Machine learning techniques for the automation of literature reviews and systematic reviews in EFSA
The present document has been produced and adopted by the bodies identified above as author(s). This task has been carried out exclusively by the author(s) in the context of a contract between the European Food Safety Authority and the author(s), awarded following a tender procedure. The present document is published complying with the transparency principle to which the Authority is subject. It may not be considered as an output adopted by the Authority. The European Food Safety Authority reserves its rights, view and position as regards the issues addressed and the conclusions reached in the present document, without prejudice to the rights of the authors.
This Report presents the results from EFSA project RC/EFSA/AMU/2016/01 related to the implementation of machine learning techniques for literature reviews and systematic reviews in EFSA. An overview of the different steps of a systematic review is provided, along with possible ways for automation. Although it was found that most steps could benefit from automation, it was also observed that some steps require more sophisticated methods than those encompassed within the machine learning framework. Availability of data and methodology allowed for the development of an automatic screening tool based on several machine learning techniques. The developed shiny R application can be used for the screening of abstracts and full texts. Properties of machine learning techniques are discussed in this Report together with their most important advantages and disadvantages. The latter discussion includes both general properties, as well as context‐specific properties based on their performance in three case studies. Although creating a universal automatic data extraction tool was considered to be infeasible in this stage, this step of the systematic review was addressed to allow the reviewer to scan the uploaded pdf files for certain words or string of words. Based on observations from the performed case studies, recommendations were made regarding which methods are preferred in specific situations. More explicitly, a discussion is made about the performance of the classifiers with respect to the magnitude of the pool of papers to be screened as well as to the amount of imbalance, referring to the proportion of relevant and irrelevant papers. Finally, it was concluded that the results presented in this report provide proof that the developed shiny application could be efficiently used in combination with other software such as DistillerSR.