Performance of Google NotebookLM for AI-assisted data extraction and consensus statement generation in a heterogenous systematic review on inflammatory bowel disease, obesity, and cardiometabolic comorbidities: A Methodological Report
AI tools like Google NotebookLM could make systematic reviews and consensus statement processes faster and less labor-intensive for IBD-related evidence, while maintaining very low rates of major factual errors.
Patients and clinicians may see accelerated synthesis of evidence that informs guidelines, but human checking remains necessary to catch omissions.
Researchers conducting systematic reviews or guideline panels in IBD and related metabolic comorbidity areas; clinical methodologists and guideline developers; clinicians who rely on rapidly produced evidence syntheses.
What To Know
This preprint reports a methodological evaluation of Google NotebookLM used to assist data extraction and to draft RAND/UCLA-style consensus statements in a systematic review covering IBD, obesity, and cardiometabolic comorbidities.
The authors organized studies into domain-specific AI notebooks, used structured (PICO) prompts to generate evidence tables, and had two independent human reviewers validate AI outputs against full texts using an error taxonomy.
Reported results across 57 studies: 1,710 extracted data cells with 91.17% cell-level accuracy and 99.77% critical accuracy (major factual errors rare). Most flagged issues were omissions or incomplete extractions.
The AI-assisted pipeline reduced estimated person-hours substantially and produced a majority of candidate consensus statements that were incorporated after expert review. What To Know This paper examines whether a Google LLM tool can speed up and assist complex evidence extraction and consensus-statement drafting for a multidisciplinary IBD-related review.
The authors report large time savings and low rates of major factual errors, but note that missing or incomplete extractions were the main limitation and required human review.
The workflow combined structured prompting (PICO) with human validation: AI produced candidate data cells and draft statements, while independent reviewers checked and corrected outputs against source full texts. Roughly half of finalized consensus statements originated from AI drafts and most were retained after expert editing.
These findings are described in a medRxiv preprint (abstract-level extraction provided); the work is methodological and does not change clinical care. The tool’s performance may vary with different review topics, prompt design, or AI updates; human
This is a preprint (medRxiv) and represents a methodological report based on an abstract-level extraction. The authors emphasize that extractive incompleteness (omissions/incomplete fields) was the primary limitation and that human review/validation was part of the workflow. Results may not generalize to all topics, LLM versions, or prompt strategies.
Summary grounded in the provided abstract; full peer-reviewed assessment is not available.
