Developing Machine Learning Models for Decision Support and Allocation Optimization in Heart Transplantation

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT The impact of heart transplantation (HTx) remains limited by donor shortages, with an estimated 250,000 adults who may benefit from transplant despite only 3,500 being performed each year in the US. Unfortunately, donor discard rates remain high at 70-80%, with substantial unexplained variability in donor evaluation and acceptance practices between centers. Recent data also demonstrate that higher risk recipients are being transplanted under the new 2018 allocation policy with worse post-transplant survival rates nationally. These trends collectively underscore current limitations in allocation policy and the ability for individual programs to assess donor quality and to pair suitable donors with appropriately selected recipients. The latter stems from a suboptimal process whereby clinicians have to make time-sensitive decisions relying solely upon experience and judgement without data-driven tools that can analyze numerous donor and recipient data and their complex interactions to provide rapid and accurate outcome projections. Existing risk models have failed to garner widespread utilization due to major limitations, including 1) narrow focus on only one of a set of relevant outcomes, 2) simplistic approach with only modest discriminatory capability (c-statistics <0.70), 3) failure to account for complex interactions between donor and recipient variables, and 4) use of only static, cross-sectional data. Our proposal seeks to advance the field by leveraging a novel, comprehensive dataset and machine learning (ML) to develop robust models that can maximize predictive performance for relevant outcomes and to better align a candidate's clinical trajectory and anticipated transplant outcome. These models will better account for complex interrelationships between donor and recipient variables, and will also account for dynamic changes in candidate and donor parameters. Optimized models will then be incorporated into a decision support system guided by key stakeholders. In addition, a previously developed artificial intelligence (AI) framework will be used to optimize heart allocation policy. We have these specific aims: 1) Establish the feasibility and usability of a stakeholder-guided, ML-derived decision support system for adult HTx; 2) Demonstrate the adaptability of a previously developed AI-based policy-optimization framework to heart allocation; and 3) Inform and evaluate the processes and outputs of Specific Aims 1 and 2 using stakeholder engagement and implementation science to refine and optimize working prototypes and promote the understanding, adoption, and use of data-driven decision support tools created for HTx. This work will optimize the allocation of scarce resources and ultimately improve outcomes of HTx.