Included studies
- In Extraction form, set each outcome, its type and the time point to extract; keep, add or remove study fields.
- Extract by hand, or start the AI pre-fill: the local model fills the form with a quotation for every value (yellow until you check it).
- Open each study, check or correct every row (✓), and confirm. Confirmed studies form the network dataset for Analysis.
Extraction form
Machine generate
0%
Manual generate
0%
Baseline information
Extraction framework
Participant
Intervention
Continuous outcomes
Outcomes to extract
| Outcome | Time point | Accepted window |
|---|
Dichotomous outcomes
Outcomes to extract
| Outcome | Time point | Accepted window |
|---|
Choose outcome type
| Outcome | Type | Time point | Accepted window |
|---|
Other information
Method and research references
- Cochrane Handbook v6.5, chapter 5 (what to collect; outcomes with timing; dual extraction, MECIR C45–C46) and chapter 6 (SD from SE or CI; multi-arm trials).
- Gartlehner G, et al. LLM-assisted data extraction with human verification vs human-only extraction (SWAR, 9,341 data elements). Ann Intern Med 2025;178:1763-71: incorrect data 9.0% vs 11.0%.
- Konet A, et al. Res Synth Methods 2024 (Claude vs GPT-4 extraction accuracy). Schmidt L, et al. F1000Research 2025 (living review of automated extraction).
- Cochrane, Campbell, JBI and CEE position statement on AI in evidence synthesis (2025): human oversight, transparent reporting.
- Arm-level long format as used by netmeta pairwise() and multinma set_agd_arm(); node definition and original labels kept for audit (PRISMA-NMA).