End-to-End Automation of Multilingual Scoping Reviews Using Agentic AI

Scoping reviews are resource-intensive, slow, difficult to scale, and suffer from reproducibility challenges, especially when integrating multilingual and heterogeneously indexed sources. This study introduces an autonomous Agentic AI framework that automates the scoping review workflow while allowing human-in-the-loop oversight when needed. The pipeline can operate end-to-end, from query formulation and multilingual database retrieval to title/abstract screening, metadata harmonization, full-text assessment, and structured evidence extraction, or begin directly at the screening stage when data collection is unnecessary. The pipeline integrates five specialized agents, coordinated via a LangChain-based orchestration layer, ensuring deterministic, reproducible execution. We evaluated the framework using a provided multilingual corpus of 52,051 oral health research records (2014–2024) and 1,092 expert-annotated samples. Because the original dataset lacked key metadata, including MeSH terms, affiliations, and full-text PDFs, we recollected enriched records for approximately 21,000 articles, including expert-annotated samples. In parallel, we retrieved 9,995 full-text PDFs and applied full-text assessment to a subset of studies labeled as Include or Uncertain during title/abstract screening (n = 3,119). Across 1,092 expert-annotated samples, the title and abstract-based screening agent achieved an F1-score of 0.87, increasing to 0.90 when using enriched metadata with MeSH terms and affiliations. The full-text agent classified 86% of eligible studies as primary research, 2% as secondary research, and 12% remained uncertain. Compared to traditional scoping reviews, which require an estimated 16 months, the automated pipeline completed harmonization, screening, and extraction in 46 h. Findings demonstrate that Agentic AI offers a scalable, accurate, and methodologically rigorous approach to multilingual scoping reviews.