The accuracy of large language models in labelling neurosurgical ‘case-control studies’ and risk of bias assessment: protocol for a study of interrater agreement with human reviewers
Bibliographic record
Abstract
Abstract Introduction Accurate identification of study designs and risk of bias (RoB) assessment is crucial for evidence synthesis in research. However, mislabelling of case-control studies (CCS) is prevalent, leading to a downgraded quality of evidence. Large Language Models (LLMs), a form of artificial intelligence, have shown impressive performance in various medical tasks. Still, their utility and application in categorising study designs and assessing RoB needs to be further explored. This study will evaluate the performance of four publicly available LLMs (ChatGPT-3.5, ChatGPT-4, Claude 3 Sonnet, Claude 3 Opus) in accurately identifying CCS designs from the neurosurgical literature. Secondly, we will assess the human-LLM interrater agreement for RoB assessment of true CCS. Methods We identified thirty-four top-ranking neurosurgical-focused journals and searched them on PubMed/MEDLINE for manuscripts reported as CCS in the title/abstract. Human reviewers will independently assess study designs and RoB using the Newcastle-Ottawa Scale. The methods sections/full-text articles will be provided to LLMs to determine study designs and assess RoB. Cohen’s kappa will be used to evaluate human-human, human-LLM and LLM-LLM interrater agreement. Logistic regression will be used to assess study characteristics affecting performance. A p -value < 0.05 at a 95% confidence interval will be considered statistically significant. Conclusion If the human-LLM agreement is high, LLMs could become valuable teaching and quality assurance tools for critical appraisal in neurosurgery and other medical fields. This study will contribute to validating LLMs for specialised scientific tasks in evidence synthesis. This could lead to reduced review costs, faster completion, standardisation, and minimal errors in evidence synthesis.
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 | 0.472 | 0.589 |
| Meta-epidemiology (narrow) | 0.008 | 0.008 |
| Meta-epidemiology (broad) | 0.013 | 0.012 |
| Bibliometrics | 0.013 | 0.012 |
| Science and technology studies | 0.007 | 0.010 |
| Scholarly communication | 0.007 | 0.009 |
| Open science | 0.008 | 0.008 |
| Research integrity | 0.013 | 0.010 |
| Insufficient payload (model declined to judge) | 0.041 | 0.018 |
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".