WHAT WE DO

Cancer research, explained to the person it is about

Peer-reviewed oncology turned into plain language, with every sentence traceable to the paper it came from

A person diagnosed with cancer is handed a body of research about their own illness that was written for somebody else. The evidence exists, it is public, and it is largely unreadable to the person it concerns most.

We built a system that takes a peer-reviewed oncology paper and produces an accurate plain-language version of it in under a minute, with every sentence linked back to the passage that supports it. It was developed with SundVision and validated at the University of Southern Denmark.

THE PROBLEM

The research is public. The understanding is not.

Open access solved the availability of cancer research. It did not solve whether anybody outside the field can read it. For a patient trying to understand a treatment decision, a freely available paper they cannot follow is not much better than one behind a paywall.

Written for other researchers
The precision that makes a paper useful to a specialist is the same precision that makes it impenetrable to the person whose treatment it describes. Neither audience is wrong to want what it wants, but only one of them is served.
Patients cannot act on what they cannot follow
Published research has repeatedly found that only a minority of cancer patients can correctly state their own tumour type or stage. That is not a failure of intelligence. It is a failure of translation, and it shows up later as a patient who cannot take part in a decision about their own treatment.
The requirement arrived before the capacity
European rules now require a plain-language summary for every clinical trial. Written by hand, each one takes hours of specialist time, and the departments that owe the most of them are the ones with the least time to spare.

WHAT IT DOES

From a peer-reviewed paper to something a patient can use

Give it an oncology paper and it returns a plain-language version in under a minute, written at a reading level you choose rather than one it guesses at. The same paper can be explained to a newly diagnosed patient and to a specialist nurse, and both versions stay true to the source.

Take a phrase like PD-L1 expression. In the paper it is a biomarker with a threshold. For the patient it becomes a marker showing who is likely to benefit more from immunotherapy, which is the form in which it can actually inform a conversation about treatment.

A plain-language version of the paper, at a stated reading level, in under a minute.
Every sentence keeps its citation, pointing back to the passage in the source that supports it.
Output where it is needed: a patient letter, the European summary template, or a chat the patient can ask follow-up questions in.
A draft for a professional to approve, not something that publishes itself.

WHY IT CAN BE TRUSTED

An explanation you can check against the paper

Simplifying medical text is the easy half. Simplifying it without quietly changing what it says is the hard half, and it is where a general-purpose chatbot fails. Four things separate a simplification from a distortion.

Citation at the level of the sentence
Each statement in the output is matched back to the sentence in the source that supports it. Coverage is measured on every run rather than assumed, so a reader can go from any claim to the text it rests on.
A second pass before anyone sees it
Every draft is checked against the source a second time. Sentences that cannot be supported are removed rather than hedged, because a hedged unsupported claim is still an unsupported claim.
It says when it does not know
A question the corpus does not answer receives that answer. The failure mode we design against is not silence, it is a confident paragraph about something the evidence does not cover.
Tested by people trying to break it
Validation included adversarial prompts written specifically to push the system into inventing findings, alongside review of the ordinary output by clinicians.

THE EVIDENCE

We are testing with users, not only on documents

Validation was run with the University of Southern Denmark. Fifty users read about cancer type and then answered a comprehension test. Half read explanations produced by the system, half read the standard written information. Two hundred oncology papers were reviewed alongside the patient study.

Patients who read the generated explanations scored around a third higher than they had at the start, an improvement several times larger than the group given the standard information.

From the 2026 validation. Results describe that study and are not a guarantee of the same effect in a different population.

WHERE IT RUNS

The data does not leave the hospital

On-premise or in a Danish cloud. It can run inside the hospital or research group’s own infrastructure, or in a Danish environment certified for health data. Which one is your choice, not a constraint of the product.
No patient data is needed. The input is a published paper. Producing an explanation does not require anybody’s record.
Reproducible. The source collection is versioned, so an explanation produced today can be regenerated and checked later against exactly the material it was drawn from.

Bring us a paper

The most useful demo is one run on a paper you already know well, because you can check the output against your own understanding of it. Send us one from your own field and we will show you what comes back, citations and all.

You can also see the patient-facing side of this work at Cancer Information, where the same approach answers questions about cancer in ordinary language.

Book a demo

Visit Cancer Information