A Locally Executed Agentic Artificial Intelligence Framework for Deduplication, Screening, and Structured Data Extraction in Spine Surgery Systematic Reviews

Abstract

Systematic reviews and meta-analyses remain time-consuming and labor-intensive. We developed and validated a locally executed agentic artificial intelligence (AI) framework for deduplication, screening, and structured data extraction in systematic reviews, reported following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)-Transparent Reporting of AI in Comprehensive Evidence Synthesis (trAIce) guidelines. A multiagent pipeline of specialized agents for deduplication, title/abstract screening, structured data extraction, and verification was executed entirely locally to ensure data governance and reproducibility. Human-in-the-loop validation compared outputs against dual independent reviewers across 3 spine surgery systematic reviews, assessing accuracy, inter-rater agreement (Cohen’s κ), time savings, and clinically critical error rates. Across 6,214 records, deduplication achieved near-perfect agreement with human reviewers (κ = 0.98), title and abstract screening yielded higher concordance than human screening (κ = 0.91) while reducing full-text review volume by 83%, and structured data extraction reached substantial agreement (κ = 0.87). The framework reduced reviewer time by 91.1% (90.8%-91.4%), a mean saving of 19.5 hours per review (p < 0.001). Clinically critical discrepancies were rare (<1%) and traceable, with no fabricated or hallucinated data introduced. A locally executed agentic AI framework has the potential to deliver accurate, efficient, and secure automation of systematic review tasks with human oversight, offering a reproducible pathway for trustworthy evidence synthesis under PRISMA-trAIce standards.