Mending a Broken Heart Allocation System with Machine Learning

点击翻译按钮获取中文摘要

PROJECT ABSTRACT Heart transplantation is a life-saving treatment for end-stage heart failure, a devastating disease which kills over 250,000 Americans each year. Unfortunately, the supply of deceased donor hearts cannot meet demand, and over a third of candidates will die or be delisted without transplant. In the context of such scarcity, allocation must make the best use of scarce deceased donor hearts by ranking candidates from most to least medically urgent. In contrast to other organ transplant systems, there is currently no objective score used to rank heart transplant candidates on the waitlist. Instead, each candidate’s priority for transplantation is based on “Status,” a designation determined by the supportive therapy prescribed by their transplant center. I have previously shown that some heart transplant centers appear to overtreat candidates with intensive therapies at far higher rates than other centers. My preliminary data demonstrates that these practices have consequences for heart allocation effectiveness. High survival benefit centers reserve intense supportive therapy for candidates who have poor prognoses without transplant, saving lives by prioritizing the sickest patients. In contrast, low survival benefit centers list stable candidates and escalate the use of supportive therapies. Based on these data, there is a clear need for a new system to fairly allocate donor hearts. The overall objective of this K08 application is to develop and simulate a novel Heart Allocation Score (HAS) designed to objectively identify the candidates who gain the greatest survival benefit from heart transplantation. Previous attempts to develop such a score using conventional statistical methods have been inaccurate, but cutting-edge machine learning (ML) techniques outperform conventional regression models in many clinical contexts. In addition, a new open-source Heart Simulated Allocation Model (HSAM) is needed to compare policy alternatives because the available program is closed-source, inflexible, outdated, and structurally unable to simulate allocation scores developed with ML. My overall hypothesis is that a HAS developed with ML will lead to policy that optimizes heart allocation. I will test this hypothesis in three Aims. In Aim 1, I will use the complete national transplant registry dataset (N = 109,315 adult candidates) to predict waitlist survival, comparing ML prediction models to the current therapy-based system. In Aim 2, I will use the same registry to predict post-transplant survival for heart recipients, comparing conventional statistical methods to ML. In Aim 3, I will develop a) a new, open-source HSAM which I will use to b) compare current policy to a novel HAS policy constructed from the best prediction models from Aim 1 & 2. My overall career goal is to save lives by designing delivery systems that fairly and efficiently distribute scarce medical resources. To accomplish this, I plan to earn a PhD in Health Services Research focused on ML, simulation modeling, and health policy. Achieving the goals of this proposal will lead to the foundation of a novel heart allocation system that has the potential to save lives and equip me with the skills needed for future R01- level applications in the field of scarce healthcare resource allocation.