Can large language models fully automate or partially assist paper selection in systematic reviews?
Bibliographic record
Abstract
BACKGROUND/AIMS: Large language models (LLMs) have substantial potential to enhance the efficiency of academic research. The accuracy and performance of LLMs in a systematic review, a core part of evidence building, has yet to be studied in detail. METHODS: We introduced two LLM-based approaches of systematic review: an LLM-enabled fully automated approach (LLM-FA) utilising three different GPT-4 plugins (Consensus GPT, Scholar GPT and GPT web browsing modes) and an LLM-facilitated semi-automated approach (LLM-SA) using GPT4's Application Programming Interface (API). We benchmarked these approaches using three published systematic reviews that reported the prevalence of diabetic retinopathy across different populations (general population, pregnant women and children). RESULTS: The three published reviews consisted of 98 papers in total. Across these three reviews, in the LLM-FA approach, Consensus GPT correctly identified 32.7% (32 out of 98) of papers, while Scholar GPT and GPT4's web browsing modes only identified 19.4% (19 out of 98) and 6.1% (6 out of 98), respectively. On the other hand, the LLM-SA approach not only successfully included 82.7% (81 out of 98) of these papers but also correctly excluded 92.2% of 4497 irrelevant papers. CONCLUSIONS: Our findings suggest LLMs are not yet capable of autonomously identifying and selecting relevant papers in systematic reviews. However, they hold promise as an assistive tool to improve the efficiency of the paper selection process in systematic reviews.
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 | 0.417 | 0.775 |
| Meta-epidemiology (narrow) | 0.005 | 0.004 |
| Meta-epidemiology (broad) | 0.007 | 0.014 |
| Bibliometrics | 0.012 | 0.012 |
| Science and technology studies | 0.002 | 0.003 |
| Scholarly communication | 0.014 | 0.014 |
| Open science | 0.006 | 0.010 |
| Research integrity | 0.005 | 0.005 |
| Insufficient payload (model declined to judge) | 0.008 | 0.004 |
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".