A Scoping Review of Adopted Information Extraction Methods for RCTs

Area of Science:

• Medical Informatics
• Computational Linguistics
• Evidence-Based Medicine

Background:

• Randomized controlled trials (RCTs) are vital for therapeutic evidence but contain vast unstructured data.
• Extracting key information from RCT reports is challenging and time-consuming.
• Automating information extraction from RCTs is essential for efficient decision-making and evidence identification.

Purpose of the Study:

• To explore methods for automating or semi-automating information extraction from RCT reports.
• To identify and review Natural Language Processing (NLP), machine learning, and deep learning techniques applied to RCT data.
• To assess the effectiveness of different information extraction approaches in the context of RCTs.

Main Methods:

• Systematic literature search of PubMed, ACM Digital Library, and Web of Science (2010-2022).
• Focus on published NLP, machine learning, and deep learning methods for information extraction in RCTs.
• Analysis of 26 selected publications detailing extraction frameworks and algorithms.

Main Results:

• 26 publications were reviewed, focusing on automatic extraction of RCT characteristics using PICO frameworks (PIBOSO, PECODR).
• 14 publications (53.8%) extracted key characteristics based on PICO, PIBOSO, and PECODR.
• Common methods included word/phrase matching, Naïve Bayes, BERT, Support Vector Machines, and Conditional Random Fields.

Conclusions:

• Determining the most powerful information extraction system is hindered by a lack of accessible software.
• Deep learning models, particularly Transformers and BERT, demonstrate superior performance in NLP tasks for RCT data.
• Further development and accessibility of NLP tools are needed to enhance information extraction from RCTs.