Developing and Evaluating a Machine-Learning Opioid Prediction & Risk-Stratification E-Platform (DEMONSTRATE)

开发和评估机器学习阿片类药物预测

• 批准号: 10442365
• 负责人: Wei-Hsuan Lo-Ciganic
• 金额: $ 65.46万
• 依托单位: UNIVERSITY OF FLORIDA
• 依托单位国家: 美国
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-07-01 至 2026-04-30
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10442365
• 关键词: Adopted; Algorithms; Ambulatory Care Facilities; American; Clinic; Clinical; Clinical Data; Clinical Research; Computer software; Criminal Justice; Data; Data Set; Data Sources; Development; Diagnosis; Dose; Electronic Health Record; Ensure; Feedback; Florida; Focus Groups; Funding; Guidelines; Health; Health Care Costs; Health system; Healthcare Systems; Homelessness; Individual; Intervention; Interview; Letters; Link; Machine Learning; Measures; Medicaid; Medicare; Methods; Naloxone; National Institute of Drug Abuse; Natural Language Processing; Nurse Practitioners; Opioid; Outcome; Overdose; Patients; Pennsylvania; Performance; Persons; Pharmacy facility; Physician Assistants; Physicians; Policy Maker; Primary Health Care; Process; Proctor framework; Productivity; Provider; Public Health; Reporting; Research; Resources; Risk; Safety; Social Behavior; System; Time; Translating; Translations; United States Centers for Medicare and Medicaid Services; Universities; Visit; Work; acceptability and feasibility; adverse outcome; base; care providers; clinical decision support; clinical practice; cohort; cost; deep neural network; design; high risk; implementation outcomes; improved; innovation; insurance claims; machine learning algorithm; machine learning prediction algorithm; model development; neural network; neural network algorithm; novel strategies; opioid overdose; opioid use; opioid use disorder; overdose risk; pilot test; post implementation; prediction algorithm; predictive modeling; prescription opioid; prescription opioid misuse; prevent; programs; prototype; recurrent neural network; response; risk mitigation; risk prediction; risk stratification; structured data; success; support tools; tool; usability; user centered design; user-friendly

Project Summary/Abstract An unprecedented rise in opioid overdose and opioid use disorder (OUD) has become a public health crisis in the US. In response, health systems, payers, and policy makers have developed or adopted measures and programs to target individuals at high-risk for overdose or OUD. However, significant gaps exist in the current approaches to identify individuals at high-risk for overdose or OUD. First, the definition of ‘high-risk’ currently used by payers and health systems varies widely (ranging from high opioid dose to the number of pharmacies or prescribers a patient has visited). Second, little is known about how accurately these measures truly identify patients with overdose or OUD, and there is some evidence showing they perform poorly, missing 70% to 90% of individuals with an actual OUD diagnosis or overdose. Third, our NIDA-funded work (R01DA044985) using national Medicare and Pennsylvania Medicaid claims data has shown that machine-learning algorithms can achieve better performance for risk prediction for opioid overdose and OUD. Thus, the immediate next step is to expand our algorithms to other data sources (e.g., electronic health records [EHR]), as well as to apply state-of- the-art longitudinal neural networks and natural language processing (NLP) to further improve prediction accuracy. In addition, we aim to translate these risk scores into a clinical decision tool to be used by health care systems to automatically analyze and visualize the relevant information regarding risk prediction and stratification for opioid overdose or OUD, using either claims data, EHR data, or both in real time. Leveraging our NIDA-funded work on developing machine-learning algorithms to predict opioid overdose and OUD, we propose to “develop and evaluate a machine-learning opioid prediction & risk-stratification e- platform (DEMONSTRATE)” that can be used by health care systems to identify patients at high risk for opioid overdose and OUD. We have 3 specific aims. Aim 1 will refine and validate prediction algorithms to identify patients at risk for opioid overdose/OUD using 3 different datasets (i.e., 2011-2020 Florida all-payer EHR, Florida Medicaid claims, and Florida Medicaid claims linked with EHR data) from the OneFlorida Clinical Research Consortium. We will expand our current algorithms by applying state-of-the-art methods (e.g., NLP) to improve prediction. In Aim 2, we will design and prototype a DEMONSTRATE clinical decision support tool to incorporate the best prediction algorithms to provide automatic warnings to primary care providers of patients at high risk of overdose/OUD. An iterative user-centered design approach will be used to enhance DEMONSTRATE’s functionality and usability. In Aim 3, we will integrate DEMONSTRATE into the University of Florida Health’s EHR system, and deploy and pilot test DEMONSTRATE in three primary care clinics. We will assess DEMONSTRATE’s usability, acceptability, and feasibility. Our proposed research is highly innovative in its expansion, translation, and application of a promising NIDA-funded machine-learning opioid prediction and risk stratification tool into a software platform to better inform clinical practice for improving safety of opioid use.

项目总结/摘要 阿片类药物过量和阿片类药物使用障碍（OUD）的空前增加已成为一场公共卫生危机， 美方作为回应，卫生系统、付款人和决策者制定或采取了措施， 计划针对高风险的个人过量或OUD。然而，目前存在重大差距， 识别过量或OUD高风险个体的方法。第一，目前“高风险”的定义 支付者和卫生系统使用的药物种类差异很大（从高阿片类药物剂量到药店数量 或者是病人曾经拜访过的开处方者）。其次，人们对这些措施的准确性知之甚少， 过量或OUD患者，有证据表明他们表现不佳，缺失70%至90% 被诊断为OUD或用药过量的人。第三，我们的NIDA资助的工作（R 01 DA 044985）使用 国家医疗保险和宾夕法尼亚州医疗补助索赔数据表明，机器学习算法可以 在阿片类药物过量和OUD的风险预测方面取得更好的表现。因此，紧接着的下一步是 将我们的算法扩展到其他数据源（例如，电子健康记录（EHR）），以及适用于 最先进的纵向神经网络和自然语言处理（NLP），以进一步改善预测 精度此外，我们的目标是将这些风险评分转化为医疗保健使用的临床决策工具 自动分析和可视化有关风险预测和分层的相关信息的系统 对于阿片类药物过量或OUD，使用真实的索赔数据、EHR数据或两者。 利用我们由NIDA资助的开发机器学习算法的工作来预测阿片类药物过量， OUD，我们建议“开发和评估机器学习阿片类药物预测和风险分层电子商务， 平台（演示）”，可用于医疗保健系统，以确定患者的高风险， 阿片类药物过量和OUD我们有三个具体目标。目标1将完善和验证预测算法， 使用3个不同的数据集识别存在阿片类药物过量/OUD风险的患者（即，2011-2020年佛罗里达所有付款人EHR， 佛罗里达州医疗补助索赔，以及与EHR数据相关的佛罗里达州医疗补助索赔） 研究联盟。我们将通过应用最先进的方法（例如，NLP） 改善预测。在目标2中，我们将设计和原型演示临床决策支持工具， 结合最佳预测算法，为患者的初级保健提供者提供自动警告， 过量/OUD的高风险。一个迭代的以用户为中心的设计方法将被用来提高 演示的功能和可用性。在目标3中，我们将把DEMONSTRATE纳入大学， 佛罗里达健康的电子健康记录系统，并部署和试点测试演示在三个初级保健诊所。我们将 评估DEMONSTRATE的可用性、可接受性和可行性。我们提出的研究是高度创新的， 它对NIDA资助的一项有前途的机器学习阿片类药物预测的扩展、翻译和应用， 将风险分层工具整合到软件平台中，以更好地为临床实践提供信息，从而提高阿片类药物使用的安全性。