A meta-analysis takes trials that tested the same thing and combines them into one result.
In the NR example in this repo, 4 trials measured systolic blood pressure and each found NR lowered it by 1.5 to 3 mmHg. None of them can show an effect on its own, since every trial's confidence interval crosses zero. Combined, with the more precise trials counting more, they show 2.5 mmHg lower (95% CI 5.1 lower to 0.2 higher). So the combined result favors a small drop, but the interval still includes no change.
metaproof does those steps for every outcome in your review. It takes each trial's means and SDs and works out the difference between the treatment and control groups (like NR and placebo) with its confidence interval. It weights each trial by how precise that difference is, which mostly comes down to sample size, and combines them into one result.
You put each trial's numbers into one spreadsheet and it gives the combined result, how much the trials disagree with each other, the range a new trial would likely fall in, and the forest and funnel plots. It also shows how much each trial counts and what happens when each one is left out, and lists any trial it couldn't use in excluded.csv with the reason.
It follows Cochrane's guide for systematic reviews, and its results were checked against a second program in R.
Made by Dr. Layth Qassem, PharmD, MSACI. Free to use under the MIT license. How to cite it is in REFERENCES.md.
On a Mac, open Terminal and run:
xcode-select --install
curl -LsSf https://astral.sh/uv/install.sh | sh
git clone https://github.com/editnori/metaproof
cd metaproof
uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python --require-hashes -r requirements.txt
On Windows, install Git from https://git-scm.com/download/win, then open PowerShell and run:
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
git clone https://github.com/editnori/metaproof
cd metaproof
uv venv --python 3.12 .venv
uv pip install --python .venv\Scripts\python.exe --require-hashes -r requirements.txt
If PowerShell can't find uv, close it and open it again. On Windows use .venv\Scripts\python wherever the commands below say .venv/bin/python.
Fill in templates/extraction_sheet.xlsx, one row per study group, outcome and time point. The data_dictionary tab says what goes in each column. Then check it for mistakes and save it as a CSV:
.venv/bin/python templates/check_extraction.py extraction_sheet.xlsx --write-csv extraction.csv
Write a short settings file listing your outcomes (examples/nicotinamide-riboside/SYNTHETIC_nr_like_config.yaml is a full example), then run it:
.venv/bin/python -m metaproof run extraction.csv --config config.yaml --out results
examples/nicotinamide-riboside/ has a search strategy, a list of candidate NR trials and a sample spreadsheet. The numbers in the sample spreadsheet are made up and labeled SYNTHETIC. Only the PubMed and ClinicalTrials.gov searches have been run, and trial list rows marked UNVERIFIED still need checking against the paper.
It uses a random-effects model, since studies differ in dose, length and population. The confidence interval gets wider when the studies disagree, and it won't pool fewer than 3 studies. If a study reports change from baseline without its SD, it fills one in with a standard assumption and reruns at two others so you can see if that choice matters. Crossover trials are analyzed as paired data, and when a trial has several doses against one placebo group, that placebo group is only counted once. Medians and IQRs get converted to means and SDs. Subgroups and the funnel plot test only run with 10 or more studies because they aren't reliable with fewer.
Every choice and its reason is in templates/protocol_template.txt, which you can use as your protocol. The templates folder also has the PRISMA 2020 flow diagram, the RoB 2 risk of bias tool and a GRADE summary of findings table.
Every check ran in Python and again in R with metafor, a widely used R package for meta-analysis. Across three published datasets, the NR example and a set of edge cases, 696 checks matched and 0 didn't. 346 automated tests check each calculation against worked examples. The core formulas were also proved mathematically, like the pooled result always falling between the smallest and largest study (validation/proofs/RECEIPTS.txt).
It only handles continuous outcomes like blood pressure or lab levels for now. Yes/no outcomes like adverse events aren't in yet.