Prompting is all you need: LLMs for systematic review screening
Bibliographic record
Abstract
Abstract Systematic reviews (SRs) are the highest standard of evidence, shaping clinical practice guidelines, policy decisions, and research priorities. However, their labor-intensive nature, including an initial rigorous article screen by at least two investigators, delays access to reliable information synthesis. Here, we demonstrate that large language models (LLMs) with intentional prompting can match human screening performance. We introduce Framework Chain-of-Thought, a novel prompting approach that directs LLMs to systematically reason against predefined frameworks. We evaluated our prompts across ten SRs covering four common types of SR questions (i.e., prevalence, intervention benefits, diagnostic test accuracy, prognosis), achieving a mean accuracy of 93.6% (range: 83.3-99.6%) and sensitivity of 97.5% (89.7-100%) in full-text screening. Compared to experienced reviewers (mean accuracy 92.4% [76.8-97.8%], mean sensitivity 75.1% [44.1-100%]), our full-text prompt demonstrated significantly higher sensitivity in four reviews (p<0.05), significantly higher accuracy in one review (p<0.05), and comparable accuracy in two of five reviews (p>0.05). While traditional human screening for an SR of 7000 articles required 530 hours and $10,000 USD, our approach completed screening in one day for $430 USD. Our results establish that LLMs can perform SR screening with performance matching human experts, setting the foundation for end-to-end automated SRs.
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 | 0.438 | 0.844 |
| Meta-epidemiology (narrow) | 0.003 | 0.004 |
| Meta-epidemiology (broad) | 0.005 | 0.006 |
| Bibliometrics | 0.011 | 0.012 |
| Science and technology studies | 0.002 | 0.004 |
| Scholarly communication | 0.012 | 0.020 |
| Open science | 0.005 | 0.014 |
| Research integrity | 0.006 | 0.006 |
| Insufficient payload (model declined to judge) | 0.025 | 0.009 |
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".