CAREER: Equitable medical decision-making

在美国，巨大的健康不平等依然存在。甚至在COVID-19大流行之前。在该国的许多地区，收入较高的人比收入最低的人多活十年。此外，大流行病本身对低收入和服务不足人口的打击尤其严重。有偏见的医疗决策助长了这种健康不平等。例如，先前的研究表明，一种广泛使用的健康风险预测算法评估非裔美国人患者的病情比同等病情的白人患者轻。这项研究将通过统计分析人类和算法做出的决定，使医疗决策更加公平。这项研究将确定偏见的来源（例如，当医学测试提供给更容易获得医疗保健的患者而不是最有可能患病的患者时），并提出解决方案（例如，将测试重新分配给预计具有最高疾病风险的患者）。这不仅会使医疗更加公平；它还可以通过将医疗资源分配到最能发挥作用的地方来提高效率。该项目还将开设一个关于如何设计公平算法的公开课程，并开展一项大规模研究，研究如何培训工程师设计更公平的算法，以提高工程人员的准备能力。由于重要的医疗决策既由人类做出，也由算法做出，因此本研究追求三个目标：1)检测人类医疗决策中的偏见，重点关注三个高风险的医疗环境：医疗测试的分配、医疗质量评估和医学图像的解释。此外，该项目还将建立决策辅助算法，通过将临床医生的注意力吸引到他们可能忽视的医学相关特征上，来减少人类的偏见。最后，该项目的目标是通过检查适合包含在医疗算法中的特征，使算法决策更加公平。该研究将与临床医生合作进行，以最大限度地提高患者的转化效益。所开发的方法利用贝叶斯推理和深度学习技术来提供偏见产生的可解释模型，更普遍地适用于包括贷款和招聘在内的一系列高风险领域的决策，因此可以影响包括法律和经济学在内的与决策公平相关的广泛领域。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Enormous health inequality persists in the United States. Even prior to the COVID-19 pandemic. In many areas of the country, people with a higher income live up to decade longer than those in the lowest income levels. Additionally, the pandemic itself has hit low income and under-served populations especially hard. Biased medical decision-making contributes to this health inequality. For example, previous work has shown that one of widely used health risk prediction algorithms assesses African-American patients as less sick than equivalently sick White patients. This research will make medical decision-making fairer by statistically analyzing the decisions made both by humans and by algorithms. The research will identify sources of bias (for example, when medical tests are given to patients with better access to healthcare rather than to patients most likely to have a disease), and propose solutions (for example, reallocating tests to patients who are predicted to have the highest disease risk). This will not only make healthcare fairer; it can also make it more efficient, by allocating medical resources where they will do the most good. The project will also create a publicly available class on how to design fair algorithms, and conduct a large-scale study of how engineers can be trained to design fairer algorithms, to improve the preparedness of the engineering workforce.Because important medical decisions are made both by humans and by algorithms, the research pursues three objectives: 1) detecting bias in human medical decision-making, focusing on three high-stakes medical settings: allocation of medical testing, healthcare quality assessment, and interpretation of medical images. Further, the project will also build algorithmic decision-aids to reduce human bias, by drawing clinicians’ attention to medically relevant features they may have overlooked. Finally, the project targets making algorithmic decision-making more equitable, by examining the features it is appropriate to include in a medical algorithm. The research will be conducted in collaboration with clinicians to maximize translational benefit to patients. The methods developed, which draw on techniques in Bayesian inference and deep learning to provide interpretable models of how bias arises, are more generally applicable to decision-making across a host of high-stakes domains—including lending and hiring—and thus can impact a wide range of fields concerned with equity in decision-making, including law and economics.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.