Assisting Systematic Review Preparation Using Automated Document Classification

描述（由申请人提供）： 在这个新的研究者发起的项目中提出的工作研究了一个假设，即基于机器学习的文本分类技术可以显著提高系统评价（SR）更新过程的效率。由于新的信息不断出现，医学也在不断变化，因此工作人员报告必须定期更新，以便正确地反映特定时间内现有的最佳医学知识。 为了支持对这一假设的研究，这里提出的工作将实现四个具体目标： 1.改进和进一步开发优化用于分类的文本分类算法 各种治疗领域的系统评价更新文献。将研究使用几种不同的机器学习技术和策略的比较分析，以及将期刊文章表示为输入到该过程的特征向量的各种方法。 2.识别和评价系统评价专家的偏好和高召回率和高精度分类系统之间的权衡。在创建工作人员代表的过程中，有几个机会可以纳入这一技术。这些应用程序中的每一个都有单独和唯一的精确度和召回率权衡阈值，将根据系统评价的益处进行研究。 3.文本分类算法的前瞻性评估。我们将验证我们的方法的性能， 预计未来的数据。 4.制定全面的黄金标准测试和培训集，以激励和评估 这方面的工作和今后的工作。 这项研究与公共卫生的长期相关性在于，自动化文档分类将 更有效地利用专家资源，创建系统性评价。这将增加 在一定程度的公众支持下，审查的数量和质量。由于最新的系统评价对于建立广泛的高质量实践标准和指南至关重要，因此整体公共卫生将从这项工作中受益。

DESCRIPTION (provided by applicant): The work proposed in this new investigator initiated project studies the hypothesis that machine learning-based text classification techniques can add significant efficiencies to the process of updating systematic reviews (SRs). Because new information constantly becomes available, medicine is constantly changing, and SRs must undergo periodic updates in order to correctly represent the best available medical knowledge at a given time. To support studying this hypothesis, the work proposed here will undertake four specific aims: 1. Refinement and further development of text classification algorithms optimized for use in classifying literature for the update of systematic reviews on a variety of therapeutic domains. Comparative analysis using several different machine learning techniques and strategies will be studied, as well as various means of representing the journal articles as feature vectors input to the process. 2. Identification and evaluation of systematic review expert preferences and trade offs between high recall and high precision classification systems. There are several opportunities for including this technology in the process of creating SRs. Each of these applications has separate and unique precision and recall tradeoff thresholds that will be studied based on the benefit to systematic reviews. 3. Prospective evaluation of text classification algorithms. We will verify that our approach performs as expected on future data. 4. Development of comprehensive gold standard test and training sets to motivate and evaluate the proposed and future work in this area. The long term relevance of this research to public health is that automated document classification will enable more efficient use of expert resources to create systematic reviews. This will increase both the number and quality of reviews for a given level of public support. Since up-to-date systematic reviews are essential for establishing widespread high quality practice standards and guidelines, the overall public health will benefit from this work.