Developing a Clinical Decision Support Tool that Assesses Risk of Opioid Use Disorder Using Natural Language Processing, Machine Learning, and Social Determinants of Health from Clinical Notes

项目总结/摘要 2017年，有170万美国人患有阿片类药物使用障碍（OUD），导致4.7万美国人死亡 阿片类药物过量社会健康决定因素（SDoH）影响患者的OUD风险水平， 医生的阿片类药物处方。医生缺乏快速准确评估SDoH相关的工具 缺乏相关干预资源的知识。临床决策支持（CDS）可以 快速评估患者与OUD风险相关的SDoH因素，并提供可行的建议， 这将减少OUD风险评估时间并解决知识差距。2018年，UCSF的研究人员 创建了《社会风险因素医学术语代码汇编》，将SDoH风险映射到 医学词汇然而，大多数SDoH都记录在临床记录中。我的长期职业目标是 研究独立性，具有以下专业知识：1）OUD风险评估，2）SDoH研究，3）干预 开发、实施和评估。与这些目标相关，本研究将使用自然语言 处理（NLP）以识别临床笔记中的SDoH，检查SDoH和OUD之间的关联， 并开发CDS工具来评估OUD风险。然后，我们将评估可用性、可接受性和可行性 在临床环境中使用CDS工具。这项研究将帮助医生快速准确地评估 OUD风险，早期干预，改善护理。我们的研究目标包括：目标1。使用NLP识别SDoH 临床记录和检查SDoH和OUD之间的关联。我们将使用Compendium和NLP 在临床记录中提取新的SDoH。两名评核人将手动验证新的SDoH，并使用描述性 统计来表征SDoH和OUD之间的关联。（培训目标1和2）。目标2：制定 CDS工具用于评估OUD风险。我们将使用目标1中的SDoH和OUD关联来开发一个有监督的 CDS工具的机器学习算法我们将通过测量CDS工具正确 评估患者EHR数据中的OUD风险（培训目标1和2）。目标3：测试可用性、可接受性和 医生使用CDS工具的可行性。将要求40名医生评估样本患者病例，然后 给出了CDS在这些案件上的结果医生将表明他们是否会遵循CDS的 建议.此外，参与者将被要求完成访谈和问卷调查， 评估可用性和可接受性。我们将通过审查招聘、实施和 元数据.（培训目标3）。这些目标是可以实现的，因为我在NLP和机器方面有经验。 学习和我的导师都是OUD研究，SDoH研究和干预设计方面的专家;并且有一个 在职业发展方面有出色的记录。这个K 01将帮助我实现研究独立， 提供a）开发OUD风险评估干预的技能; B）新兴SDoH领域的专业知识; c）创新的试验就绪可扩展干预;以及d）R 01的初步数据。

PROJECT SUMMARY/ABSTRACT In 2017, 1.7 million Americans suffered from opioid use disorders (OUD), which led to 47,000 American deaths from opioid overdose. Social determinates of health (SDoH) affect patients' OUD risk level and physicians' opioid prescribing. Physicians lack the tools to quickly and accurately assess SDoH associated with OUD, and lack knowledge of relevant resource for intervention. Clinical decision support (CDS) could quickly assess a patients' SDoH factors associated with OUD risk and provide actionable recommendations, which would reduce OUD risk assessment time and address knowledge gaps. In 2018, UCSF researchers created the Compendium of Medical Terminology Codes for Social Risk Factors that maps SDoH risks to medical vocabularies. However, most SDoH are documented in clinical notes. My long-term career goal is research independence with expertise in: 1) OUD risk assessment, 2) SDoH research, and 3) intervention development, implementation, and evaluation. Related to these goals, this study will use natural language processing (NLP) to identify SDoH in clinical notes, examine associations between SDoH and OUD, and develop a CDS tool to assess OUD risk. We will then assess usability, acceptability, and feasibility of using the CDS tool in clinical settings. This research will help physicians quickly and accurately assess OUD risk, intervene earlier, and improve care. Our research aims include: Aim 1. Use NLP to identify SDoH in clinical notes and examine associations between SDoH and OUD. We will use the Compendium and NLP to extract new SDoH in clinical notes. Two raters will manually validate the new SDoH, and use descriptive statistics to characterize associations between SDoH and OUD. (training goals 1 and 2). Aim 2: Develop a CDS tool to assess OUD risk. We will use SDoH and OUD associations from aim 1 to develop a supervised machine learning algorithm for our CDS tool. We will validate the CDS tool by measuring its ability to correctly assess OUD risk in patients' EHR data (training goals 1 and 2). Aim 3: Test the usability, acceptability, and feasibility of physicians' use of the CDS tool. 40 physicians will be asked to assess sample patient cases, then given CDS results on those same cases. Physicians will indicate whether they would follow the CDS's recommendations. Additionally, participants will be asked to complete an interview and questionnaire to evaluate usability and acceptability. We will assess feasibility by examining recruitment, implementation, and metadata. (training goal 3). These aims are achievable because I have experience in NLP and machine learning and my mentors are experts in OUD research, SDoH research, and intervention design; and have an outstanding record in career development. This K01 will help me achieve researcher independence by providing a) skills to develop an OUD risk assessment intervention; b) expertise in a novel growing SDoH field; c) an innovative trial-ready scalable intervention; and d) preliminary data for an R01.