Semi-Automated Abstract Screening for Comparative Effectiveness Reviews

描述(由申请者提供)：在这个为期三年的项目中，我们的目标是应用最先进的信息分析技术来帮助产生系统评价和荟萃分析，这些评价和荟萃分析越来越多地被用作循证医学(EBM)和比较有效性评价的基础。我们计划开发一种人工引导的计算机化摘要筛选工具，以极大地减少执行繁琐但关键的步骤的需要，即手动筛选由文献搜索产生的数千篇摘要，以检索可能与进一步分析相关的一小部分。该工具将把经过验证的机器学习技术与一个新的开放源码工具结合起来，以实现对筛选过程的管理。与目前的人工过程相比，这项新技术将使调查人员能够在一小部分时间内筛选摘要。它将减少产生系统审查的时间和成本，提供该过程的明确文件，并有可能更准确地执行任务。随着循证医学的接受和对系统审查的需求日益增加，迫切需要工具来协助产生新的系统审查并对其进行更新。这一需求再紧迫不过了。最近通过了《美国复苏和再投资法案》，并拨出11亿美元用于比较有效性研究，这对系统审查产生了前所未有的需求，并创造了机会，以改进其开展工作的方法和效率。 在此，我们建议开发新的、开源的软件，以帮助系统评审员更好地 处理这些洪流般的数据。该工具的研究和开发将由经验丰富的系统审查调查人员团队与塔夫茨大学的计算机科学家进行，塔夫茨大学的计算机科学家去年因塔夫茨获得NIH临床翻译科学奖(CTSA)之一而开始合作。我们将通过多种渠道传播这项新技术，包括但不限于出版、在会议上发表演讲、探索医疗保健研究和质量机构(AHRQ)循证实践中心(EPC)计划对其采用的兴趣、Cochrane协作、CTSA网络和其他进行系统评估的组织，以及制作教程材料。我们的目标是： 1.研究设计并实现了一个基于机器学习的半自动系统 以信息检索的方法识别相关摘要，以提高系统评价的准确性和效率。 2.开发Abstrackr，这是一个具有用于筛选摘要的图形用户界面的开放源码系统，它应用目标1中开发的方法自动排除不相关的摘要/文章。 3.评估目标1中开发的主动学习模型的性能以及目标1的功能 Abstrackr在AIM 2中开发，通过应用于手动筛选的数据集的集合 生物医学摘要，随后将公开提供，用作存储库，以刺激 在机器学习和信息检索社区进行研究。

DESCRIPTION (provided by applicant): In this three-year project, we aim to apply state-of-the-art information analysis technologies to assist the production of systematic reviews and meta-analyses that are increasingly being used as a foundation for evidence-based medicine (EBM) and comparative effectiveness reviews. We plan to develop a human guided computerized abstract screening tool to greatly reduce the need to perform a tedious but crucial step of manually screening many thousands of abstracts generated by literature searches in order to retrieve a small fraction potentially relevant for further analysis. This tool will combine proven machine learning techniques with a new open source tool that enables management of the screening process. This new technology will enable investigators to screen abstracts in a small fraction of the time compared to the current manual process. It will reduce the time and cost of producing systematic reviews, provide clear documentation of the process and potentially perform the task more accurately. With the acceptance of EBM and increasing demands for systematic reviews, there is a great need for tools to assist in generating new systematic reviews and in updating them. This need cannot be more pressing. The recent passage of the American Recovery and Reinvestment Act and the $1.1 billion allocated for comparative effectiveness research have created an unprecedented need for systematic reviews and opportunities to improve the methodologies and efficiency of their conduct. We herein propose the development of novel, open-source software to help systematic reviewers better cope with these torrents of data. The research and development of this tool will be carried out by a highly experienced team of systematic review investigators with computer scientists at Tufts University who began to collaborate last year as a result of Tufts being awarded one of the NIH Clinical Translational Science Awards (CTSA). We will pursue dissemination of the new technology through numerous channels including, but not limited to publication, presentation at conferences, exploring interest in its adoption by the Agency for Healthcare Research and Quality (AHRQ) Evidence-based Practice Center (EPC) Program, Cochrane Collaboration, CTSA network, and other groups conducting systematic reviews, and production of tutorial material. Our aims are: 1. Conduct research to design and implement a semi-automated system using machine learning and information retrieval methods to identify relevant abstracts in order to improve the accuracy and efficiency of systematic reviews. 2. Develop Abstrackr, an open-source system with a Graphical User Interface (GUI) for screening abstracts, that applies the methods developed in Aim 1 to automatically exclude irrelevant abstracts/articles. 3. Evaluate the performance of the active learning model developed in Aim 1 and the functionality of Abstrackr developed in Aim 2 through application to a collection of manually screened datasets of biomedical abstracts that will subsequently be made publicly available for use as a repository to spur research in the machine learning and information retrieval communities.