Empowering Systematic Literature Reviews: Utilizing Generative AI for Comprehensive Literature Screening From Titles and Abstracts to Full-Text
Author(s)
Singh B1, Kaur R2, Rai P3
1Pharmacoevidence, London, UK, 2Pharmacoevidence, Mohali, India, 3Pharmacoevidence, Mohali, PB, India
Presentation Documents
OBJECTIVES: Recent advancements in generative AI, are transforming systematic literature reviews (SLR) by automating and expediting literature screening. This study aims to develop and evaluate an automated system that utilizes advanced language models and embedding techniques for rapid and accurate literature screening.
METHODS: A Python-based interface was developed utilizing the Claude 3.5 Sonnet generative AI model to automate the SLR process in two distinct phases. Initially, a data input sheet containing relevant details (especially titles and abstracts) and eligibility criteria was uploaded to commence the first-stage screening. Upon completion of the first stage, full-text publications were uploaded to begin the second-stage screening. The full texts of the publication were uploaded as small chunks and analysed against the eligibility criteria to complete the second stage screening. A subject matter expert (SME) with over a decade of experience refined the final prompts used in both screening stages. The SME oversaw quality control throughout the entire process.
RESULTS: The first-stage screening of the publications, using title and abstracts, was performed by the both human reviewer and Claude 3.5 Sonnet. The generative AI demonstrated an outstanding performance achieving an accuracy rate of 96.72%, a sensitivity rate of 90.00%, and a specificity of 97.13%. In subsequent steps, the AI interface effectively interacted with uploaded full texts. The model utilized full texts against the eligibility criteria and achieved screening efficiency comparable to that of the first stage. Together, screenings across both stages resulted in saving approximately 12 hours compared to traditional human review processes, despite the small sample size (518 publications).
CONCLUSIONS: The integration of Claude 3.5 Sonnet in a generative AI interface demonstrated high performance and efficiency. Its seamless full-text interaction handling and consistent efficiency in subsequent stages accelerate and streamline systematic reviews, yielding significant time savings over traditional manual methods.
Conference/Value in Health Info
Value in Health, Volume 27, Issue 12, S2 (December 2024)
Code
MSR140
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
No Additional Disease & Conditions/Specialized Treatment Areas