无金标准下评价认知功能多状态量表准确度的贝叶斯统计方法研究

The scale is the preferred method for diagnosing cognitive function with multiple states. Since the gold standard cannot be obtained, it is urgent to develop new methods to correct bias in evaluating the accuracy of scales. The existing evaluation methods have limitations in measuring the correlation among different scales, screening the covariates and parameter identification when diagnosing cognitive function with multi-state in the absence of a gold standard. In the early stage, the applicant carried out some methodology study on evaluating the accuracy of diagnostic scales based on the binary latent true disease status in the absence of a gold standard. This project intends to study the methodology on evaluating the accuracy of scales for multi-state outcomes. Firstly, based on the scale data, the Copula function is used to measure the complex correlation structure among multiple scales, and the new indicators of accuracy evaluation are proposed, and the correlation is established for the different scales. Then the LASSO method is used to screen the variables, and we establish the model to describe the quantitative relationship between covariates and the accuracy of the scale; Secondly, the Bayesian algorithm is used to estimate the parameters in the models. Through a large number of simulation experiments, the optimization parameters are obtained under different correlation degrees. We use the cross-validation and Bootstrap method to verify the model. Finally, we use the26-years cohort study in Beijing elderly population and US open database for elderly cognitive function to validate our model. It provided a scientific basis for the evaluation of the accuracy of scales for the cognitive function with multi-state correctly in the absence of a gold standard, and provided statistical methodological support for similar research.

量表是诊断认知功能多状态的首选方法，由于无法得到金标准而亟需发展评价量表准确度新方法来校正偏倚。无金标准下认知功能多状态现有诊断评价方法在不同量表相关性度量、协变量筛选与参数识别方面均存在局限性。申请人前期开展了无金标准二分类量表准确度的评价方法学研究，本项目拟深化到对多状态结局量表准确度评价方法研究。首先基于量表数据，用Copula函数度量多个量表间复杂的相关结构，提出准确度评价新指标，为不同量表进行诊断评价建立相关性；然后用LASSO方法筛选变量，建立模型描述协变量与量表准确度的数量关系；其次基于贝叶斯算法对所建模型进行参数估计，通过大量模拟实验得到不同相关程度下的优化参数，采用交叉验证与Bootstrap方法进行模型验证；最后基于北京市老年人群26年队列与美国老年认知功能公开数据库进行实例验证，为无金标准下正确评价认知功能多状态量表准确度提供科学依据，并为同类研究提供统计方法学支撑。

量表是认知功能多状态诊断的首选工具，由于无法得到认知功能多状态金标准的诊断结果，导致现有量表准确度评价存在偏倚。无金标准下评价量表准确度过程中不同量表间相关性度量、协变量筛选均存在局限性。本项目基于北京市老年人群队列与美国ADNI公开数据库，研究简易精神状态检测量表(MMSE)、蒙特利尔认知评估量表(MoCA)及阿尔茨海默病评定量表(ADAS)诊断认知功能多状态的准确度，提出FGM(Farlie-Gumbel-Morgenstern) Copula新的相关结构度量MMSE与MoCA诊断结果间的相关关系。基于lasso变量筛选，建立带有协变量的FGM Copula模型描述协变量对准确度的影响效应。利用贝叶斯马尔科夫蒙特卡洛(MCMC)中Gibbs-Sampling算法估计所提出新模型中的参数。通过大量数值模拟验证模型参数估计的效果，比较了FGM Copula新的相关结构模型与两类其他相关关系模型在不同相关强度下的诊断性能。数据分析结果显示：无金标准下MoCA量表从轻度认知损害和正常人群中鉴别65岁以上老年人是否患有AD的总体诊断性能最高，其次为MMSE量表，ADAS总体诊断性能最低。用ROC（Receiver Operating Characteristic）曲面下的体积VUS(Volume under ROC Surface)指标来评价MMSE、MoCA和ADAS对认知功能三种状态的整体准确性，结果分别为：0.639,0.629和0.723。通过单因素分析和lasso变量选择，筛选出有意义的协变量因素，建立带有协变量的FGM Copula模型，结果显示:ApoE-ε4基因、年龄、教育程度对三种量表准确度的影响效应有统计学意义，而人种因素对三种量表准确度的影响效应无统计学意义。本研究能够在无金标准下正确评价量表对认知功能多状态的诊断准确度提供科学依据，并为同类研究提供统计方法学支撑。