A105 PILOT STUDY ON THE ACCURACY OF CHATGPT IN ARTICLE SCREENING FOR SYSTEMATIC REVIEWS IN GASTROENTEROLOGY
Bibliographic record
Abstract
Abstract Background Systematic reviews synthesize extant research to answer a research question in a way that minimizes bias. After articles for potential inclusion are identified by sensitive searches, screening requires human expert review, which may be time-consuming and subjective. Large language models such as ChatGPT may have potential for this application. Aims This pilot study aims to assess the accuracy of ChatGPT 3.5 in screening of articles for systematic reviews in gastroenterology by (1) identifying if articles were correctly included and (2) excluding articles reported by authors as difficult to assess. Methods We searched the Cochrane Library for gastroenterology systematic reviews (January 1, 2022 to May 31, 2023) and selected the 10 most cited studies. The test set used to determine the accuracy of Open AI’s ChatGPT 3.5 model for included studies was the final list of included studies for each Cochrane review. The test set used for studies challenging to assess was the “excluded studies” list as defined in the Cochrane Handbook. Figure 1 shows the prompt used for the screening query. Articles were omitted if they did not have digital sources, abstracts or methods. Each article was screened 10 times to account for variability within ChatGPT’s outputs. Articles with ≥5 inclusion results were counted as an included study. Results ChatGPT correctly identified included studies at rates ranging from 60% to 100%. ChatGPT correctly identified exlcuded studies at rates ranging from 0% to 50% (Table 1). A total of 265 articles were screened. Conclusions In this pilot study, we demonstrated that ChatGPT is accurate in identifying articles screened for inclusion in Cochrane reviews; however, it is inaccurate in excluding articles described by the authors as being difficult to assess. We hypothesize that the GPT 3.5 model can read for keywords and broad interventions but is unable to reason cognitively, as an expert would, as to why a study may be excluded. We aim to review reasons for exclusion in future work. Table 1. Screening Results of ChatGPT Figure 1. ChatGPT Screening Prompt Funding Agencies None
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 | 0.569 | 0.879 |
| Meta-epidemiology (narrow) | 0.002 | 0.002 |
| Meta-epidemiology (broad) | 0.003 | 0.009 |
| Bibliometrics | 0.007 | 0.010 |
| Science and technology studies | 0.002 | 0.003 |
| Scholarly communication | 0.006 | 0.011 |
| Open science | 0.003 | 0.007 |
| Research integrity | 0.004 | 0.003 |
| Insufficient payload (model declined to judge) | 0.006 | 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".