Developing a Donor-Candidate Risk Prediction System to Optimize Lung Allocation and Transplant Outcomes

项目摘要/摘要 肺移植分配系统不是以循证策略为指导的，这种策略解释了复杂性 捐赠者和候选人特征的相互作用错失了最大化生存收益的机会 利用严重有限的器官供应。为了克服这一缺陷，我们将开发一个捐赠者候选人风险 基于传统回归统计技术和现代机器学习的预测系统 人工学习技术侧重于揭示捐赠者特征、术后变化的影响。 移植存活，以及供者和候选人的相互作用。这一目标将通过实施 遵循三个目标。在目标1中，我们将检验这样一个假设，即结合捐赠者的特征可以改善 受体移植后存活率预测模型的准确性。我们将使用基于回归和机器的 学习方法，并比较结果生存模型的准确性。在目标2中，我们将确定如何 供者和候选者的特征相互作用，在移植后的存活率中引入差异。基于回归的 并将使用机器学习方法来识别和评估交互、集群和效果 根据等待名单时间、疾病严重程度和功能状态进行修改。在目标3中，我们将开发一个机器学习/ 人工智能算法，用于通知器官分配和接受决定。生存权衡将是 特征是利用机器学习模型构建的人工智能分配算法 与历史上的决定相比。总而言之，当前的美国肺分配系统尚未考虑 供者因素对移植后风险预测的贡献，这可能解释为什么LAS-派生的估计 生存福利是不准确的。改进的风险预测将允许优化捐赠者和候选人 匹配，为基于兼容性的系统奠定框架，具有改进供体的潜力 利用率、等待存活率和移植后存活率。使用分阶段建模策略结合传统的 基于回归的方法以及现代机器学习和人工智能方法将鼓励 美国肺分配问题的创新解决方案。这项提议的创新之处进一步增加了一个 独一无二的合格的多学科研究团队，具有复杂系统分析和美国肺的专业知识 分配政策。

Project Summary/Abstract The lung transplant allocation system is not guided by an evidence-based strategy that accounts for the complex interactions of donor and candidate characteristics missing an opportunity to maximize survival benefit from utilization of the severely limited organ supply. To overcome this deficit, we will develop a donor-candidate risk prediction system guided by traditional regression-based statistical techniques and modern machine learning and artificial learning techniques focused on uncovering the impact of donor characteristics, variation in post- transplant survival, and donor and candidate interactions. This goal will be accomplished by carrying out the following three aims. In Aim 1, we will test the hypothesis that incorporating donor characteristics improves accuracy of prognostic models of recipient post-transplant survival. We will use regression-based and machine learning approaches and compare the accuracy of the resultant survival models. In Aim 2, we will determine how donor and candidate characteristics interact to introduce variation in post-transplant survival. Regression-based and machine learning approaches will be used to identify and evaluate interactions, clustering, and effect modification by waitlist time, illness severity, and functional status. In Aim 3, we will develop a machine learning/ artificial intelligence algorithm to inform organ allocation and acceptance decisions. Survival trade-offs will be characterized using machine learning models to build an artificial intelligence allocation algorithm which will be compared to historical decisions. In summary, the current US lung allocation system does not yet consider the contribution of donor factors to post-transplant risk predictions which may explain why LAS-derived estimates of survival benefit are inaccurate. Improved risk predictions would permit optimization of donor and candidate matching to lay the framework for a system based on compatibility which has the potential to improve donor utilization, waitlist survival, and post-transplant survival. Use of a staged modeling strategy combining traditional regression-based approaches and modern machine learning and artificial intelligence methods will encourage innovative solutions to problems in US lung allocation. This proposal's innovation is further augmented by a uniquely qualified multi-disciplinary research team with expertise in analysis of complex systems and US lung allocation policies.