[论文解读] Fitting and projecting HIV epidemics: Data, structure and parsimony

Brian Williams|arXiv (Cornell University)|Nov 22, 2014

HIV/AIDS Research and Interventions参考文献 14被引用 4

一句话总结

本文提出一种结构化且简洁的建模方法，用于拟合并预测一般性HIV流行病趋势，数据来源为肯尼亚。研究评估了模型结构（如风险异质性、行为变化、威布尔生存分布和人口学细节）的影响，发现生存分布的选择显著影响发病率和死亡率的预测结果，而人口学复杂性带来的收益相对于计算成本而言微乎其微。

ABSTRACT

Understanding historical trends in the epidemic of HIV is important for assessing current and projecting future trends in prevalence, incidence and mortality and for evaluating the impact and cost-effectiveness of control measures. In generalized epidemics the available data are of variable quality among countries and limited mainly to trends in the prevalence of HIV among women attending ante-natal clinics. In concentrated epidemics one needs, at the very least, time trends in the prevalence of HIV among different risk groups, including intravenous drug users, men-who-have-sex-with-men, and commercial sex workers as well as the size of each group and the degree of overlap between them. Here we focus on the comparatively straight forward problems presented by generalized epidemics. We fit data from Kenya to a susceptible-infected model and then successively add structure to the model, drawing on our knowledge of the natural history of HIV, to explore the effect that different structural aspects of the model have on the fits and the projections. Both heterogeneity in risk and changes in behaviour over time are important but easily confounded. Using a Weibull rather than exponential survival function for people infected with HIV, in the absence of treatment, makes a significant difference to the estimated trends in incidence and mortality and to the projected trends. Allowing for population growth has a small effect on the fits and the projections but is easy to include. Including details of the demography adds substantially to the complexity of the model, increases the run time by several orders of magnitude, but changes the fits and projections only slightly and to an extent that is less than the uncertainty inherent in the data. We make specific recommendations for the kind of model that would be suitable for understanding and managing HIV epidemics in east and southern Africa.

研究动机与目标

• 开发一种简洁但结构化明确的模型，用于拟合并预测一般性流行病中的HIV流行趋势。
• 评估不同结构组件（如风险异质性、行为改变和生存分布）对模型拟合和预测的影响。
• 评估模型复杂性与准确性的权衡，特别是人口学细节的影响。
• 为数据质量参差不齐的资源有限环境中的HIV流行病建模提供实用建议。

提出的方法

• 将易感-感染者分层模型拟合至肯尼亚产科诊所的HIV患病率数据。
• 系统性地引入结构化要素：风险异质性、随时间变化的行为、威布尔分布与指数分布的生存函数比较。
• 比较逐步复杂化的模型版本在模型性能和预测结果上的差异。
• 引入人口增长和详细的人口学结构，以评估其对模型拟合和计算成本的影响。
• 采用基于似然的方法评估模型拟合度及对结构假设的敏感性。
• 将模型应用于不同结构假设下，预测发病率、患病率和死亡率趋势。

实验结果

研究问题

• RQ1 在无治疗条件下，HIV生存分布假设（指数分布 vs. 威布尔分布）对估计的发病率和死亡率趋势有何影响？
• RQ2 行为改变和风险异质性在多大程度上会混淆对一般性流行病中HIV患病率趋势的解读？
• RQ3 引入人口增长如何影响模型拟合和预测准确性？
• RQ4 在拟合度和预测能力方面，增加详细的人口学结构所能带来的边际收益是多少？
• RQ5 哪种模型结构在准确性、计算效率和数据需求之间提供了最佳平衡，适用于公共卫生规划？

主要发现

• 与使用指数分布相比，采用威布尔生存分布显著改变了HIV发病率和死亡率趋势的估计结果，尤其在无治疗条件下更为明显。
• 风险异质性和随时间的行为改变均具有重要意义，但在模型拟合中容易相互混淆，因此需要仔细的结构化设定。
• 引入人口增长对模型拟合和预测结果有微小但可测量的影响，且计算实现简单直接。
• 增加详细的人口学结构会使模型复杂度和运行时间提高数个数量级，但对拟合度和预测结果的改善仅在数据不确定性范围内，提升有限。
• 本研究结论认为，一种中等结构化的模型——整合风险异质性、行为改变和威布尔生存分布——在东非和南部非洲的HIV流行病建模中，提供了准确性与实用性的最佳平衡。