AgentSLR: Revolutionizing Systematic Literature Reviews in Epidemiology with AI

Systematic literature reviews play a critical role in consolidating scientific knowledge, but they often present challenges such as high costs, scalability issues, and lengthy timelines. These factors can hinder the development of evidence-based policies. Our research investigates the potential of large language models to fully automate the systematic review process, which includes retrieving articles, screening them, extracting data, and synthesizing reports. Focusing on reviews related to nine priority pathogens identified by the WHO, we validated our open-source tool, AgentSLR, against expert-curated benchmarks. The findings indicate that AgentSLR can match the performance of human researchers while drastically reducing the time needed for reviews from around seven weeks to just 20 hours, achieving a remarkable 58-fold increase in speed. We also analyzed five leading models and found that the effectiveness of systematic literature reviews is influenced more by the unique abilities of each model than by their size or operational costs. Through a human-in-the-loop approach, we pinpointed critical areas where errors occur. Our study highlights the potential of agentic AI to significantly enhance the efficiency of scientific evidence synthesis in specialized fields.