PlatyPeer
An iAIMS tool

Structured
peer review
that saves time.

You have a manuscript in one window and an hour between clinics. This instrument holds the method (line numbers, sections, severity, the AI checklist) so your attention stays on the science.

Start a review → See the method

Works in the browser. Readings stay on this device until you copy them out.

Example output · PDFScroll to read ↓
Peer-review reportMajor revision
Peer-review report

MS-2026-0417 · Chen et al. · AI checklist run

Overall impression

A promising deep-learning model for early detection of pancreatic ductal adenocarcinoma, evaluated retrospectively on a single-centre cohort. The clinical question is relevant and the reporting is generally clear, but the evaluation design does not yet support the clinical claims made in the abstract.

Major concernsMajor
Methods
L87Methods
Train and test split appears to be by study, not by patient. If any patient contributes to both, the reported performance is inflated.MAJ
L118Methods
No sample-size justification is given for the test cohort. With 61 positive cases, the confidence interval around the reported AUC is too wide to support the claim of non-inferiority to radiologist reading.MAJ
Results
L203Results
Point estimates are reported without confidence intervals throughout Table 2. Discrimination, calibration and clinical utility are three different claims; only the first is shown.MAJ
PlatyPeerPage 1 / 3
Major concernsMajor
Discussion
L256Discussion
The discussion moves from a retrospective AUC to a bedside-benefit claim without external validation. The clinical claim outruns what was tested.MAJ
Minor pointsMinor
Introduction
L31Introduction
Reference 12 is a preprint and is superseded by the 2024 published version.MIN
Methods
L142Methods
Scanner models are listed in the supplement but the protocol for harmonising acquisition parameters is not described.MIN
L155Methods
Define “early detection” in the Methods — the term is used three different ways across the manuscript.MIN
Results
L214Results
Figure 3 legend omits the error-bar definition. State whether bars are SD or 95% CI.MIN
AI checklist checks passed
T1Passed
Did they keep a proper held-out test set? — yes, patient-level split confirmed in supplement.
T7Passed
Did they analyse failures, subgroups, and fairness? — subgroup results by sex and age provided.
PlatyPeerPage 2 / 3
Recommendation
Major revisionSubstantive rework
PlatyPeerPage 3 / 3

Illustrative example · the report your readings build, filed as a PDF

Six stations. Two reads kept apart.

Get to a final peer review report with six easy steps, of which two are optional. We guide you through so you can focus on the science.

01

Invite

Ask yourself three easy questions to decide whether you should accept this peer review: expertise, conflicts of interest and time.

02

First read

Read the manuscript entirely and jot down anything that comes to mind. Don't worry about nice wording, minor/major classification, or whether you will include it in the review letter eventually.

03

Second read

Read the manuscript again after having detached from it (e.g. read it the next day). This phase is about filtering your initial notes, adding new ones, rewriting the ones you keep, and labelling them.

04

AI Checklist

Are you reviewing a paper involving AI in healthcare? This optional step contains a list of questions, based on EQUATOR checklists, that aims to filter out major flaws in AI methodology and reporting.

Optional
05

Report

View your entire report as it will appear as PDF. Make final adaptations, or choose for AI polishing in step 6.

06

AI polishing

If you would like to have your text refined with the help of AI, you can provide your own model key or have us generate a prompt for you to copy into your favourite chat app.

Optional
AI Glossary

A term you don't know never stops the reading.

Every step carries an AI Glossary button — a searchable list of AI and research-ethics terms, from algorithmic bias to TRIPOD+AI, one click away whenever a word in the paper (or in your own head) needs defining.

Generative AI: where we let it in.

AI can sharpen how you say what you decided. It must never decide for you. Here is how this tool guides a reviewer through that maze — and exactly where the barrier is drawn.

Bring your own key

Paste the API base URL and key for a model you already use — OpenAI-compatible or Gemini-compatible endpoints. The tool fetches that endpoint's own model list and you pick one.

Request goes machine → endpoint Key held in this browser, clearable in one click Output previewed beside your draft before anything is applied

Copy a ready prompt out

Generate a complete prompt with your finished report already embedded, fenced and ruled. Paste it into whichever assistant you prefer, then paste the result back to compare against your draft.

No key, no endpoint, no account Built locally and offline Same side-by-side comparison when you paste the result back

The barrier is drawn where the manuscript ends.

Peer review is conducted in confidence. Journals generally forbid sharing an unpublished manuscript with third parties, and that includes model providers. So this tool does not ask you to trust a privacy policy.

The polish step sees the draft report you would submit — nothing else. That report is built from your readings, so your own wording travels inside it. What does not travel is the manuscript: the tool has no field for manuscript text, so it cannot send what it has never held. The only manuscript-derived string in a request is the identifier you typed into the report header yourself.

If the journal's rules leave no room for even that, take the copy-out route instead. It builds the prompt locally and sends nothing at all.

Verify it yourself: the request is made from your own browser, and the developer tools will show you exactly what was in it.