The Center for Predictive Computational Phenotyping-1 Overall

描述(申请人提供)：随着我们在监测、记录、存储和整合表征人类生物学和健康的信息的能力方面的进步，生物医学科学正在发生根本性的变化，范围从单个分子到大量受试者。这些丰富的信息有可能极大地促进我们对人类生物学的理解，以及我们改善人类健康的能力。也许利用生物医学数据的最核心和最普遍的方法是使用机器学习和统计建模的方法来推断预测模型。这样的模型将表示某一感兴趣对象的可观测数据作为输入，并产生关于该对象的特定不可观测属性的预测作为输出。这种方法已被证明对广泛的生物医学任务具有很高的价值，但为了充分发挥预测建模的潜力，仍有许多重大挑战需要解决。为了应对这些挑战，我们建议建立预测计算表型中心(CPCP)。我们提议的中心将专注于广泛的问题，这些问题可以作为计算表型。虽然一些表型很容易测量和解释，并以可获得的形式提供，但广泛的科学和临床上重要的表型不符合这些标准。在这种情况下，需要计算表型方法来(I)从复杂的数据源或异类数据源的集合中提取相关的表型，(Ii)在它们展示之前预测临床上重要的表型，或者(Iii)在同一应用中同时进行这两项工作。

DESCRIPTION (provided by applicant): The biomedical sciences are being radically transformed by advances in our ability to monitor, record, store and integrate information characterizing human biology and health at scales that range from individual molecules to large populations of subjects. This wealth of information has the potential to substantially advance both our understanding of human biology and our ability to improve human health. Perhaps the most central and general approach for exploiting biomedical data is to use methods from machine learning and statistical modeling to infer predictive models. Such models take as input observable data representing some object of interest, and produce as output a prediction about a particular, unobservable property of the object. This approach has proven to be of high value for a wide range of biomedical tasks, but numerous significant challenges remain to be solved in order for the full potential of predictive modeling to be realized. To address these challenges, we propose to establish The Center for Predictive Computational Phenotyping (CPCP). Our proposed center will focus on a broad range of problems that can be cast as computational phenotyping. Although some phenotypes are easily measured and interpreted, and are available in an accessible format, a wide range of scientifically and clinically important phenotypes do not satisfy these criteria. In such cases, computational phenotyping methods are required either to (i) extract a relevant phenotype from a complex data source or collection of heterogeneous data sources, (ii) predict clinically important phenotypes before they are exhibited, or (iii) do both in the same application.