sievesafe

Title and abstract screening for systematic reviews that removes records only below a threshold fixed in advance.

You give it your search export and your eligibility criteria. For each record it estimates how likely it is that the record should go on to full-text review. It needs no training data and no labelled examples. Then it works in one of two modes:

• rank (the default) keeps every record, puts the most likely ones first and flags the ones below a safe threshold;
• exclude (opt-in) removes the records below that threshold before human screening, and hands you a random sample of them so you can check the result on your own review.

In both modes your records come back in the format they came in, ready to import into Rayyan, Covidence, EndNote or Zotero, together with a report that gives the PRISMA 2020 count and a draft methods paragraph.

What it promises, and what it doesn't

The safe threshold was set on 20 reviews and then tested, without any change, on 23 other reviews and on Cochrane reviews from a second collection, CLEF TAR. Across all of them it lost one finally included study, in one review, while a large share of the records fell below it. That record is a cardiology paper listed as included in a review about colorectal cancer, and the review itself does not cite it, while it cites the other 48 included studies (label-check.md). So it is probably a labelling mistake, but that CLEF set still failed the rule we had fixed for it. In the benchmarks, the records it removed rarely contained a study that ended up in the review.

It does not promise to match human title and abstract screening. Some records that human screeners sent on to full text do fall below the threshold. In the benchmark all of them were later excluded at full text, but a benchmark is not a guarantee for your review, so check exclude mode on your own data before relying on it, as the RAISE recommendations ask.

on 23 held-out SYNERGY+ reviews (33,001 records)	sievesafe (zero-shot, no labels)	ASReview 3.0.8 (active learning, 1+1 labelled priors)
work saved at 95% recall (median WSS@95)	0.758	0.723
work saved at 100% recall (median WSS@100)	0.791	0.711
reviews where sievesafe saves as much or more (WSS@95 / WSS@100)	18/23 / 18/23

safe threshold (0.06), frozen on 20 other reviews
records below it (flagged in rank mode, removed in exclude mode)	29% of 33,001
finally included studies below it	0 of 597 (recall 100% in 23/23 reviews)
leave-one-review-out, 43 reviews: reviews keeping ≥98% / 100% of their includes	43/43 / 42/43 (threshold 0.06–0.08; 95% upper bound on reviews losing any include: 11%)
records humans passed at title/abstract stage kept above it (12 reviews)	96% on average (lowest 78%); of the rest, 0 finally included
run-to-run correlation of scores (model pinned)	0.996
cost	about US$ 0.05 per 1,000 records

After that, the threshold was tested again on CLEF TAR, still unchanged, with each analysis plan committed to git before any record was scored (benchmark/external/):

CLEF TAR Cochrane reviews, pre-specified plans, threshold 0.06 unchanged	Intervention (20 reviews), criteria: objectives and selection criteria	Intervention (20 reviews), criteria: objectives only	diagnostic test accuracy (8 reviews), criteria: objectives and selection criteria	2017–2018 topics (40 reviews), criteria: objectives and selection criteria
reviews keeping every finally included study	18/18	18/18	8/8	38/39
finally included studies kept	510 of 510	510 of 510	250 of 250	848 of 849
records below the threshold (weighted)	52%	34%	45%	49%
95% upper bound on reviews losing any include	15%	15%	31%	12%

WSS (work saved over sampling) is the share of records a reviewer can skip, reading in the tool's order, and still find 95% or 100% of the included studies. The numbers in these tables are written by the scripts in benchmark/ from public data and the model answers kept in the repository; benchmark/results.md has the full tables.

Install

You need Python 3.11 or newer. With pipx:

pipx install sievesafe

sievesafe uses TypeSafe's Jev model, which needs a paid API key. Set it as an environment variable (TYPESAFE_API_KEY) or keep it in a file:

mkdir -p ~/.config/sievesafe
echo 'TYPESAFE_API_KEY=your-key' > ~/.config/sievesafe/typesafe.env
chmod 600 ~/.config/sievesafe/typesafe.env

Use

To work in the browser, start the local page. Nothing leaves your computer except what is sent to TypeSafe (see Limits).

sievesafe serve

Or from the terminal:

sievesafe screen search.ris --criteria criteria.txt --title "Emicizumab prophylaxis in children with haemophilia A"
sievesafe screen search.ris --criteria criteria.txt --title "…" --mode exclude --budget 1.00

Before spending anything it tells you how many records there are and what they should cost, and asks you to confirm. --budget is a limit in US$: a call is only sent if its estimated cost, added to what was spent and to the calls still running, fits within it. Any record it could not score within the budget is kept at the top of the list and is never removed. Running it again on the same file only pays for records that have not been scored yet.

It reads RIS (Scopus, Web of Science, Rayyan, EndNote, Zotero and others), PubMed/MEDLINE .nbib, Web of Science plain text, and CSV or TSV files separated by , or ; in the usual encodings. The results are written in the same format and encoding:

file	what it contains
to-screen.<ext>	the records for human screening, most likely first, left untouched
ranked.csv	rank, original position, probability, flag and title of every record
excluded.<ext>	exclude mode only: the records that were removed
validation-sample.<ext>	exclude mode only: 50 of the removed records, drawn at random, to screen as a local check
report.md	counts for the PRISMA 2020 diagram, a methods paragraph, the validation evidence and your criteria

Records without an abstract are ranked but never flagged or removed, because the threshold was only validated on records that had one. Duplicates (same DOI or title) are scored once and every copy is kept; removing duplicates is still a job for your reference manager.

Limits

• Title and abstract stage only. It does not read full texts, and it is tuned to let doubtful records through.
• English-language literature. Most of the benchmark reviews restricted their searches to English. Other languages have not been tested.
• A paid, closed model. Each record costs a fraction of a cent, and its title and abstract, along with your review title and criteria, are sent to TypeSafe's API. The model is proprietary. sievesafe pins the version the threshold was calibrated with (jev-1.13.0, the version current at the time); if that version is retired, the threshold has to be recalibrated, and benchmark/ shows how.
• Two benchmarks are not a guarantee. SYNERGY+ and CLEF TAR are made of published reviews with published criteria. In the leave-one-review-out test one review lost one included study; in the external tests one review lost one, and the other three sets lost none. When we gave the model only the review's objectives, it removed fewer records but did not lose more studies; that is a single test, though.
• One test review was seen while choosing the question's wording (Donners 2021, in a four-review pilot). Without it the SYNERGY+ results are the same in substance; see benchmark/results.md.
• A known flaw in how the threshold was set, explained in benchmark/README.md. The corrected procedure gives a slightly higher threshold; sievesafe keeps the original, more cautious one.

How it works

For each record, sievesafe asks Jev one yes/no question: should this record move on to full-text review for this review? Jev is one of TypeSafe's System One models, which return a probability rather than generated text. The model sees the review's title and criteria, the record's title and abstract, and a short policy saying that missing a relevant study is much worse than reading an extra one. Answers are cached on disk by model, criteria and record text. The threshold was set on 20 calibration reviews so that all of their included studies stayed above it.

How to cite

Please cite the software as archived on Zenodo, doi:10.5281/zenodo.22992077 (all versions; each release also has its own DOI), or use CITATION.cff; GitHub's "Cite this repository" button turns it into APA or BibTeX. A preprint describing the benchmark is on its way, and this section will link to it once it is posted.

Help and contributing

For questions or bugs, open an issue in this repository. If you want to contribute, CONTRIBUTING.md explains how to run the checks and why no number in the repository is typed by hand.

License

AGPL-3.0-or-later.