ctcae.ai launches for site teams, with a de-identifier that runs in the browser

Week 38 of 2026 (September 14 to 20) launched ctcae.ai, a free site where research site teams look up any CTCAE term in v5.0 and v6.0 side by side and grade what they gathered in both versions. Its de-identifier checks a clinical note on the user's own computer, and the person reading it has the final say on what comes out.

The week also redesigned the CTCAE by Burna AI provider app around an Inbox, and Healthnotes questionnaire answers now file as one typed row per item. It serves coordinators, research nurses and investigators at trial sites, and the sponsors who receive what those sites record.

The ctcae.ai launch: a free CTCAE lookup for research site teams

ctcae.ai is a free site for coordinators, research nurses, data managers and everyone else who works adverse events at a trial site. It uses the same grading engine Burna AI builds for research sites and sponsors.

• Term search answers from the publications. A search that looks like a term returns that term from both CTCAE v5.0 and v6.0, read straight from the two vocabularies with no model involved, so the answer is the publication's own wording.
• Anything else starts a conversation, where a scenario comes back graded in both versions, one row per adverse event, with the criterion and the words each grade rests on.
• One account. Sign-in is an email and a one-time code, on the same account system as Burna AI and Protocol Studio, and it opens in a panel over the current page so nobody loses their place.
• Open to everyone who does the work. Monitors, sponsor staff and students are welcome. Every sign-up question is optional, and the role someone gives is never used to turn them away.

The site's copy describes the work as it happens: the coordinator researches the grade, the investigator reviews, modifies and signs, and every grade is a suggestion. In grading contexts the site asks for the safety information a coordinator gathered rather than for "the note", because the office visit note is the clinician's document.

A de-identifier that runs in the browser

ctcae.ai/de-identify finds identifiers in a clinical note on the user's own computer. The detection model downloads once; after that, the note is checked in the browser and is not sent anywhere while it is checked.

How it works:

1. Paste a note. Pattern rules mark fixed-format identifiers at once, and the clinical model reads the note again as soon as it has loaded.
2. Review each highlight. Every identifier is marked where it sits, with three choices: keep it in this note, remove it, or always keep it.
3. Choose the output. The clean copy uses placeholders such as [PATIENT_1] or realistic stand-ins. Stand-ins shift every date back by the same number of days so intervals survive, and a note that has to meet Safe Harbor exactly uses placeholders.
4. Grade the clean copy. Grade this note sends the cleaned version, never the original.

What stays in matters as much as what comes out. CTCAE v5.0 and v6.0 term names, NCT numbers and protocol codes stay in the note, and grades, lab values, blood pressure and cycle days stay unmarked, because a cleaned note still has to be gradable.

The detector is built around the 18 identifiers that HIPAA's Safe Harbor method lists in 45 CFR 164.514(b)(2). Three candidate models were compared on synthetic oncology research-visit notes, and the clinical model chosen runs beside pattern rules for fixed-format identifiers such as ZIP codes, Social Security numbers and vehicle identification numbers. The person reviewing the note has the final say on every highlight. Signed-in users can also drop a batch of .txt, .md or .docx notes and take back a clean copy of each.

The CTCAE by Burna AI app, redesigned around an Inbox

The CTCAE by Burna AI app, which clinicians and research coordinators use while they work adverse events at the visit, was rebuilt to a new design.

• An Inbox in three sections: safety flags, signatures waiting and handoffs. A section count shows when more items are waiting than the page holds, so open safety work is never understated.
• The longest-open safety flag stays in view, however many newer flags have been resolved since.
• A failed read shows as an error, never as an empty Inbox.
• New clinical components built to the same design: grade chips, the evidence span behind a grade, a respond-by clock for safety flags and an e-signature sheet.
• Signing out forgets the device, so a signed-out phone stops receiving pages.

Healthnotes questionnaire answers filed as typed rows

When a participant claims a PRO token in Healthnotes, the participant app sponsors deploy through sites, they are enrolled as a study subject with a subject number, and the questionnaires the trial's published schedule calls for are set up for them. A trial's PRO links deliver questionnaires once its assessment schedule is published.

Each answer is filed as its own typed row, keyed to the questionnaire item and the instrument version, so every answer carries its own record anchor. The site's review screen reads those rows back against the version the participant answered, and the data arrives at item level, the shape a sponsor's questionnaire dataset is built from.

Grading evaluations that cannot score a missing measurement

Burna AI's evaluation workspace, where the team runs grading agents against datasets and reads the evidence behind each run, gained tracing and a review inbox. Every measurement now takes one of four states: applicable, inapplicable, unmeasured or failed. A run that nothing could check records a lost measurement rather than a perfect score, so missing evidence can never read as quality.

Also in week 38

• Files opens on Knowledge documents, the reference library the grading and chat agents read from, with a Category filter beside Type.
• Capture controls in one place. On both capture surfaces, the contact-type and CTCAE-version pickers sit beneath the note, and Recent Activity shows an empty state rather than a blank panel.
• Trial headers fit their numbers. A long protocol number wraps inside its column and leaves the Enrolled count readable.

Since then: In week 39 ctcae.ai moved its session to a first-party cookie on ctcae.ai itself, and added dictation that transcribes a spoken note on the computer and hands it to the de-identifier before anything is sent.

AI suggests, clinicians decide. Human-in-the-loop, always.

If your site team wants the platform behind ctcae.ai, with grading, attribution and sign-off in one place, book a 15-minute conversation.