Enhancing Systematic Reviews: AI Literature Review Integration for Efficiency and Rigor

AI Literature Review Integration is transforming systematic literature reviews (SLRs), as outlined in this preprint. It provides methodological guidance for incorporating artificial intelligence (AI) into SLRs with a human-in-the-loop approach to maintain scientific rigor. AI accelerates resource-intensive stages like screening and data extraction while tackling risks such as bias, hallucinations, and lack of transparency. The core argument is that AI complements human judgment, enabling faster, scalable evidence synthesis aligned with PRISMA and Cochrane standards, without replacing expert oversight.

Tackling SLR Workload Bottlenecks

There are operational challenges in traditional SLRs, which require months of manual effort to screen thousands of records amid exploding scientific output. This leads to extended timelines, workforce strain, and outdated reviews post-publication. AI Literature Review Integration delivers key advantages, including over 50% reductions in completion time across 17 of 25 studies, five- to sixfold decreases in abstract screening duration, and 55%–64% fewer abstracts needing human review, per a pragmatic review. These efficiencies enable scalability for large evidence bases, like sourcing parameters for economic models in health economics and outcomes research (HEOR), while reducing human error in repetitive tasks and improving pattern detection. Vigilant human validation counters AI biases from training data, preserving methodological integrity.

Human Oversight in AI-Driven SLR Stages

The preprint emphasizes human oversight to address AI limitations, offering structured guidance for AI Literature Review Integration across SLR stages, drawn from authors’ experience and literature. It follows the Cochrane Handbook’s process—from research question definition to reporting—via a detailed table. AI prioritizes records for title/abstract screening through relevance ranking refined by human feedback, generates search strings with synonyms and MeSH terms, and extracts structured PICO elements, validated against gold-standard datasets. Rigor uses metrics like precision, recall, F1 score, pilot testing, and inter-reviewer reliability (Cohen’s kappa 0.77–0.88). Tools like RobotReviewer handle risk-of-bias assessments, while NLP-ML platforms aid PRISMA diagrams. Protocols ensure transparency with dual AI-human screening, spot checks, and PRISMA standards.