Novel Algorithmic Fairness Tools for Reducing Health Disparities in Primary Care

项目摘要：医疗保健系统的差异是巨大的，导致更糟糕的健康结果 以及对边缘化群体的护理质量。这些差异反映出我们目前的卫生系统 不公平的均衡医疗保健数据中隐藏着社会偏见，包括种族主义和 为社会经济背景低和农村地区的个人提供护理。但不少 算法方法不足以解决健康差异，因为算法不 评估或优化这些组的性能。现有的工具，以改善不同的性能， 在现实的医疗保健环境中，多个边缘化群体极为有限。我们的创新方法， 健康差异中的数据和算法偏见问题是创造一个前所未有的 算法公平性框架为多个边缘化群体。在初始阶段，我们将专注于数据 转换-干预数据，以便“去偏置”它，以代表所需的平衡，而不是 强化不公平的均衡。第二阶段建立新的公平回归估计，以加强公平性 预测的限制。我们的目标是创建可重复使用的工具，以促进健康的公平提供 在乎我们将通过开发遵循道德管道的可推广方法来实现这一目标， 由健康框架的社会决定因素指导的算法。我们的具体目标是：（1）开发和测试 新的数据转换方法，依赖于微观模拟去偏置医疗保健数据，（2）发展 并测试新的公平惩罚回归方法优化多组，（3）测试的性能， 慢性肾脏疾病优先级排序中高影响力初级保健应用的新算法框架 公平对待面临健康差距的多个种族和民族群体，以及（4）创建开源 计算工具，教程插图，以及用于可重复研究的合成数据资源， 传播。拟议的研究将产生一个统计创新的可重用算法的公平 统一数据转换和公平回归的框架，以减少健康差距， 慢性肾脏病护理质量研究。这个初级保健应用程序将利用丰富的注册表数据， 包括对健康的社会决定因素的测量，在通常的护理环境中收集， 在地理上，种族，和种族多样化的人口跨越多个付款人。我们的方法中心 采用严格的方法设计的稳健性，包括与现有替代估计量的比较， 综合模拟研究和国家真实世界登记册数据的标准做法。解决健康 初级保健是一个持续、协调的保健中心， 通过卫生保健系统改善公共卫生。我们框架的广泛适用性和 可重复使用的计算工具将有助于在许多实际环境中的部署。

PROJECT SUMMARY: Disparities in the health care system are substantial, leading to worse health outcomes and quality of care for marginalized groups. These disparities reflect that our current health system has an inequitable equilibrium. Imbedded within health care data are societal biases, including racism and barriers in access to care for individuals from low socioeconomic backgrounds and rural areas. However, many algorithmic approaches are inadequate for addressing health disparities because the algorithms do not evaluate or optimize performance in these groups. Existing tools to ameliorate differential performance for multiple marginalized groups in realistic health care settings are extremely limited. Our innovative approach to the data and algorithmic bias problems in health disparities is to create a first-of-its-kind overarching algorithmic fairness framework for multiple marginalized groups. In the initial phase, we will focus on data transformations—intervening on the data in order to ‘de-bias’ it to represent a desired equilibrium rather than reinforcing the unfair equilibrium. The second stage builds novel fair regression estimators to enforce fairness constraints for prediction. Our goal is to create reusable tools that advance the equitable provision of health care. We will accomplish this by developing generalizable methodology that follows an ethical pipeline for algorithms guided by a social determinants of health framework. Our specific aims are to: (1) develop and test novel data transformation methods that rely on microsimulations for de-biasing health care data, (2) develop and test new fair penalized regression approaches optimized for multiple groups, (3) test the performance of the new algorithmic framework for a high-impact primary care application in chronic kidney disease prioritizing fairness for multiple racial and ethnic groups facing health disparities, and (4) create open-source computational tools, tutorial vignettes, and a synthetic data resource for reproducible research and dissemination. The proposed research will yield a statistically innovative reusable algorithmic fairness framework unifying data transformations and fair regression to reduce health disparities with robust testing in a chronic kidney disease study of quality of care. This primary care application will leverage rich registry data, including measurements of social determinants of health, collected in usual care settings from a geographically, racially, and ethnically diverse population across multiple payers. Our approach centers robustness with rigorous methodological design, including comparisons to alternative existing estimators and standard practice in comprehensive simulation studies and national, real-world registry data. Addressing health disparities in primary care—a hub of continuous, coordinated care—has the potential for substantial impact on improving public health via the health care system. The broad applicability of our framework and creation of reusable computational tools will facilitate deployment in many practical settings.