Guide

Data extraction for systematic reviews, quoted from the papers

Jev Reviewer reads a study's files in your browser and answers your extraction questions with sentences and table rows copied from them, each with its file and page. You check every answer, write down the value, and export a table with one row per study. It is free, open source, and keeps your files on your computer.

Open Jev Reviewer Source on GitHub

What it does

• Asks every file of a study at once: the trial report, its appendix, the protocol, the statistical analysis plan, the data tables, a conference slide deck.
• Quotes, never paraphrases. The model points at lines; the app copies them word for word with their file and page, paragraph, row or slide, and highlights them where they sit.
• Says when something is not reported instead of guessing: an answer reads Not found, or Unclear with the closest lines.
• Keeps a whole review together as a project: its studies, their files, the questions, the answers and your checks, in your browser.

A review, step by step

1. Remove duplicates from the searches, before screening, if you like: Deduplicate search results takes the exports of every database searched. A pair of records found by both the rules (the same DOI, PubMed id, or title) and Jev is removed; a pair only one of them finds is listed for you to decide. The records that remain download as RIS, with a log of what was removed and why, or go straight on to screening.
2. Screen titles and abstracts. Write the eligibility criteria, one per line. Jev reads each record's title and abstract and says, criterion by criterion, whether the record meets it, does not, or does not say; the likeliest records come first. You decide each record with a press or a key, see where you and Jev disagree, and can exclude in bulk the records Jev finds clearly ineligible, which the methods paragraph reports. The included records become the project's studies, each with its abstract until the full text comes.
3. Create a project for your review in the column on the left.
4. Import the included studies from your reference manager or database export (RIS, BibTeX, EndNote XML, PubMed, Web of Science, CSL JSON, or CSV and Excel with a title column), picked together with the PDFs, as a zip, or as the folder the reference manager exported. Each reference becomes a study named the way reviews cite it, such as Smith 2024, and its files are matched by their attachment names, DOI, title, or first author and year. A preview shows what matched before anything is created. A reference without its full text comes in with its abstract, so it can be asked about until the PDF arrives; importing the list again with the PDFs adds them. Every study is checked for retractions (Retraction Watch, Crossref, OpenAlex, PubMed), and a study with an open access copy in PubMed Central gets a PMC link beside its DOI, with its article and supplementary files one press away. A study added from a PDF gets its full citation from the DOI it prints.
5. Set the questions. Upload your extraction form as CSV, Excel or text, one question per row, or start from a template. A guidance column holds your coding manual, shown under each question in every study. A question you ask in one study can join the list for all of them.
6. Ask them of every study. Each study is asked only what it lacks: a question it never answered, a question reworded since, or every question once a file has been added.
7. Check each answer. Read the quotes against the highlighted place in the file, press the green tick on the right quote (it turns green, becomes your answer, and the other quotes fold away), edit that answer with the pencil if your form needs other words (the numbers in the quotes are one press away), or write it yourself and press Check. A question that does not apply to a study is marked Not applicable. Find looks up exact words at once, to confirm that something really is missing.
8. Exclude what does not belong. Under the file tabs, Exclude asks for a reason; the study stays, struck through, runs skip it, and the table counts the PRISMA flow: full reports assessed, excluded with reasons, included.
9. Judge the risk of bias. With a RoB 2, ROBINS-I or QUADAS-2 template asked, the shield under the file tabs lists each domain with your answers beside it; judge each one, and the overall judgment is suggested from the most serious. The table's Risk of bias tab shows a traffic-light grid and exports it for robvis, which draws the figures.
10. Follow the extraction table: every study against every question, marked reported, unclear, not found, checked, or due to be asked again. A cell opens the study at that answer.
11. Report. The Report tab draws the PRISMA 2020 flow diagram from the project's own numbers (searches, duplicates, screening, full texts and exclusions), as SVG or as the PRISMA2020 package's template for its app, builds the table of included studies from your checked answers, and drafts the methods paragraph.
12. Export and back up. The table exports one row per study, with each question's value and quotes, or one row per quote with its file, place and score. A backup keeps the whole project in one zip, with both sheets as CSV, and restores in any browser.

Templates for extraction forms and risk of bias

Each template asks for the quotes a reviewer needs, in plain words the model can search for. The judgments stay with you and your co-reviewers.

• Trial characteristics (18 questions): design, setting, eligibility, the number randomized, baseline age and sex, interventions, outcomes, follow-up, randomization and blinding methods, attrition, funding and registration.
• Risk of bias in randomized trials, RoB 2 (14): the quotes behind each domain, from the randomization process to the selection of the reported result.
• Risk of bias in non-randomized studies, ROBINS-I (10): confounding, selection, classification of interventions, deviations, missing data, measurement and reporting.
• Diagnostic accuracy studies, QUADAS-2 (12): patient selection, the index test, the reference standard, and flow and timing.
• Intervention description, TIDieR (12): what was given and why, by whom, how, where, when and how much, tailoring, modifications and fidelity.
• Transparency and reproducibility (7): competing interests, funding and the funder's role, registration and its timing, the protocol, data sharing, code sharing, and the reporting guideline followed.
• Outcome data for meta-analysis (10): time points, the measure and its direction, numbers analyzed, means and standard deviations or medians, events, and the reported effect with its confidence interval. The means and the events are answered arm by arm, in a grid the numbers in the quotes fill with one press each, and export in the long layout the meta and netmeta packages for R read.

How the quotes are found

The model is Jev, TypeSafe's System One model (jev-1.13.0). Jev does not write text: it answers typed questions with probabilities. Each file is split into sentences and table rows. First, every stretch of two or three pages is screened for all the questions at once: which line, if any, answers each one. Then each promising line is checked on its own: does this line itself answer the question? The lines that pass become the answer, joined when they are neighbors, with a table row's label kept. The app copies them from the file, so nothing is paraphrased and nothing can be invented.

What it reads

PDF with a text layer, Word (.docx and .doc), Excel (.xlsx and .xls), PowerPoint (.pptx), OpenDocument text, spreadsheets and slides, RTF, saved web pages, CSV, TSV, plain text and Markdown, up to 26 files per study. A scanned PDF needs OCR first; the app warns when it finds almost no text.

Your files stay on your computer

Projects, studies, files, answers and checks are kept in your browser's storage on your device, and are never uploaded. To answer a question, the text of the files and the question go to TypeSafe through the site's relay; the files themselves do not. The site counts visits with Google Analytics, which sees neither your files, your questions nor the names of your projects.

Cost and limits

Using the site is free: questions go through a shared TypeSafe key with a daily limit, or through your own key, pasted in Settings. On the sample study, an 18-question form over three files took 27 requests, under five seconds and about a cent. The app counts requests and cost for the session, in all, and for each project, so the total for a review is there when you need it. Figures are images and are not read, though their captions are; English works best; a spreadsheet is read up to 5,000 rows.

Reporting it in your review

PRISMA 2020 asks reviews to describe how data were collected, including any automation tools. The exports record, for every answer, the model version, the date it was asked, and whether a reviewer checked it, and Write the methods paragraph, in the extraction table's Report tab, drafts the text from your project's own numbers. A sentence like this one can go in your methods, adapted to what you did:

Data were extracted with the help of Jev Reviewer (https://jevreviewer.xera.ac, model jev-1.13.0), which proposes verbatim quotes from each report with their location; every proposed answer was checked against the report by a reviewer, who recorded the extracted value.

Questions

Open source

Jev Reviewer is MIT licensed. The code, its tests and the notes on how it was measured are on GitHub, and it runs on your own computer with your own TypeSafe key.

Open Jev Reviewer