Artificial Intelligence and Counterfactually Actionable Responses to End HIV (AI-CARE-HIV)

摘要 佛罗里达州是美国人类免疫缺陷病毒(HIV)感染发生率最高的州， 社会和种族差异。在佛罗里达州，大约40%的艾滋病毒携带者没有达到无法检测到的病毒载量，而且 非洲裔美国人受到的影响最大。除了众所周知的社会人口因素导致 不利的结果和差距，其中的一部分仍然无法解释，也无法采取行动。预付款 在人工智能(AI)和大型真实世界数据(RWD)数据库的可用性不断增加方面，例如电子数据 健康记录(EHR)和管理索赔数据是开发精确健康模型的理想选择。 然而，人工智能的全部能力仍然受到这样一个事实的阻碍，即EHR没有很好地与其他 相关数据来源，包含关于健康的社会和行为决定因素的信息，特别是 对艾滋病毒护理、获取和结果很重要。此外，艾滋病毒结果的一个强有力的决定因素是耻辱，这是 没有在EHR的结构化字段中捕获，但可以通过自然语言处理在临床笔记中识别。 事实上，许多其他背景和个人层面的SDoH可以从EHR的临床叙述中提取出来。 建立在RWD基础上的人工智能的另一个关键问题是，由于EHR等观测数据的固有偏见， 人工智能模型可能会识别干预的错误影响。因此，另一种预测(即反事实) 天真的人工智能系统可能是错误的，可能会导致伤害。因果推理方法正在被 越来越多地与人工智能结合起来，以解决这种偏见。该项目的总体目标是开发“AI-CARE-- 艾滋病毒，“一个可操作的反事实的RWD人工智能框架，以改善佛罗里达州的艾滋病毒结果，特别是 通过解决SDoH来缩小差距。我们假设有一部分无法解释的系统性 差异可以通过结合因果推理和利用复杂交互作用的人工智能模型来阐明 在个人级别和情景级别的SDoH之间。然后，可以使用此框架来开发无偏见的(在 某些假设)，可用于规划和实施临床和公共卫生的可操作模式 干预措施。我们将通过One佛罗里达+临床研究联盟开发该项目，该联盟负责整理 RWD数据来自1680万佛罗里达州人，特别是One佛罗里达州+艾滋病毒队列(现在N=71,363)。我们的项目 旨在：(1)通过整合大规模SDoH(9,000+)来增强队列，并前瞻性地验证新的 SDoH，包括污名，使用NLP；(2)从SDoH创建多社会风险评分，确定人群层面的原因 SDOH条件干预对HIV结局的影响，并发展个性化的反事实人工智能 艾滋病毒结果模型，经过校准以减少差异；(3)计划--与医疗保健提供者、国家官员、 公民科学家-以我们的反事实人工智能模型为基础的有针对性的临床和公共卫生干预措施， 使用实施科学、标准化的协议(例如CONTORT-AI)。我们的团队包括多学科 (方法、临床、定性)由One佛罗里达+、Fl卫生部和少数族裔支持的专业知识- 服务实体。我们预计会在多个层面产生影响，从基础设施的加强到公共健康的好处。

ABSTRACT Florida has the highest incidence of Human Immunodeficiency Virus (HIV) infections in the US, with marked social and racial disparities. About 40% of people with HIV in Florida do not reach undetectable viral load, and Black African Americans are the most affected. Besides well-known sociodemographic factors contributing to unfavorable outcomes and disparities, part of such remains unexplained and cannot be actioned upon. Advances in artificial intelligence (AI) and increasing availability of large real-world data (RWD) databases, e.g., electronic health records (EHRs) and administrative claims data, are ideal for developing models for precision health. However, the full capabilities of AI are still hampered by the fact that EHRs are not well integrated with other relevant data sources, containing information on social and behavioral determinants of health (SDoH), especially important for HIV care access and outcomes. Further, a strong determinant of HIV outcomes is stigma, which is not captured in structured fields of EHRs, but can be identified in clinical notes via natural language processing. In fact, many other contextual- and individual-level SDoH can be extracted from clinical narratives in EHRs. Another critical problem with AI built on RWD is that, due to inherent bias in observational data like EHRs, the AI models might identify wrong effects for interventions. Thus, alternative predictions (i.e., counterfactuals) of naïve AI systems might be mistaken, potentially leading to harm. Causal inference methods are being increasingly coupled with AI to address such bias. The overarching goal of this project is to develop “AI-CARE- HIV,” an actionable counterfactual RWD AI framework to improve HIV outcomes in Florida, in particular reducing disparity through addressing SDoH. We hypothesize that a portion of the unexplained systemic disparity can be elucidated by combining causal inference and AI models that exploit complex interactions between individual- and contextual-level SDoH. This framework can then be used to develop an unbiased (under certain assumptions), actionable model usable for planning and implementing clinical and public health interventions. We will develop the project through the OneFlorida+ Clinical Research Consortium, which collates RWD data from >16.8M Floridians, and specifically the OneFlorida+ HIV cohort (now N=71,363). Our project aims to: (1) Enhance the cohort by incorporating large-scale SDoH (9,000+) and prospectively validate new SDoH, including stigma, using NLP; (2) Create polysocial risk scores from SDoH, identify population-level causal effects of SDOH-conditioned interventions on to HIV outcomes, and develop individualized counterfactual AI models for HIV outcomes, calibrated to reduce disparity; (3) Plan –with healthcare providers, State officials, citizen scientists– targeted clinical and public health interventions anchored on our counterfactual AI models, using implementation science, standardized protocols (e.g., CONSORT-AI). Our team includes multidisciplinary (methodological, clinical, qualitative) expertise supported by OneFlorida+, Fl Dept of Health, and minority- serving entities. We expect impact at multiple levels, from infrastructure enhancement to public health benefit.