Federated learning methods for heterogeneous and distributed Medicaid data

异构分布式医疗补助数据的联邦学习方法

• 批准号: 10590354
• 负责人: Lu Tang
• 金额: $ 18.93万
• 依托单位: UNIVERSITY OF PITTSBURGH AT PITTSBURGH
• 依托单位国家: 美国
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-04-01 至 2025-03-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10590354
• 关键词: Address; Big Data; Buprenorphine; Cells; Characteristics; Clinical; Collection; Complex; Computer software; Confidence Intervals; Data; Data Sources; Decision Making; Dedications; Dose; Ensure; Environment; Formulation; Goals; Health; Health Policy; Heterogeneity; Hospitals; Individual; Learning; Left; Low income; Machine Learning; Medicaid; Meta-Analysis; Methodology; Methods; Modeling; Opioid; Outcome; Overdose; Overdose reduction; Participant; Pennsylvania; Pharmaceutical Preparations; Policies; Population; Population Heterogeneity; Population Research; Precision Health; Recommendation; Research; Research Personnel; Safety; Sampling; Scheme; Site; Source; Subgroup; System; Testing; Translational Research; US State; Uncertainty; Variant; Vulnerable Populations; buprenorphine treatment; data integration; demographics; distributed data; diverse data; effectiveness clinical trial; empowerment; federated learning; heterogenous data; improved; improved outcome; individualized medicine; information model; innovation; insight; learning strategy; medication for opioid use disorder; mortality; multiple data sources; novel strategies; opioid overdose; opioid use disorder; overdose death; overdose prevention; personalized approach; precision medicine; programs; randomized, clinical trials; repository; social health determinants; success; treatment choice; treatment effect; user friendly software; user-friendly

Project Summary The broad objective of this project is to develop federated learning approaches that can efficiently reduce uncertainty and improve generalizability when assessing treatment effects based on multiple data sources. The proposal is motivated by a study of the Medicaid Outcome Distributed Research Network (MODRN) of eleven states in assessing the quality and access of medications for opioid use disorder (OUD). The collection of Medicaid claims data accounts for 40% of the OUD population in the US and covers a wide array of treatment choices, making it an ideal data source for understanding subgroup-specific treatment effects and developing precision health strategies. We leverage this large-scale distributed research network (DRN) to investigate the heterogeneous treatment effect (HTE) of buprenorphine, an opioid-based medication, on overdose mortality. However, the extra source of heterogeneity across states due to variation in state policy environments, which is largely unobserved, has presented great challenge in the assessment of HTE. Existing approaches such as meta-analysis are inadequate and underpowered to address the translational research needs in understanding the complex interactions among treatments, clinical characteristics and social determinant of health, especially, under the heavy influence of unexplainable heterogeneity across states. A suite of novel approaches will be developed to address a wide range of analytical requests that support data-driven precision health research under the framework of federated learning, where states collaboratively build analytical models under the orchestration of a coordinating state without pooling individual-participant data. With a central goal of modeling for different levels of heterogeneity in DRNs, this project focuses on the following aims: 1. To develop and evaluate a high-precision HTE estimator for buprenorphine for Pennsylvania by incorporating modeling information from ten other states; 2. To develop and evaluate a generalizable treatment recommendation system that protects vulnerable populations and is robust to policy variation across states. The methods will be rigorously tested and delivered as user friendly statistical software. The proposed methods extend well beyond MODRN and easily find applications in other common DRNs, such as hospital data networks and mobile data networks.

项目摘要 这个项目的广泛目标是开发联合学习方法，以有效地减少 在评估基于多个数据源的治疗效果时，这种方法具有不确定性，并提高了泛化能力。这个 该提案的动机是对医疗补助结果分布式研究网络(MODRN)的一项研究，该网络由11人组成 在评估治疗阿片类药物使用障碍的药物的质量和获取方面。收藏品 医疗补助计划声称，数据覆盖了美国40%的肥胖人口，并涵盖了广泛的治疗 选择，使其成为了解亚组特定治疗效果和发展的理想数据源 精准的健康战略。我们利用这个大规模的分布式研究网络(DRN)来调查 阿片类药物丁丙诺啡对过量用药死亡率的异质性治疗作用。 然而，由于州政策环境的变化而导致的跨州异质性的额外来源，这 在很大程度上没有被观察到，这给HTE的评估带来了巨大的挑战。现有的方法，如 元分析不足以解决翻译研究在理解方面的需求 治疗、临床特征和健康的社会决定因素之间的复杂相互作用，特别是， 在无法解释的跨州异质性的严重影响下。一套新的方法将是 旨在满足支持数据驱动的精确健康研究的广泛分析要求 在联邦学习框架下，各州在 协调状态的编排，而无需汇集单个参与者的数据。以建模为中心目标 针对DRN中不同程度的异构性，本项目侧重于以下目标：1.开发和 结合模型评估宾夕法尼亚州丁丙诺啡的高精度hte估计器 来自其他十个州的信息；2.开发和评估可推广的治疗建议 保护弱势群体并对各州政策变化保持稳健的制度。方法将是 经过严格测试，并作为用户友好的统计软件提供。建议的方法远远超出了 MODRN，并轻松找到其他常见DRN中的应用，如医院数据网络和移动数据 网络。