0.85-0.89)，在PLCO中低估了14%，E/O=0.86 （95% CI: 0.82-0.90）。更新后的模型总体上校准良好：NIH-AARP的E/O=1.03 (95% CI: 1.00-1.05)， PLCO的E/O=1.01 （95% CI: 0.97-1.06）。PLCO的子集分析表明，当预测期限制在2003-2006年之间时，Gail模型在PLCO中得到了很好的校准。人们对确定SNP基因型是否可以为BCRAT等模型增加重要的预测价值很感兴趣。使用从美国4项队列研究和波兰1项病例对照研究中收集的5590例受试者和5998例50 - 79岁的对照受试者的数据，我们根据BCRAT中使用的危险因素信息和10种与乳腺癌相关的常见遗传变异拟合绝对风险模型。我们的结论是，新发现的遗传因素的纳入适度地增加了受体操作特征曲线下的面积，但在添加这些SNP基因型后，大多数女性的乳腺癌预测风险水平变化不大。另一项研究表明，在BCRAT中添加10个snp只会在以下应用中产生微小的改善：决定是否服用他莫昔芬预防乳腺癌；决定是否做乳房x光检查；在没有足够的资金对所有女性进行筛查的情况下分配乳房x光检查。相关工作表明，可预见的全基因组关联研究不太可能发现足够多的具有足够强关联的snp，从而进一步改进风险预测。我们分析了华盛顿德系犹太人研究（Washington Ashkenazi Study）的数据，以确定来自已知携带BRCA1或BRCA2基因突变的高风险家族的女性，即使被发现没有携带突变，患乳腺癌的风险是否高于平均水平。由于乳腺癌风险的大多数家族相关性不是由于BRCA1或BRCA2突变，并且由于大多数高风险家族是由于几个成员受到影响而确定的，因此有理由相信，这样的女性仍然比一般人群具有更高的风险，即使风险不像突变携带者那么高。我们的数据和文献回顾表明，这种残留的家族风险可以影响临床管理。BRCAPRO是一种基于乳腺癌和卵巢癌家族史来预测家族中谁携带BRCA1或BRCA2突变的模型。我们正在扩大BRCAPRO，以解释可能被误报的家族史。我们利用NCI-AARP和前列腺、肺癌、结直肠癌和卵巢癌（PLCO）筛查试验队列的数据，分别建立了乳腺癌、卵巢癌和子宫内膜癌的绝对风险预测模型。我们与哈佛大学的研究人员合作，在护士健康研究中验证了这些模型。手稿正在准备中。我们正在构建一个基于HPV检测的模型，以帮助指导有宫颈癌风险的妇女的诊断检测和治疗。我们估计了在北加州凯撒医疗机构接受筛查的33万名妇女的风险。我们正在开发一个统计模型，将筛选方案从自然历史中分离出来。这将使我们能够估计任何筛查方案的风险。我们提出了两个标准来评估预测疾病发病率风险的模型对筛查和预防的有用性，或评估疾病诊断后管理的预后模型的有用性。第一个标准，遵循PCF的病例比例(q)，是指将发病的个体的比例，包括在最高风险人群中个体的比例q中。第二个标准是需要随访的比例，即PNF(p)，即人们需要随访的处于最高风险的一般人群的比例，以便在注定成为病例的人群中有一定比例的人得到随访。PCF(q)评估一个项目的有效性，该项目跟踪了100%的高危人群。PNF(p)通过指出必须跟踪多少高危人群来评估覆盖100%病例的可行性。我们展示了这两个准则与洛伦兹曲线及其逆曲线的关系，并提出了PCF和PNF估计的分布理论。我们开发了基于影响函数的新方法，用于单个风险模型的推断，以及比较两个风险模型的pcf和pnf，这两个模型都是在相同的验证数据中进行评估的。在某些病例中，当死因信息缺失时，我们开发了预测SEER中死于意外癌症的绝对风险的归算方法。

We continued to develop, refine and evaluate the National Cancer Institutes Breast Cancer Risk Assessment Tool (BCRAT). Using data from the Asian American Breast Cancer Study, we obtained ethnicity-specific relative risks and attributable risks for Asian American women, and we coupled these with age- and ethnicity-specific breast cancer incidence rates from SEER to produce absolute risks. We are in the final stages of evaluating this model before publication and incorporation into BCRAT. With a summer Fellow , Mateo Banegas, who is a pre-doctoral student at the 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. BCRAT (also called Gail model 2) uses age-specific breast cancer incidence rates and competing mortality rates from SEER from 1983-1987 to predict the absolute risk of invasive breast cancer. Motivated by changes in breast cancer incidence during the 1990s, we evaluated the model's calibration in Caucasian, postmenopausal women from the NIH-AARP cohort (1995-2003), and the PLCO Screening Trial (1993-2006). We assessed calibration by comparing the number of breast cancers expected (E) from the Gail model with that observed (O). We then evaluated the calibration using an updated model that combined Gail model relative risks with SEER invasive breast cancer incidence rates from the period corresponding to our cohorts, 1995-2003. Overall, the Gail model significantly underpredicted the number of invasive breast cancers by 13% in NIH-AARP, E/O=0.87 (95% CI: 0.85-0.89) and by 14% in PLCO, E/O=0.86 (95% CI: 0.82-0.90). The updated model was well-calibrated overall: E/O=1.03 (95% CI: 1.00-1.05) in NIH-AARP and E/O=1.01 (95% CI: 0.97-1.06) in PLCO. Subset analyses in PLCO suggested that the Gail model was well-calibrated in PLCO when prediction period was restricted to between 2003-2006. There is interest in determining whether SNP genotypes can add important predictive value to models such as BCRAT. Using data collected from 5590 case subjects and 5998 control subjects between 50 to 79 years of age from four U.S. cohort studies and one case-control study from Poland, we fit models of absolute risk based on information about risk factors used in BCRAT and 10 common genetic variants associated with breast cancer. We concluded that the inclusion of newly discovered genetic factors modestly increased the area under the receiver-operating characteristic curve, but the level of predicted breast-cancer risk among most women changed little after the addition of these SNP genotypes. Another study showed that adding ten SNPs to BCRAT produced only minimal improvements in the following applications: deciding whether to take tamoxifen to prevent breast cancer; deciding whether to have a mammogram; and allocating mammograms when there is not enough money to screen all women. Related work showed that it is unlikely that foreseeable genome-wide association studies will discover enough SNPs with sufficiently strong associations to make substantial further improvements in risk prediction. We analyzed data from the Washington Ashkenazi Study to address whether a woman from a high risk family known to carry mutations in BRCA1 or BRCA2 genes had above average risk of breast cancer even if she was found not to carry a mutation. Because most of the familial correlation in breast cancer risk is not due to BRCA1 or BRCA2 mutations, and because most high risk families are ascertained because several members are affected, there is reason to believe that such a woman remains at higher risk than the general population, even though the risk is not as high as for a mutation carrier. Our data and review of the literature indicate that such residual familial risk can affect clinical management. BRCAPRO is a model to predict who in a family carries a mutation in BRCA1 or BRCA2, based on family history of breast and ovarian cancer. We are extending BRCAPRO to account for potentially misreported family history. We developed separate absolute risk prediction models for breast cancer, ovarian cancer, and endometrial cancer using data from the NCI-AARP and the Prostate, Lung, Colorectal and Ovarian Cancer (PLCO) Screening Trial cohorts. We validated the models in the Nurses Health Study, in collaboration with investigators at Harvard University. A manuscript is in preparation. We are constructing a model based on HPV testing to help guide diagnostic testing and treatment of women at risk of cervical cancer. We estimated risk from 330,000 women screened in Kaiser-Permanente Northern California. We are developing a statistical model that will separate the screening protocol from natural history. This will allow us to estimate risk for any screening protocol. We proposed 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. PCF(q) assesses the effectiveness of a program that follows 100q% of the population at highest risk. PNF(p) assess the feasibility of covering 100p% of cases by indicating how much of the population at highest risk must be followed. We showed the relationship of those two criteria to the Lorenz curve and its inverse, and present distribution theory for estimates of PCF and PNF. We developed new methods, based on influence functions, for inference for a single risk model, and also for comparing the PCFs and PNFs of two risk models, both of which were evaluated in the same validation data. We developed imputation methods for projecting absolute risk of dying from an incident cancer in SEER when cause of death information is missing in some cases.