Not Ready for Prime Time: Limitations of a Retrieval-Augmented Generation Large Language Model in Assessing Risk of Bias in Observational Studies
Bibliographic record
Abstract
Background: Current research has focused on the use of large language models (LLMs) to augment systematic reviews. LLMs are limited by their vulnerability to "hallucinations"; retrieval-augmented generation (RAG) reduces these by limiting the model's source knowledge to user-provided material. The purpose of this study was to evaluate the accuracy and reliability of a RAG-LLM in quality assessment of observational studies in pediatric orthopaedic literature as compared to manual review. Methods: Previously published systematic reviews of observational studies in pediatric orthopaedics containing reported Newcastle-Ottawa Scale (NOS) scores from our group were included. After uploading observational study source files, NotebookLM (Google, Mountain View, CA) evaluated each of the included studies using the NOS scoring sheet. Agreement among scores across all NotebookLM trials was determined using a two-way random, average measures, absolute agreement intraclass correlation coefficient [ICC(2,k)]. Agreement among individual scores generated by each NotebookLM instance (LM1, LM2, LM3, and LM4) and ground truth (published manual review score) was calculated using a two-way random, single measures, absolute agreement intraclass correlation coefficient [ICC(2,1)]. Results: Two systematic reviews comprising a total of 27 observational studies were included. ICC across all measurements (ICC(2,k)-Reviewer-LM1,2,3,4) was 0.69 (95% CI: 0.46-0.84), indicating moderate agreement. ICC comparing individual NotebookLM scores to ground truth demonstrated poor agreement [ICC(2,1) LM1-Reviewer = 0.27 (95% CI: -0.064 to 0.57), LM2-Reviewer = 0.18 (95% CI: -0.12 to 0.48), LM3-Reviewer = 0.081 (95% CI: -0.24 to 0.41), and LM4-Reviewer = 0.23 (95% CI: -0.14 to 0.55)]. Percent agreement ranged from 14.8% to 29.6%. Single measures ICCs comparing individual NotebookLM scores across multiple trials demonstrated moderate-to-poor agreement. Conclusions: NotebookLM demonstrated low reliability and accuracy in performing quality assessment of observational studies. Caution should be taken when implementing LLMs to augment research efforts in pediatric orthopaedics. Key Concepts: (1)NotebookLM (Google, Mountain View, CA) demonstrated low reliability and accuracy in performing quality assessment of observational studies.(2)Caution should be taken when implementing artificial intelligence tools such as large language models (LLMs) to augment research efforts, even retrieval-augmented generation (RAG)-LLM models that reduce hallucinations.(3)Until emerging artificial intelligence technologies are further validated it remains essential that researchers and clinicians continue to critically appraise new studies independently. Level of Evidence: IV.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.792 | 0.919 |
| Meta-epidemiology (narrow) | 0.005 | 0.004 |
| Meta-epidemiology (broad) | 0.010 | 0.021 |
| Bibliometrics | 0.016 | 0.017 |
| Science and technology studies | 0.003 | 0.009 |
| Scholarly communication | 0.013 | 0.013 |
| Open science | 0.008 | 0.012 |
| Research integrity | 0.004 | 0.005 |
| Insufficient payload (model declined to judge) | 0.005 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; the direct Gemma label and the distilled Codex classifier agree on what is shown here.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".