基于认知功能多维测量和潜在结构的个体化痴呆风险动态预测模型研究

As a major public health issue, dementia will not only endanger health of the elderly, but put burdens heavily on patients, caregivers, societies and health care systems as well. Personalized cognitive degradation risk dynamic prediction models may enable early detection of the subjects that are subject to risk of cognitive decline so as to lay the theoretical foundation for precision medicine and optimally allocating the limited health-care resources. Based on the cognition epidemiological cohort study among the community-dwelling elderly people, we built multi-state models to determine the different effects of known risk factors in the cognitive impairment process and to make prediction. The following issues should be considered simultaneously. Cognition which is the central longitudinal process in dementia natural history is not directly observed but measured by multiple psychometric tests collected repeatedly at cohort visits, which is longitudinal process. In addition, cognitive decline is very associated with onset of dementia or death, which is survival process. A further issue that should be accounted for is the strong unobservable heterogeneity between subjects that may not be explained by the covariates. A population of elderly subjects usually mixes groups of subjects with different types of cognitive trajectory (i.e. latent classes). This study aims to build a joint latent class mixed model for jointly analyzing longitudinal process and survival process to further explain heterogeneity between subjects by latent classes, to calculate individual dementia risk dynamic prediction probabilities based on cognitive measures, and to evaluate the predictive abilities for different model built strategies. This project will provide statistical support for health-care providers to identify subjects who may have unrecognized cognitive impairment or undiagnosed dementia among the elders. It will also provide model built references for multiple longitudinal markers with latent trait character and multiple terminal events.

老年痴呆是严重危害老年人生命健康并造成极大负担的重大公共卫生问题，个体化痴呆风险动态预测有助于高危人群早期筛查，是精准医学和有限医疗卫生资源最优化分配的依据。项目组前期基于社区老年人群认知流行病学队列研究，采用多状态模型探讨了传统危险因素在痴呆自然史不同阶段的不同作用以及基于危险因素的转归预测。考虑到具有潜变量特质的认知功能只能由多个神经心理学测试量表在不同时间点重复测量间接观察，构成纵向过程；认知与终点事件（痴呆/死亡）有关，构成生存过程；同时老年人群是一个具有强异质性的总体，协变量不能完全解释该异质性。本项目构建基于潜在结构的纵向过程和生存过程的联合模型，通过潜在类别进一步解释总体异质性，计算基于认知功能多维测量的个体痴呆风险动态预测概率，并进行不同建模策略预测效能比较。研究结果可为识别认知干预重点人群提供统计决策支持，并为其他具有潜在特质属性的纵向标记物和多终点事件统计建模提供参考。

老年痴呆特别是阿尔茨海默病（Alzheimer's Disease，AD）已成为严重危害老年人生命健康并造成极大负担的重大公共卫生问题。基于前期社区老年人群的认知流行病学队列研究和美国阿尔茨海默病神经影像学倡议（ADNI）公开数据库，本项目研究内容包括：（1）老年人潜在认知衰退过程：基于老年人群不同潜在类别的协变量和结局变量的特点，分析协变量对认知评估量表（阿尔茨海默病评估量表ADAS-11，临床痴呆评分总和量表CDR-SB，功能活动调查表FAQ及简易精神状态量表MMSE）结果的影响差异，揭示认知评估量表的不同应用特点；探讨老年人认知衰退的轨迹，评估协变量对认知变化的影响。（2）阿尔茨海默病风险预测：基于AD疾病自然史，采用贝叶斯网络模型、联合潜在类别模型和Landmark模型，实现疾病因果推断和个体风险预测，探讨老年人认知功能的影响因素，比较不同纵向标记物对AD风险预测的准确性。（3）痴呆患者生命质量评价：采用结构方程模型和纵向多分类项目反应理论，从量表的维度和条目两水平分析AD患者生命质量评价中的反应转移发生情况，并评价生命质量真实的改变；对帕金森病（Parkinson’s Disease，PD）患者构建以日常生活能力和抑郁作为中介变量的链式多重中介模型和自回归中介模型，探讨PD患者生命质量的影响机制。（4）痴呆患者照料者负担评价：采用潜调节结构方程方法和偏差校正的非参数百分位Bootstrap方法，并引入多重中介效应，分别深入分析AD和PD患者因素、照顾者因素和社会因素对照顾者负担的影响过程和作用机制。本项目旨在探讨老年认知衰退轨迹和影响因素，进行个体化痴呆风险动态预测；同时对痴呆患者生命质量及痴呆患者照料者负担进行评价，研究结果可为老年人认知损害个体化预防指导和早期干预提供可靠的流行病学评价，为疾病监测和公共卫生政策制定提供强有力的统计决策支持。