Predicting post-transplant kidney function from pre-transplant biopsy and clinical biomarkers using a convolutional neural network

项目总结/摘要 目前，对可移植肾脏的需求远远超过了其可用性， 已列入候补名单，但每年进行的移植手术不到22，000例。超过8,000名患者被从 由于死亡或健康状况恶化而等待录取。尽管这一迫切需要， 器官，近20%的回收肾脏最终被丢弃。最常报告的原因是 丢弃供体活检的组织学结果不佳。进行这些移植前活检是为了 评估器官的质量，并且通常用作预测植入后同种异体移植物性能的工具。 不幸的是，活检结果的预后意义是有争议的，并有越来越多的关注 关于来自活检解释的数据的可靠性和可重复性， 可变性最近的证据表明，受体移植物的结果只与供体活检相关 由经验丰富的肾脏病理学家进行解读。然而，大多数移植中心没有更多的 而不是少数专门的肾脏病理学家;考虑到器官恢复往往发生在远程 在深夜或周末的医院，活检通常由随叫随到的病理学家解释， 专门的肾脏组织学培训。这些提供者倾向于高估慢性病变的严重性， 导致不适当地丢弃原本可接受的器官。 卷积神经网络（CNN）是一种机器学习技术，可以等同或超过人类 以自动化、客观的方式执行视觉分析任务。我们建议利用这一新的 技术来实现以下目标：（1）开发一种可靠准确地预测后 数字化采购活检载玻片的移植物功能以及供体和受体指标， 移植受体科学登记（SRTR）数据集;（2）比较我们的预测准确性 CNN到目前可用的供体风险评分;和（3）定性评估CNN的可采用性，可接受性， 和实用性。考虑到我们团队在机器学习、肾脏研究和生物医学方面的专业知识， 移植和SRTR数据分析。 我们假设我们可以建立一个CNN，为移植医生提供准确的术前真实的- 移植后移植成功的时间估计，以帮助指导患者咨询。如果目标是 实现，从我们的CNN反馈可以防止不适当的丢弃数以千计的肾脏， 通过增加在全国范围内进行的移植手术数量来降低等待名单上的死亡率。通过 进行这项研究，Dr. Malleson将培养一套技能，包括国家登记数据分析， 定性方法和机器学习：正在迅速使用的重要现代技术 在整个医学领域，并将在她作为一名独立的外科医生-科学家的职业生涯中为她提供良好的服务。

PROJECT SUMMARY/ABSTRACT The need for transplantable kidneys far exceeds their availability with over 90,000 candidates currently waitlisted but less than 22,000 transplants performed annually. Over 8,000 patients are removed from the waitlist each year due to death or deterioration in health while awaiting an offer. Despite this critical need for organs, nearly 20% of recovered kidneys are ultimately discarded. The most commonly reported reason for discard is unfavorable histology on donor biopsy. These pre-transplant biopsies are performed in order to assess the quality of the organ and are often used as a tool to predict post-implantation allograft performance. Unfortunately, the prognostic significance of biopsy findings is controversial and there is growing concern regarding the reliability and reproducibility of data derived from biopsy interpretation due to inter-pathologist variability. Recent evidence demonstrates that recipient graft outcomes correlate only with donor biopsy interpretation performed by an experienced renal pathologist. However, most transplant centers have no more than a handful of dedicated expert renal pathologists; given that organ recovery often occurs at remote hospitals late at night or on weekends, biopsies are usually interpreted by on-call pathologists without dedicated training in renal histology. These providers tend to overestimate the severity of chronic lesions, resulting in the inappropriate discard of otherwise acceptable organs. Convolutional neural networks (CNNs), a machine learning technique, can equal or exceed human performance in visual analysis tasks in an automated, objective fashion. We propose to leverage this new technology to accomplish the following aims: (1) To develop a CNN that reliably and accurately predicts post- transplant graft function from digitized procurement biopsy slides and donor and recipient metrics in the Scientific Registry of Transplant Recipients (SRTR) dataset; (2) To compare the predictive accuracy of our CNN to currently available donor risk scores; and (3) To qualitatively evaluate CNN adoptability, acceptability, and utility by clinicians. These aims are highly feasible given our group's expertise in machine learning, kidney transplantation, and analysis of SRTR data. We hypothesize that we can build a CNN that provides transplant physicians with accurate pre-operative real- time estimates of post-transplant graft success to help guide patient counseling. If the proposed aims are achieved, feedback from our CNN could prevent the inappropriate discard of thousands of kidneys and decrease waitlist mortality by increasing the number of transplants performed across the country. By conducting this research, Dr. Eagleson will cultivate a skillset that includes national registry data analysis, qualitative methods, and machine learning: important modern techniques that are rapidly becoming used throughout medicine and will serve her well throughout her career as an independent surgeon-scientist.