Verbal Autopsy: Reimagining Data & Automated Cause Assignment (using ALPHA Network data)

 描述（由申请人提供）：通过衡量有多少人因各种疾病而结束或影响生命，确定全球卫生优先事项并评估进展情况。大多数低收入和中等收入国家没有足够的生命统计数据，导致全球三分之二的死亡没有登记，也没有原因。当无法进行全面尸检和死亡证明时，口头尸检（VA）有可能成为评估死因的可行且负担得起的方法。现有的 VA 方法无法产生可重复、可比较的死因分配，并且速度可能慢得令人无法接受。该项目旨在开发负担得起的、稳健的、校准的方法，通过完成以下四项活动来为个人死亡分配原因，并描述死亡原因的分布：(1) 开发一种新的标准、自动化统计方法，根据 VA 数据 - InSilicoVA 为死亡分配原因，(2) 使用额外的基于设施的数据和去偏医生分配的原因来创建新的、改进的症状原因信息模块，InSilicoVA 使用该模块来通知原因分配，(3) 进行详尽、公平的比较现有用于从 VA 数据中分配原因的自动化方法，包括 InSilicoVA，以及 (4) 通过在非洲的多个长期实地地点进行实地考察来减少 VA 数据的测量误差，以了解和标准化 VA 访谈，然后测试新的访谈设计和基于这种理解的语言处理程序的有效性。 InSilicoVA 将通过以下方式改进现有方法：(1) 完全概率性；(2) 保持个体和群体水平一致；(3) 报告两个水平上一致的不确定性；(4) 利用 A 访谈叙述部分的信息；(5) 考虑测量误差；(6) 纳入分层结构，以便跨广泛区域进行汇总和比较。在此项目结束时，VA 用户社区将拥有 (1) 对 InSilicoVA 的完整开源访问权限，(2) 用于 InSilicoVA 和其他方法的有价值的新症状原因信息模块，(3) 描述现有自动化 VA 原因分配方法性能的迫切需要的客观信息，以及 (4) 创新的新标准化 VA 访谈设计和标准操作程序，用于在 VA 访谈期间和之后处理语言和 VA 数据。总体效果将是增加使用具有可量化准确性的可重复方法确定的具有可比原因的死亡人数。反过来，这将为世界部分地区改善国家、区域和全球人口健康指标提供机会，而无需 适当的生命统计系统。有了更好、更及时的人口健康信息，就可以更有效、更有效地分配资源并采取有针对性的干预措施。更多人的生活将得到改善和延长。

 DESCRIPTION (provided by applicant): Global health priorities are set and progress assessed by measuring how many lives are ended or affected by various diseases. Most low and middle income countries do not have adequate vital statistics, with the result that two-thirds of global deaths are not registered and have no cause. Verbal autopsy (VA) has the potential to become a feasible, affordable method for assessing cause of death when full autopsy and death certification are not possible. Existing VA methods do not yield reproducible, comparable cause of death assignments and can be unacceptably slow. This project aims to develop affordable, robust, calibrated methods to assign causes to individual deaths and characterize the distribution of deaths by cause by completing the following four activities: (1) develop a new standard, automated, statistical method to assign causes to deaths from VA data - InSilicoVA, (2) use additional facility-based data and debiased physician-assigned causes to create new, improved symptom-cause information modules used by InSilicoVA to inform the assignment of causes, (3) conduct an exhaustive, fair comparison of existing automated methods for assigning causes from VA data, including InSilicoVA, and (4) reduce measurement error in VA data by conducting fieldwork at multiple long-term field sites in Africa to understand and standardize the VA interview and then test the effectiveness of a new interview design and language processing procedures informed by that understanding. InSilicoVA will improve on existing methods by (1) being fully probabilistic, (2) keeping individual and population levels consistent (3) reporting consistent uncertainty at both levels, (4) utilizing information from the narrative section of the A interview, (5) accounting for measurement error, and (6) incorporating a hierarchical structure to allow aggregation and comparison across broad regions. At the end of this project, the community of VA users will have (1) full, open-source access to InSilicoVA, (2) valuable, new symptom-cause information modules for use in InSilicoVA and other methods, (3) urgently needed, objective information describing the performance of existing automated VA cause-assignment methods, and (4) an innovative new standardized VA interview design and standard operating procedures for handling language and VA data during and after the VA interview. The overall effect will be to increase the number of deaths with comparable causes determined using a reproducible method with quantifiable accuracy. This, in turn, will provide an opportunity to improve national, regional and global population health indicators in parts of the world without adequate vital statistics systems. With better and more timely information on population health, resources can be allocated and interventions targeted more efficiently and with greater effect. More lives will be improved and extended.