The RCMI Program in Health Disparities at Meharry Medical College - Supplement

项目摘要 物质使用障碍（SUD）是一个主要的公共卫生问题，最近在美国增加了一倍多。 在过去几年中，美国人的患病率从2018年的2000万上升到4630万 2021年受SUD影响的人不成比例地经历健康的负面社会决定因素（SDoH）， 包括安全住房、交通、教育、就业机会和营养不足， 食品这些SDoH与低自尊，自我效能和SUD禁欲的失败尝试有关。 收集和整合SDoH信息，作为患者电子健康记录（EHR）的一部分，用于临床 建模可以帮助揭示与SUD相关的患者体验和行为，但大多数SDoH都是 嵌入非结构化自由文本。此外，SUD诊断在结构EHR中的代表性不足。 因此，需要一种自动化且更准确的方法来提取SDoH并识别SUD。自然 语言处理（NLP）可以解锁临床叙述中传达的信息，从而发挥关键作用 在现实世界的研究中。目前正在开发各种方法和工具，以便利这种提取;然而，这些工具 对于特定于队列的样本仍在研究中，非结构化文本分析很复杂。以促进健康 差异研究，并提高对服务不足的SUD患者的患者特征的理解 在Meharry的人口中，探索机器学习工具以提取SDoH因素并识别 SUD诊断。在本提案中，我们将解决SDoH提取和SUD识别的挑战 从非结构化的临床记录或患者调查中生成一致的框架， 识别、理解、治疗和预测SUD和相关结局（例如复发）。我们 我将通过开发两个NLP管道来解决这一挑战，一个用于SDoH提取，一个用于SUD 识别.我们将重点关注可卡因，大麻和阿片类药物使用障碍的SUD患者。SDoH NLP 管道将挖掘和丰富CDC定义的五个SDoH领域的数据，包括经济稳定，教育， 卫生保健可及性、邻里环境以及社会和社区背景。命名实体识别 方法将在SUDs NLP管道中进行研究。我们将在一个队列中开发和测试这两条管道， 200-500例患者，并将我们的管道导出结果与手动审查结果进行比较，以衡量NLP模型 性能我们假设将开发一个高性能的SDoH NLP管道， 应用，并将识别更多SUD患者。我们研究的最终目标是确定患者在 SUD风险，识别与健康结局相关的风险因素，并改善患者健康结局 预测，这将可能有助于临床决策支持和医疗管理。为了实现 因此，我们将整合从两个NLP中获得的SDoH相关因素和SUD诊断 管道，沿着从EHR结构化数据中提取的其他表型风险因素。我们假设 SUD预测模型的性能将在包括SDoH之后增加。

PROJECT SUMMARY Substance use disorders (SUDs) are a major public health issue that has recently more than doubled in prevalence among Americans in the last few years, from affecting 20 million in 2018 to 46.3 million Americans in 2021. Those affected by SUD disproportionately experience negative social determinants of health (SDoH), including inadequate access to safe housing, transportation, education, employment opportunities, and nutritious foods. These SDoH are connected with low self-esteem, self-efficacy, and failed attempts at SUD abstinence. Collecting and integrating SDoH information as part of patients’ electronic health records (EHR) for clinical modeling could help uncover patient experiences and behaviors related to SUD, but the majority of SDoH are embedded in unstructured free text. Additionally, the SUD diagnosis is under represented in structural EHR. Therefore, an automated and more accurate approach is needed to extract SDoH and identify SUDs. Natural language processing (NLP) can unlock the information conveyed in clinical narratives, thus playing a critical role in real-world studies. Methods and tools are being developed to facilitate such extractions; however, these tools are still under study for cohort-specific samples and unstructured text analytics is complex. To advance health disparity studies and improve understanding of patient characteristics of the SUD patients from the underserved population at Meharry, it is a high priority to explore machine learning tools to extract SDoH factors and identify SUDs diagnosis. In this proposal, we will address the challenge of SDoH extraction and SUDs identification from unstructured clinical notes or patient surveys to generate a consistent framework that can aid in identifying, understanding, treating, and predicting SUDs and associated outcomes (e.g. relapse). We will solve this challenge by developing two NLP pipelines, one for SDoH extraction and one for SUDs identification. We will focus on SUD patients with cocaine, cannabis, and opioid use disorders. The SDoH NLP pipeline will mine and enrich data in five SDoH domains defined by CDC including economic stability, education, health care access, neighborhood environment, and social and community context. Name entity recognition methods will be investigated in the SUDs NLP pipeline. We will develop and test the two pipelines in a cohort of 200-500 patients and compare our pipeline-derived results to manual review outcome to measure the NLP model performance. We hypothesize that a high performance SDoH NLP pipeline will be developed that will fit to our application, and more SUD patients will be identified. The ultimate goal of our studies is to identify patients at risk of SUDs, identify risk factors associated with health outcomes, and improve patient health outcome prediction that will potentially help clinical decision support and healthcare management. In order to accomplish this, we will integrate the SDoH associated factors and SUD diagnosis that we obtained from the two NLP pipelines, along with other phenotype risk factors being extracted from EHR structured data. We hypothesize that the performance of the SUD predictive model will be increased after SDoH are included.