Leveraging an electronic medical record infrastructure to identify primary care patients eligible for genetic testing for hereditary cancer and evaluate novel cancer genetics service delivery models

增刊摘要 本申请是根据NOT-OD-22-026作为拟议的行政补充提交的 犹他州大学(犹他州)/纽约大学(NYU)U01题为《利用电子医疗 记录基础设施，以确定有资格进行遗传性癌症基因检测的初级保健患者和 评估新的癌症遗传学服务提供模式“(U01 CA232826)。母公司U01正在使用 基于电子健康记录(EHR)的临床决策支持(CDS)基础设施：(I)识别未受影响的患者 犹他州和纽约大学医疗系统中有资格获得癌症遗传学服务的初级保健患者 关于当前指导方针(目标1)；和(2)比较两种癌症遗传学服务提供模式，以确定 患者参加随机临床试验(目标2和3)。在父母的研究中，我们确定了22,208人的队列 两个医疗系统中有资格获得癌症基因服务的初级保健患者，我们是 规划该项目的可持续性。然而，我们之前的数据显示了种族、民族和 池中由当前CDS算法标识的语言首选项与基础主要项的比较 护理病人群体。我们发现，至少一个促成因素是系统的差异 结构化电子病历领域中癌症家族史信息的可用性和全面性 这是当前算法所基于的。这些差异引发了与由此产生的偏见相关的关键伦理问题 来自CDS算法的集成。因此，我们建议：(补充目标1)调查是否 将自然语言处理(NLP)工具结合到CDS算法中可以减少 确定符合条件的初级保健患者。我们已经创建了一种NLP增强算法，该算法结合了 自由文本评论和识别的患者比当前的CDS算法多54%。我们将研究是否 NLP增强算法的使用会影响按种族、民族、语言进行识别的差异 偏好，以及农村/边境居住地与潜在患者人口的比较。(补充目标2) 确定几乎符合癌症基因评估标准的患者，并探索扩展到 这些患者来自医疗服务不足的社区。我们将开发一个统计模型，可以确定 几乎符合癌症基因评估标准的患者，这将使我们能够针对 收集其他家族历史信息。我们还将举办两个社区参与工作室，其中一个在 英语和西班牙语，以检查通过算法识别后直接联系的可接受性 在来自医疗服务不足的社区的个人中。这些补充目标将共同调查 两种可能的方法来解决观察到的患者识别方面的差异。这些发现将直接 通知制定明确包括监控CDS算法和人工智能工具影响的策略 关于患者的结果以及在测试和实施的迭代阶段中这些结果的差异。

SUPPLEMENT ABSTRACT This application is being submitted in response to NOT-OD-22-026 as a proposed administrative supplement to the University of Utah (Utah)/New York University (NYU) U01 entitled “Leveraging an electronic medical record infrastructure to identify primary care patients eligible for genetic testing for hereditary cancer and evaluate novel cancer genetics service delivery models” (U01 CA232826). The parent U01 is employing an electronic health record (EHR)-based clinical decision support (CDS) infrastructure to: (i) identify unaffected primary care patients in the Utah and NYU healthcare systems who qualify for cancer genetics services based on current guidelines (Aim 1); and (ii) compare two models of cancer genetics services delivery for identified patients in a randomized clinical trial (Aims 2 and 3). In the parent study, we have identified a cohort of 22,208 primary care patients in the two healthcare systems who are eligible for cancer genetic services, and we are planning for sustainability of this project. However, our prior data has shown disparities by race, ethnicity, and language preference in the pool identified by the current CDS algorithm compared to the underlying primary care patient populations. We have found that at least one contributing factor is systematic disparities in availability and comprehensiveness of cancer family history information available in the structured EHR fields upon which the current algorithm is based. These disparities raise critical ethical issues related to bias resulting from integration of the CDS algorithm. We therefore propose to: (Supplemental Aim 1) Investigate whether incorporating natural language processing (NLP) tools into the CDS algorithm reduces disparities in identification of eligible primary care patients. We have created an NLP-augmented algorithm that incorporates free-text comments and identifies 54% more patients than the current CDS algorithm. We will examine whether use of the NLP-augmented algorithm impacts disparities in identification by race, ethnicity, language preference, and rural/frontier residence compared with the underlying patient population. (Supplemental Aim 2) Identify patients who nearly meet criteria for cancer genetic evaluation and explore acceptability of outreach to those patients from medically underserved communities. We will develop a statistical model that can identify patients who nearly meet criteria for cancer genetic evaluation, which would allow us to target outreach to collect additional family history information. We will also conduct two community engagement studios, one in English and one in Spanish, to examine the acceptability of direct outreach after identification by an algorithm among individuals from medically underserved communities. Together these supplemental aims will investigate two potential approaches to address observed disparities in patient identification. These findings will directly inform the development of policies that explicitly include monitoring the impact of CDS algorithms and AI tools on patient outcomes and disparities in those outcomes during iterative phases of testing and implementation.