The Literature Review Network: An Explainable Artificial Intelligence for Systematic Literature Reviews, Meta-analyses, and Method Development

AI can significantly reduce the time burden of systematic reviews: LRN completed a literature review process in 288.6 minutes over 5 days compared to 19,920 minutes (over 11 months) for manual expert review, demonstrating potential to address resource constraints that currently limit high-quality evidence synthesis.

Non-expert training can produce reliable AI models for literature screening: A non-expert successfully trained LRN models that achieved 84.78-85.71% accuracy and demonstrated high interrater reliability (k = 0.4953), suggesting that explainable AI tools may democratize access to systematic review processes beyond specialized experts.

Explainability is essential for validating AI-assisted reviews: LRN's ability to identify and explain which specific terms (such as "reduce," "accident," "sharp," and "double-gloving") drove its classification decisions enabled researchers to understand and verify the AI's reasoning, supporting responsible deployment in evidence synthesis.

AI models may identify relevant literature that humans miss: One LRN model covered 91.51% of relevant literature despite disagreeing with the non-expert reviewer's judgments, suggesting AI can complement human review by catching studies humans might overlook, though this requires careful validation.

Adherence to research standards is critical for responsible AI in evidence synthesis: LRN was designed to meet PRISMA 2020 standards and used blinded comparisons between AI and expert results, demonstrating that responsible AI development requires alignment with established methodological guidelines and rigorous validation protocols.