Using System Dynamics Modeling to Foster Real-time Connections to Care

项目摘要 我们目前资助的HD2A创新项目的两个目标是：(1)实施新的、可扩展的、 基于证据的干预(例如，我们的远程医疗平台RecoveryPad)，将过量服药的人联系起来 有了治疗阿片使用障碍(MOUD)的药物、减少伤害的服务和康复支持， 以及(2)收集与部署本系统相关的过程和结果的高质量数据 该平台可以与我们现有的系统动力学(SD)模型集成，以确定 以及未来应该实施哪些干预措施。以这种方式，我们的数据(即输入和输出 来自SD模型)在持续反馈中推动我们的行动(即，提供和改进RecoveryPad 循环播放。这一行政补充将允许将伦理人工智能(AI)框架整合到 对母项目的远程健康干预进行细化和评估。具体地说，我们将评估 父项目的数据集、模型假设、算法输入、开发和性能 潜在的偏见，特别是在加剧弱势群体与联合国系统有关的成果差距方面 人口。通过系统地检测和缓解算法偏差，我们将增强 人工智能强化干预的公平性，促进不同领域的公平治疗参与 人口统计数据。我们获得的见解不仅将优化我们自己的RecoveryPad平台和系统 动力学模型，但也将有助于更广泛的伦理人工智能应用于医疗保健。此外，我们的工作 致力于改善有OUD的个人的结果，并支持各国解决阿片类药物危机的努力。 具体地说，我们提出以下补充目标：1)目标1-评估部门内部的偏见和公平性 模型：这个目标寻求将人工智能公平评估方法转化为对 现有的系统动力学模型。通过利用我们的集成团队，其中包括生物伦理学家、人工智能专家、数据 科学家、临床医生和有实际经验的人，我们将检查关键的模型输入及其潜力 对敏感人口统计属性的模型输出的偏向影响。此外，我们打算确保 表示，并减轻任何算法偏差。2)目标2-评估以下方面的偏见和总体公平 RecoveryPad：目标2a)数据集的公正性评估和期间的简短协商面谈(BNI)过程 RecoveryPad开发：我们将通过分析我们的人口水平 机器学习算法显示了对不同组的Moud参与度的不同预测 历史电子健康记录数据，其中ED患者从亲自健康的人那里收到了BNI 推广倡导者。目标2b)RecoveryPad的偏差和公平性评估：我们将评估偏差和 通过模拟和实时参与者对话会议实现RecoveryPad内的公平性，利用 现有的框架用于评估对话人工智能的偏见和毒性。

Project Summary The two objectives of our currently funded HD2A Innovation Project are: (1) to implement a novel, scalable, evidence-based, intervention (i.e., our telehealth platform RecoveryPad) that links people who have overdosed with access to medication for opioid use disorder (MOUD), harm reduction services, and recovery supports, and (2) to collect high-quality data about the processes and outcomes associated with deployment of this platform that can be integrated with our existing system dynamics (SD) model to determine if, where, when, and what interventions should be implemented in the future. In this manner, our data (i.e., input and output from the SD model) drives our action (i.e., provision and refinement of RecoveryPad) in a continuous feedback loop. This administrative supplement will allow integration of an ethical artificial intelligence (AI) framework into the refinement and evaluation of the telehealth intervention of the parent project. Specifically, we will evaluate datasets, model assumptions, algorithmic inputs, development, and performance of the parent project for potential biases, particularly in relation to exacerbating disparity of OUD-related outcomes among vulnerable populations. Through the systematic detection and mitigation of algorithmic biases, we will enhance the fairness of AI-augmented interventions, promoting equitable treatment engagement across diverse demographics. The insights we gain will not only optimize our own RecoveryPad platform and system dynamics model but will also contribute to wider ethical AI applications in healthcare. Moreover, our work stands to improve outcomes for individuals with OUD and support national efforts to address the opioid crisis. Specifically, we propose the following supplemental aims: 1) Aim 1 - to assess bias and fairness within the SD model: This aim seeks to translate AI fairness assessment methodologies into iterative refinement of the existing system dynamics model. By leveraging our integrated team that includes a bioethicist, AI experts, data scientists, clinicians, and people with lived experience, we will examine key model inputs and their potential bias implications on model outputs for sensitive demographic attributes. Furthermore, we intend to ensure representation and mitigate any algorithmic bias. 2) Aim 2 - to assess bias and overall fairness of RecoveryPad: Aim 2a) Fairness evaluation of datasets and the brief negotiated interview (BNI) process during RecoveryPad Development: We will assess potential biases by analyzing whether our population-level machine learning algorithms exhibit differential predictions for MOUD engagement across diverse groups using historical electronic health record data, where ED patients have received a BNI from an in-person health promotion advocate. Aim 2b) Bias and Fairness Assessment of RecoveryPad: We will evaluate bias and fairness within RecoveryPad through simulated and real-time participant conversational encounters, leveraging existing frameworks for assessing conversational AI for bias and toxicity.