Rountree Lillian, Lin Yi-Ting, Liu Chuyu, Salvatore Maxwell, Admon Andrew, Nallamothu Brahmajee, Singh Karandeep, Basu Anirban, Bu Fan, Mukherjee Bhramar
Department of Biostatistics, School of Public Health, University of Michigan, Ann Arbor, MI, United States.
Department of Epidemiology and Biostatistics, College of Human Medicine, Michigan State University, East Lansing, MI, United States.
Online J Public Health Inform. 2025 Mar 19;17:e66598. doi: 10.2196/66598.
Clinical risk prediction models integrated into digitized health care informatics systems hold promise for personalized primary prevention and care, a core goal of precision health. Fairness metrics are important tools for evaluating potential disparities across sensitive features, such as sex and race or ethnicity, in the field of prediction modeling. However, fairness metric usage in clinical risk prediction models remains infrequent, sporadic, and rarely empirically evaluated.
We seek to assess the uptake of fairness metrics in clinical risk prediction modeling through an empirical evaluation of popular prediction models for 2 diseases, 1 chronic and 1 infectious disease.
We conducted a scoping literature review in November 2023 of recent high-impact publications on clinical risk prediction models for cardiovascular disease (CVD) and COVID-19 using Google Scholar.
Our review resulted in a shortlist of 23 CVD-focused articles and 22 COVID-19 pandemic-focused articles. No articles evaluated fairness metrics. Of the CVD-focused articles, 26% used a sex-stratified model, and of those with race or ethnicity data, 92% had study populations that were more than 50% from 1 race or ethnicity. Of the COVID-19 models, 9% used a sex-stratified model, and of those that included race or ethnicity data, 50% had study populations that were more than 50% from 1 race or ethnicity. No articles for either disease stratified their models by race or ethnicity.
Our review shows that the use of fairness metrics for evaluating differences across sensitive features is rare, despite their ability to identify inequality and flag potential gaps in prevention and care. We also find that training data remain largely racially and ethnically homogeneous, demonstrating an urgent need for diversifying study cohorts and data collection. We propose an implementation framework to initiate change, calling for better connections between theory and practice when it comes to the adoption of fairness metrics for clinical risk prediction. We hypothesize that this integration will lead to a more equitable prediction world.
集成到数字化医疗信息系统中的临床风险预测模型有望实现个性化的一级预防和护理，这是精准医疗的核心目标。公平性指标是评估预测建模领域中性别、种族或民族等敏感特征之间潜在差异的重要工具。然而，临床风险预测模型中公平性指标的使用仍然不常见、不规律，且很少进行实证评估。
我们试图通过对两种疾病（一种慢性病和一种传染病）的流行预测模型进行实证评估，来评估公平性指标在临床风险预测建模中的应用情况。
2023年11月，我们使用谷歌学术对近期关于心血管疾病（CVD）和新冠肺炎的临床风险预测模型的高影响力出版物进行了范围界定文献综述。
我们的综述产生了一份23篇聚焦心血管疾病的文章和22篇聚焦新冠肺炎疫情的文章的入围名单。没有文章评估公平性指标。在聚焦心血管疾病的文章中，26%使用了按性别分层的模型，在有种族或民族数据的文章中，92%的研究人群中超过50%来自同一个种族或民族。在新冠肺炎模型中，9%使用了按性别分层的模型，在包含种族或民族数据的模型中，50%的研究人群中超过50%来自同一个种族或民族。两种疾病的文章都没有按种族或民族对模型进行分层。
我们的综述表明，尽管公平性指标能够识别不平等并指出预防和护理中的潜在差距，但用于评估敏感特征差异的情况却很少见。我们还发现，训练数据在很大程度上仍然在种族和民族上同质化，这表明迫切需要使研究队列和数据收集多样化。我们提出了一个实施框架来启动变革，呼吁在采用临床风险预测公平性指标时，理论与实践之间建立更好的联系。我们假设这种整合将带来一个更加公平的预测世界。
