National NLP Clinical Challenges (n2c2): Challenges in Natural Language Processing for Clinical Narratives

项目概要和摘要 电子健康记录（EHR）的叙述包含难以自动提取的有用信息， 索引、搜索或解释。自然语言处理（NLP）技术可以提取这些信息， 将其转换为计算机系统更容易访问的结构化格式。但 自然语言处理系统的发展取决于对相关数据的访问，而电子病历是出了名的难以获得 因为隐私的原因。尽管最近努力去识别和发布叙事EHR用于研究， 这些数据仍然非常稀少。因此，临床NLP作为一个领域已经落后。为了解决这个问题， 自2006年以来，我们组织了13项共同任务，并举办了讲习班和出版期刊。十二 这些共同的任务集中在临床NLP系统的开发和其余的一个在 这些系统的可用性。我们已经在共享任务、准备任务方面涵盖了深度和广度 它研究来自多个机构的各种EHR数据的尖端NLP问题。我们的共同任务是 运行时间最长的一系列临床NLP共享任务，不断增长的EHR数据集，任务和参与。 我们最流行的三个数据集被引用了495次（2010年数据）、284次（2006年de-id数据）和274次（2009年数据） 次，分别代表了数百篇文章，这些文章来自这三个数据集。我们 本提案的目标是继续我们在2006年开始的i2 b2共享任务挑战（i2 b2，NIH NLM U 54 LM 008748，PI：Kohane和R13 LM 011411，PI：Uzuner）去识别EHR，用金色注释它们- 临床NLP任务的标准注释，并将其发布给研究社区进行开发 和临床NLP系统的头对头比较，以促进最新技术的发展。 我们在国家自然语言处理临床挑战（n2 c2）的基础上，在健康数据科学计划的努力 作为哈佛医学院新成立的生物医学信息学系，我们的目标是建立合作伙伴关系 与社区一起，以几种方式增加共享任务的工作量：（1）增加可用的去识别化EHR数据 通过伙伴关系，可以有助于增加数据的数量和种类，以及（2）增加可用的 在NLP任务的深度和广度方面的黄金标准注释。鉴于这些目标和伙伴关系，我们 计划举行一系列的共同任务。我们将通过举办研讨会来补充这些共同任务， 与美国医学信息学协会秋季研讨会和杂志特别 这些问题的解决，可以加快最先进技术的进步，子孙后代可以在过去的基础上再接再厉。

Project Summary and Abstract Narratives of electronic health records (EHRs) contain useful information that is difficult to automatically extract, index, search, or interpret. Natural language processing (NLP) technologies can extract this information and convert it in to a structured format that is more readily accessible by computerized systems. However, the development of NLP systems is contingent on access to relevant data and EHRs are notoriously difficult to obtain because of privacy reasons. Despite the recent efforts to de-identify and release narrative EHRs for research, these data are still very rare. As a result, clinical NLP, as a field has lagged behind. To address this problem, since 2006, we organized thirteen shared tasks, accompanied with workshops and journal publications. Twelve of these shared tasks have focused on the development of clinical NLP systems and the remaining one on the usability of these systems. We have covered both depth and breadth in terms of shared tasks, preparing tasks that study cutting-edge NLP problems on a variety of EHR data from multiple institutions. Our shared tasks are the longest running series of clinical NLP shared tasks, with ever growing EHR data sets, tasks, and participation. Our most popular three data sets have been cited 495 (2010 data), 284 (2006 de-id data), and 274 (2009 data) times, respectively, representing hundreds of articles that have come out of these three data sets alone. Our goal in this proposal is to continue the efforts we started in 2006 under i2b2 shared task challenges (i2b2, NIH NLM U54LM008748, PI: Kohane and R13 LM011411, PI: Uzuner) to de-identify EHRs, annotate them with gold- standard annotations for clinical NLP tasks, and release them to the research community for the development and head-to-head comparison of clinical NLP systems, for the advancement of the state of the art. Continuing our efforts under National NLP Clinical Challenges (n2c2) based at the Health Data Science program of the newly established Department of Biomedical Informatics at Harvard Medical School, we aim to form partnerships with the community to grow the shared task efforts in several ways: (1) grow the available de-identified EHR data sets through partnerships that can contribute to the volume and variety of the data, and (2) grow the available gold-standard annotations in terms of depth and breadth of NLP tasks. Given these aims and partnerships, we plan to hold a series of shared tasks. We will complement these shared tasks with workshops that meet in conjunction with the Fall Symposium of the American Medical Informatics Association and with journal special issues so that advancement of the state of the art can be sped up and future generations can build on the past.