Assessing risk of bias of cohort studies with large language models
Bibliographic record
Abstract
Abstract This study aims to explore the feasibility and accuracy of utilizing large language models (LLMs) to assess the risk of bias (ROB) in cohort studies. We conducted a pilot and feasibility study in 30 cohort studies randomly selected from reference lists of published Cochrane reviews. We developed a structured prompt to guide the ChatGPT-4o, Moonshot-v1-128k, and DeepSeek-V3 to assess the ROB of each cohort twice. We used the ROB results assessed by three evidence-based medicine experts as the gold standard, and then we evaluated the accuracy of LLMs by calculating the correct assessment rate, sensitivity, specificity, and F 1 scores for overall and item-specific levels. The consistency of the overall and item-specific assessment results was evaluated using Cohen’s kappa (κ) and prevalence-adjusted bias-adjusted kappa. Efficiency was estimated by the mean assessment time required. This study assessed three LLMs (ChatGPT-4o, Moonshot-v1-128k, and DeepSeek-V3) and revealed distinct performance across eight assessment items. Overall accuracy was comparable (80.8%–83.3%). Moonshot-v1-128k showed superior sensitivity in population selection (0.92 versus ChatGPT-4o’s 0.55, P < 0.001). In terms of F 1 scores, Moonshot-v1-128k led in population selection ( F = 0.80 versus ChatGPT-4o’s 0.67, P = 0.004). ChatGPT-4o demonstrated the highest consistency (mean κ = 96.5%), with perfect agreement (100%) in outcome confidence. ChatGPT-4o was 97.3% faster per article (32.8 seconds versus 20 minutes manually) and outperformed Moonshot-v1-128k and DeepSeek-V3 by 47–50% in processing speed. The efficient and accurate assessment of ROB in cohort studies by ChatGPT-4o, Moonshot-v1-128k, and DeepSeek-V3 highlights the potential of LLMs to enhance the systematic review process.
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 | 0.681 | 0.821 |
| Meta-epidemiology (narrow) | 0.003 | 0.003 |
| Meta-epidemiology (broad) | 0.006 | 0.020 |
| Bibliometrics | 0.011 | 0.006 |
| Science and technology studies | 0.002 | 0.004 |
| Scholarly communication | 0.007 | 0.007 |
| Open science | 0.004 | 0.009 |
| Research integrity | 0.004 | 0.003 |
| 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".