论文标题
一个开源集成框架，用于在系统评论中进行引文收集和筛选的自动化
An open-source integrated framework for the automation of citation collection and screening in systematic reviews
论文作者
论文摘要
科学生产的指数增长使二级文献缩写越来越苛刻。我们介绍了一个新的开源框架，用于系统评价，可大大减少时间和工作量，以收集和筛选科学文献。该框架提供了三个主要工具：1）一种自动引用搜索引擎和管理器，该引擎和经理通过统一的查询语法从多个在线来源收集记录，2）基于迭代的人机交互，基于人机交互的贝叶斯，主动机器学习工具，以提高预测准确性，3）一个半功率，数据驱动的查询生成器，从而从现有的cileties创建新的搜索Queries，以创建新的搜索设置。为了评估自动筛选器的性能，我们估计了使用贝叶斯模拟的后验敏感性和效率[90％可信间隔]，以预测未发现潜在相关记录的分布。在一个示例主题上进行了测试，该框架通过引文经理收集了17,755个独特的记录。 766记录需要评估人类，而其余的则由自动分类器排除；理论效率为95.6％[95.3％，95.7％]，灵敏度为100％[93.5％，100％]。从标记的数据集生成了一个新的搜索查询，并收集了82,579个记录。自动筛查后，只有567个记录需要人类审查，并发现了六个额外的积极匹配。总体预期灵敏度降至97.3％[73.8％，100％]，而效率提高到98.6％[98.2％，98.7％]。该框架可以通过简化引用收集和筛查，同时证明出色的灵敏度来大大减少进行大量文献审查所需的工作量。这样的工具可以改善系统评价的标准化和可重复性。
The exponential growth of scientific production makes secondary literature abridgements increasingly demanding. We introduce a new open-source framework for systematic reviews that significantly reduces time and workload for collecting and screening scientific literature. The framework provides three main tools: 1) an automatic citation search engine and manager that collects records from multiple online sources with a unified query syntax, 2) a Bayesian, active machine learning, citation screening tool based on iterative human-machine interaction to increase predictive accuracy and, 3) a semi-automatic, data-driven query generator to create new search queries from existing citation data sets. To evaluate the automatic screener's performance, we estimated the median posterior sensitivity and efficiency [90% Credible Intervals] using Bayesian simulation to predict the distribution of undetected potentially relevant records. Tested on an example topic, the framework collected 17,755 unique records through the citation manager; 766 records required human evaluation while the rest were excluded by the automatic classifier; the theoretical efficiency was 95.6% [95.3%, 95.7%] with a sensitivity of 100% [93.5%, 100%]. A new search query was generated from the labelled dataset, and 82,579 additional records were collected; only 567 records required human review after automatic screening, and six additional positive matches were found. The overall expected sensitivity decreased to 97.3% [73.8%, 100%] while the efficiency increased to 98.6% [98.2%, 98.7%]. The framework can significantly reduce the workload required to conduct large literature reviews by simplifying citation collection and screening while demonstrating exceptional sensitivity. Such a tool can improve the standardization and repeatability of systematic reviews.
