Development of a Machine Learning Model for Liver Transplantation

项目概要/摘要 目前，有近13，000名患者等待肝脏移植，但只有三分之二的患者将获得 2018年，约有2，500名患者因医疗原因死亡或从等待名单中删除。 恶化可供移植的供体肝脏的短缺导致了边缘肝脏的使用 比一般的供体肝脏风险更高，但可以安全地移植到精心挑选的接受者身上。 这些包括老年供体（>70岁）、脂肪肝和通过心脏移植后捐赠获得的肝脏。 死亡由于其风险性，这些边缘肝脏通常会下降，但多达84%的患者 在移植等待名单上死亡的人在死亡前拒绝了一个或多个边缘肝脏。鉴于此，某些 等待名单上的候选人可能从接受移植中获得生存益处， 器官，而不是留在等待名单上（即他们会在移植后存活更长时间， 边缘肝比他们会在等待名单上生存）。 目前，关于特定边缘肝脏是否适合特定候选人的决定是基于 临床完形或简单的亚组分析，使用传统的回归模型，这可能不完全 近似供体、受体和移植因素之间的复杂相互作用。为了解决这个问题，我们 将利用机器学习（可以包含复杂的高阶交互）来预测 特定候选者将从接受具有特定边缘肝脏的移植中获得存活益处， 并采访移植候选人和外科医生，以了解如何最好地将这些预测转化为 立即临床有用的决策援助。 为了实现这一目标，我们将利用国家移植登记处（SRTR）的数据， （n= 293，140），并使用机器技术（随机森林）来解决以下目标：（1）预测 等待名单上的肝移植候选人的等待名单生存率;（2）预测肝移植后的生存率 边缘肝脏的移植受者;以及（3）创建一个决策辅助工具， 存活率和具有边缘肝的特定移植候选者的预测移植后存活率。这些 鉴于我们小组在肝移植方面的专业知识，对国家登记数据的分析， 和机器学习技术。 我们假设，利用SRTR和机器学习，我们可以准确地预测移植后的生存率。 对于具有特定边缘肝脏的特定候选人，以及对于相同候选人的等待名单存活率， 没有肝脏。我们还假设我们的决策辅助可以实时用于通知临床 决策的如果实现了所提出的目标，我们的决策辅助可以用于改善临床 通过将高质量的风险预测直接带给患者和移植专业人员， 通知实时临床决定候选人是否应该接受边缘肝脏移植。

PROJECT SUMMARY/ABSTRACT Currently, there are nearly 13,000 patients waitlisted for liver transplant, yet only two-thirds will receive a transplant, and in 2018 approximately 2,500 patients died or were removed from the waitlist due to medical deterioration. This shortage of donor livers available for transplant has led to the use of marginal livers – livers that are higher risk than typical donor livers but may be safely transplantable in carefully selected recipients. These include older donors (>70 years), steatotic livers, and livers procured through donation after cardiac death. Due to their riskiness, these marginal livers are often declined, yet as many as 84% of patients who died on the transplant waitlist declined one or more marginal livers prior to death. In light of this, certain waitlisted candidates might have derived a survival benefit from undergoing transplantation with a marginal organ rather than remaining on the waitlist (i.e. they would have survived longer after a transplant with a marginal liver than they would have survived on the waitlist). Currently, decisions about whether a particular marginal liver is suitable for a particular candidate are based on clinical gestalt or simple subgroup analysis using traditional regression models, which likely do not fully approximate the complex interactions between donor, recipient, and transplant factors. To account for this, we will utilize machine learning (which can incorporate complex, higher-order interactions) to predict whether a specific candidate would derive a survival benefit from undergoing transplantation with a specific marginal liver, and interview transplant candidates and surgeons to understand how best to translate these predictions into an immediately clinically-useful decision aid. To accomplish this, we will leverage Scientific Registry of Transplant Recipient (SRTR) national data (n=293,140) and use a machine technique (random forests) to address the following aims: (1) To predict waitlist survival for waitlisted liver transplant candidates; (2) To predict post-transplant survival for liver transplant recipients of a marginal liver; and (3) To create a decision aid that compares predicted waitlist survival and predicted post-transplant survival for a specific transplant candidates with marginal liver. These aims are highly feasible given our group’s expertise in liver transplantation, analysis of national registry data, and machine learning techniques. We hypothesize that utilizing SRTR and machine learning, we can accurately predict post-transplant survival for a particular candidate with a particular marginal liver, as well as waitlist survival for that same candidate without a liver. We also hypothesize that our decision aid could be utilized in real-time to inform clinical decision-making. If the proposed aims are achieved, our decision aid could be utilized to improve clinical practice by bringing high-quality risk prediction directly to patients and transplant professionals to directly inform the real-time clinical decision of whether a candidate should undergo transplant with a marginal liver.