Novel AI applications in systematic review: GPT-4 assisted data extraction, analysis, review of bias
Bibliographic record
Abstract
OBJECTIVE: To assess custom GPT-4 performance in extracting and evaluating data from medical literature to assist in the systematic review (SR) process. DESIGN: A proof-of-concept comparative study was conducted to assess the accuracy and precision of custom GPT-4 models against human-performed reviews of randomised controlled trials (RCTs). SETTING: Four custom GPT-4 models were developed, each specialising in one of the following areas: (1) extraction of study characteristics, (2) extraction of outcomes, (3) extraction of bias assessment domains and (4) evaluation of risk of bias using results from the third GPT-4 model. Model outputs were compared against data from four SRs conducted by human authors. The evaluation focused on accuracy in data extraction, precision in replicating outcomes and agreement levels in risk of bias assessments. PARTICIPANTS: Among four SRs chosen, 43 studies were retrieved for data extraction evaluation. Additionally, 17 RCTs were selected for comparison of risk of bias assessments, where both human comparator SRs and an analogous SR provided assessments for comparison. INTERVENTION: Custom GPT-4 models were deployed to extract data and evaluate risk of bias from selected studies, and their outputs were compared to those generated by human reviewers. MAIN OUTCOME MEASURES: Concordance rates between GPT-4 outputs and human-performed SRs in data extraction, effect size comparability and inter/intra-rater agreement in risk of bias assessments. RESULTS: When comparing the automatically extracted data to the first table of study characteristics from the published review, GPT-4 showed 88.6% concordance with the original review, with <5% discrepancies due to inaccuracies or omissions. It exceeded human accuracy in 2.5% of instances. Study outcomes were extracted and pooling of results showed comparable effect sizes to comparator SRs. A review of bias assessment using GPT-4 showed fair-moderate but significant intra-rater agreement (ICC=0.518, p<0.001) and inter-rater agreements between human comparator SR (weighted kappa=0.237) and the analogous SR (weighted kappa=0.296). In contrast, there was a poor agreement between the two human-performed SRs (weighted kappa=0.094). CONCLUSION: Customized GPT-4 models perform well in extracting precise data from medical literature with potential for utilization in review of bias. While the evaluated tasks are simpler than the broader range of SR methodologies, they provide an important initial assessment of GPT-4's capabilities.
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
Direct model labels (unvalidated)
Per-model category and study-design labels from the labeling rounds. They are machine output, unvalidated, and the disagreement between models ships as data. No study design here is MEDLINE-validated yet.
| Model arm | Categories | Study design | Confidence |
|---|---|---|---|
| gemma | Metaresearch Domain: Methods · Genre: Empirical About the Canadian research system: no · About a Canadian topic: no | Simulation or modeling | low |
| gpt | Metaresearch Domain: Methods · Genre: Methods About the Canadian research system: no · About a Canadian topic: no | Systematic review | medium |
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 | 0.232 | 0.522 |
| Meta-epidemiology (narrow) | 0.004 | 0.004 |
| Meta-epidemiology (broad) | 0.005 | 0.013 |
| Bibliometrics | 0.013 | 0.012 |
| Science and technology studies | 0.002 | 0.002 |
| Scholarly communication | 0.009 | 0.006 |
| Open science | 0.004 | 0.009 |
| Research integrity | 0.004 | 0.003 |
| Insufficient payload (model declined to judge) | 0.024 | 0.006 |
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, unvalidatedLabeled directly by 2 models reading the full record.
The models disagree on parts of this classification; every voice is preserved in the section at the end of the page.
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".