Using natural language processing and machine learning to identify potentially preventable hospital admissions among outpatients with chronic lung diseases

项目摘要 患有慢性肺部疾病(CLD)的患者经常因潜在的可预防因素而入院治疗。 原因。这样的入院可能与患者的偏好不一致和/或代表着低价值的 卫生系统资源。为了预测这样的入院情况，本fi中现有的临床预测模型通常 提出一种“全因”风险评估，即使它是准确的，也忽略了ADMIS背后的可行机制-- Sion风险，因此未能确定规定的应对措施。这一限制可能只能解释-充其量是适度的 -在大多数已测试的干预捆绑包中看到的住院和再入院人数的减少 在这群人中。因此，有机会预测住院风险，同时识别 患者的表型(即一些社会、人口、临床和其他特征的星座) 已知的预防性干预措施是存在的。拟议的研究试图克服这些限制，并利用 利用这个机会(1)对住院的慢性腰椎病患者进行半结构化访谈，以及他们的 护理人员和临床医生，以直接识别莫迪fi的风险及其相关的表型驱动医院广告- 任务；(2)使用自然语言处理技术建立Classifi阳离子模型 临床病历非结构化文本中的细微差别叙述、社交和临床信息，以识别 具有这些表型的患者；以及(3)建立以可操作表型为重点的风险预测模型 广泛的传统回归和机器学习方法，同时也整合了大量 从文本数据中提取预测变量，并考虑时变趋势。候选人的前期工作 应用基础fi技术显著提高住院患者临床预测模型的识别率 人口推动了这种方法论的方法。与日俱增的住院负担和费用 CLD以及联邦支付者越来越多的关注，突显了这项工作的关键性质。完成 这项研究将建立在候选人过去的培训基础上，其中包括卫生政策科学硕士 在NHLBI T32支持下获得的研究，并将提供经验、教育和指导，以允许 成为一名完全独立的调查员的候选人。根据候选人量身定做的培训计划，他将 通过课程学习获得混合方法研究、自然语言处理和试验设计方面的高级技能，密切合作- 指导和监督，指导实践。这些技能将使他处于提交成功R01测试的理想状态 建议的临床预测模型在真实世界环境中的部署。候选人的主要导师， 合作者和顾问将确保遵守建议的时间表和目标，并提供支持- IVE环境为他发展了独立的研究生涯，测试了临床的真实部署 减少慢性腰椎间盘突出症患者低价值和偏好不一致护理的预测模型。

Project Summary Patients living with chronic lung diseases (CLDs) are frequently admitted to the hospital for potentially preventable causes. Such admissions may be discordant with patient preferences and/or represent a low-value allocation of health system resources. To anticipate such admissions, existing clinical prediction models in this field typically produce an “all-cause” risk estimate which, even if accurate, overlooks the actionable mechanisms behind admis- sion risk and therefore fails to identify a prescribed response. This limitation may explain the only modest – at best – reductions in hospital admissions and readmissions seen in most intervention bundles that have been tested in this population. An opportunity exists, therefore, to predict hospitalization risk while simultaneously identifying patient phenotypes (i.e. some constellation of social, demographic, clinical, and other characteristics) for which known preventive interventions exist. The proposed study seeks to overcome these limitations and capitalize on this opportunity by (1) conducting semi-structured interviews with hospitalized patients with CLDs, and their caregivers and clinicians, to directly identify modifiable risks and their associated phenotypes driving hospital ad- missions; (2) using natural language processing techniques (NLP) to build classification models that will leverage nuanced narrative, social, and clinical information in the unstructured text of clinical encounter notes to identify patients with these phenotypes; and (3) building risk prediction model focused on actionable phenotypes with a wide-array of traditional regression and machine learning approaches while also incorporating large numbers of predictor variables from text data and accounting for time-varying trends. The candidate's preliminary work using basic NLP techniques to significantly improve the discrimination of clinical prediction models in an inpatient population has motivated this methodologic approach. The rising burden and costs of hospitalizations associated with CLDs, and the increasing attention from federal payers, highlights the critical nature of this work. Completion of this research will build upon the candidate's past training, which includes a Masters of Science in Health Policy Research obtained with NHLBI T32 support, and will provide the experience, education, and mentorship to allow the candidate to become a fully independent investigator. Based on the candidate's tailored training plan, he will acquire advanced skills in mixed-methods research, NLP, and trial design all through coursework, close men- toring and supervision, and direct practice. The skills will position him ideally to submit successful R01s testing the deployment of the proposed clinical prediction models in real-world settings. The candidate's primary mentor, collaborators, and advisors will ensure adherence to the proposed timeline and goals and provide a support- ive environment for him to develop an independent research career testing the real-world deployment of clinical prediction models to reduce low-value and preference-discordant care for patients with CLDs.