Development of Machine Learning Composite Measures for Graft Outcome Selection in Pediatric Liver Transplantation

摘要 儿科肝移植(PLT)的结果不受捐赠器官供应的限制。孩子们快死了 等待器官，即使这些死亡是完全可以通过适当的器官选择来预防的。而不是 临终时，孩子们可以通过适当选择供移植的肝脏来度过完整而活跃的一生。至关重要的是 实现零等待死亡率和长期移植收益是及时干预的能力 有合适的器官和移植物类型。继续进行PLT的决定是复杂的，最终是基于 移植团队经验、临床评估和器官可用性的一致性。在一个器官短缺的时代， 技术变异(TV)移植物的使用，包括分离式肝移植和活体供体肝移植，已经 扩大移植物选择的可能性，并使手术干预更及时。大多数移植计划都有 既往的电视移植物经验具有较低的患者死亡率和良好的移植结果。然而，一些人 移植计划之前在电视移植方面的经验有限，报告了许多糟糕的结果 接受电视移植的患者。尽管总体结果有所改善，但国家登记数据 证实不同移植中心在等待死亡率、电视移植物使用和移植后方面存在显著差异 结果。与这一变化密切相关的是移植决策的复杂性。集体，捐赠者 而受贿接受、候选人的优先顺序和分配政策描绘了一个复杂的场景。多过 在一个捐赠者和受赠者“最佳匹配”的决定中可以考虑100个变量，这有主观性的风险 和不匹配，因为人类的局限性，不应该被低估。认识到这些限制， 包括机器学习和深度学习在内的人工智能分类器因其 有可能支持或确认移植领域的决策。尽管如此，总体上还是由数据驱动 缺乏对最佳嫁接选择和传播嫁接决策的支持。机会，以及 发现的影响是很大的。该项目将产生一个综合决策支持软件工具，它使用 机器学习使用移植前死亡率、移植后死亡率来预测和建模患者的最佳生存 移植结果和先前的中心经验。可以建立决策支持工具来补充 目前在PLT中的移植物选择实践。我们预计，基于复合度量的建模将 在电视移植物的接受者中展示同样的结果。我们将开发一种算法来优化儿科 将通过Starzl儿科卓越网络商业化使用的移植物类型选择 在进一步的多中心验证后，它将可用于所有的儿科移植计划。 我们将通过以下三个目标实现我们的目标。一、确定最优特征空间 患者和PLT移植物存活率的预测变量。第二，开发生存预测模型，“PSELECT” 保留在等待名单上或接受各种移植物类型。三、论证模拟技术可行性 以消除基于先前搁置数据的PSELECT性能的等待名单死亡率。

ABSTRACT Outcomes in pediatric liver transplantation (pLT) are not limited by the donated organ supply. Kids are dying waiting for organs even when these deaths are completely preventable through proper organ selection. Instead of dying, children can live a full and active lifetime with a properly selected liver graft for transplant. Critical to achieving zero waitlist mortality and long-term transplant benefit is the capacity to intervene in a timely manner with a suitable organ and graft type. Decisions to proceed with pLT are complicated, ultimately based on the alignment of transplant team experience, clinical assessment, and organ availability. In an era of organ shortages, the use of technical variant (TV) grafts, including split liver transplantation and living donor liver transplant, has the potential to expand graft choice and enable timelier surgical intervention. Most transplant programs that have prior experience with TV grafts have low patient mortality and excellent transplant outcomes. However, some transplant programs that have limited prior experience with TV grafts have reported many poor outcomes for patients receiving TV transplants. Despite improvements in overall outcomes, national registry data have confirmed significant variation among transplant centers in waitlist mortality, TV graft use, and post-transplant outcomes. Integrally linked to this variation is the intricacy of transplant decision making. Collectively, donor and graft acceptance, prioritization of candidates, and allocation policies depict a complex scenario. More than 100 variables can be considered in a single donor-recipient ‘‘best matching’’ decision, with a risk of subjectivity and mismatch because of human limitations that should not be underestimated. Recognizing these limitations, artificial intelligence classifiers, including machine learning and deep learning, have been recognized for their potential to support or confirm decision making within the field of transplantation. Still, overall data-driven support for optimal graft selection and dissemination of graft decision support is lacking. Opportunities for, and the impact of, discovery are high. This project will result in a composite decision support software tool that uses machine learning to predict and model the best survival for the patient using pre-transplant mortality, post- transplant outcomes, and prior center experience. The decision support tool can be established to supplement current graft selection practices in pLT. We anticipate that modeling based on composite measures will demonstrate equivalent outcomes in recipients of TV grafts. We will develop an algorithm for optimal pediatric graft-type selection that will be commercialized for use through the Starzl Network for Excellence in Pediatric Transplantation and after further multi-center validation it will be available for all pediatric transplant programs. We will accomplish our objective through the following three aims. One, determine the optimal feature space for predictive variables for patient and pLT graft survival. Two, develop survival prediction models, “PSELECT,” for remaining on the waitlist or receiving various graft types. Three, demonstrate the simulated technical feasibility to eliminate the waitlist mortality based on the PSELECT performance on previously held-out data.