Developing an innovative statistical framework to integrate multiple verbal autopsy datasets to estimate cause-specific mortality

项目摘要 死亡原因数据对于了解疾病负担、新出现的卫生需求以及 公共卫生干预措施的有效性。低收入和中等收入国家（LMIC）很少有足够的生命力。 提供高质量死因统计数据的统计系统。口头尸检（VA）是一种广泛采用的 在无法进行全面尸检和颁发死亡证明时收集死因信息的工具。的 目前VA的分析方法受到缺乏普遍性的严重限制。现有的VA方法产生 死亡原因分配不准确，部署时对死亡分布的估计有偏差 与开发模型所基于的群体不同的群体。在这个项目中， 我们将开发强大的，域自适应的，计算上可行的方法来分配个人的原因， （一）建立统计模型， 表征多个异构VA数据集;（ii）开发和评估域自适应算法 （iii）扩展统一的艾德域适应框架 到常规VA分析管道。这个新的框架将通过利用完整的 多个人群的参考死亡中的可用信息，以实现对数据转移的稳健性 人口。该框架还将把复杂的依赖关系纳入所收集的迹象， 以可解释的方式描述症状，并允许快速和简化的实施， 退伍军人事务部问卷。我们将开发第一个统一的艾德框架，用于领域自适应死因分配 使用VA数据，并对VA收集的体征和症状之间的关系提供重要见解， 死因该项目将为未来的研究奠定基础，例如将VA与其他 从病史或组织样本中收集的协变量和生物标志物信息，并设计系统 使用大规模VA调查进行死因监测和监督。

Project Abstract Cause of death data are essential for understanding the burden of disease, emerging health needs, and the effectiveness of public health interventions. Few low- and middle-income countries (LMIC) have adequate vital statistics systems that produce high quality statistics on causes of death. Verbal autopsy (VA) is a widely adopted tool to collect information on causes of death when full autopsy and death certification are not possible. The current analytical methods for VA are significantly limited by the lack of generalizability. Existing VA methods yield inaccurate cause-of-death assignment and biased estimates of the distribution of deaths when they are deployed to populations that are different than the populations based on which the models are developed. In this project, we will develop robust, domain adaptive, and computationally feasible methods to assign causes to individual deaths and estimate cause-specific mortality, by completing the following aims: (i) to develop statistical models to characterize multiple heterogeneous VA datasets; (ii) to develop and evaluate domain adaptive algorithms for cause-of-death assignment in new populations; and (iii) to extend the unified domain adaptation framework to routine VA analysis pipeline. This new framework will improve on existing VA methods by utilizing the full information available in reference deaths from multiple populations to achieve robustness to data shift across populations. The framework will also incorporate the complex dependence relationship in the collected signs and symptoms in an interpretable manner, and allow fast and streamlined implementation compatible with standard VA questionnaires. We will develop the first unified framework for domain adaptive cause-of-death assignment using VA data and offer critical insights into the relationship between the signs and symptoms collected by VA and causes of death. The project will lay the groundwork for future research, such as integrating VAs with additional covariates and biomarker information collected from medical history or tissue samples, and designing systematic cause-of-death monitoring and surveillance using large-scale VA surveys.