Dynamic prediction incorporating time-varying covariates for the onset of breast cancer

项目总结 由于癌症的人口负担，准确的风险评估是肿瘤学的首要任务。乳腺癌 是全球女性中领先的癌症诊断，因此拥有最长和最广泛的关注点 关于风险预测。大多数传统预测模型只使用已知与以下因素相关的基线因素 乳腺癌风险。最近的模型扩展到更加重视基因组风险因素。然而， 添加基因组风险标记的主要举措结合了一种与时间不变的度量(基于 SNPs)，并不一定能解决改善乳腺癌风险分类的挑战。这个 随着时间的推移，患者之间和患者内部的内在异质性在一定程度上反映在时间变化的协变量上 这可能为乳腺癌风险的预测提供重要信息。积累 在乳腺癌方面有充分的记录，非常适合于结合时间的方法- 不同协变量。这项提案的目标是提供新的统计模型，该模型可以将 以个性化、动态的方式实现患者异质性，从而实现更准确的风险预测方案。 建议的算法包含创新的函数方法，以全面表征 通过一组独立于结果/无监督的结果和结果- 从属/受监督的功能。这组特定于个体的要素将包含有关观察到的 而不是现有方法中的一次性曝光，从而产生更高的预测能力。 动态预测模型将以逐步的方式建立，从单个时变协变量开始， 并扩展到多变量设置，以适应多个时变协变量。建议数 方法将应用于护士健康研究，并在Mayo进行进一步的外部评估 乳房X光摄影健康研究。所有建议的方法都将伴随着用户友好的开源 软件。

PROJECT SUMMARY Accurate assessment of risk is a top priority in oncology due to the population burden of cancer. Breast cancer is the leading cancer diagnosis among women worldwide and accordingly has the longest and broadest focus on risk prediction. Most traditional prediction models only utilize baseline factors known to be associated with breast cancer risk. More recent models expand to place greater emphasis on genomic risk factors. However, the predominant move of adding genomic risk markers incorporates a measure that is invariant to time (based on SNPs) and do not necessarily solve the challenge of improving breast cancer risk classification. The intrinsic heterogeneity between and within patients over time are reflected in part, by the time-varying covariate trajectories, which may provide important information for the prediction of breast cancer risk. The accumulation of cancer risk over life, well documented for breast cancer, is ideally suited to methods that incorporate time- varying covariates. Theobjective of this proposal is toprovide novel statistical models that can incorporate patient heterogeneity in a personalized, dynamic manner leading to a more accurate risk prediction scheme. The proposed algorithms encompass innovative functional approaches to comprehensively characterize the changing pattern of the longitudinal trajectories by a set of outcome-independent/unsupervised and outcome- dependent/supervised features. The set of individual-specific features will contain information on the observed time-varying `pattern' rather than one-time exposure in existing methods, leading to a higher predictive power. The dynamic prediction models will be built in a stepwise fashion, starting with a single time-varying covariate, and extended to the multivariate settings, to accommodate multiple time-varying covariates. The proposed methods will be applied to the Nurses' Health Study and further assessed externally in the Mayo Mammography Health Study. All of the proposed methods will be accompanied with user-friendly open-source software.