Stroke Decision Copilot

MAAI 901/902 group project · UOWD · 2026

Can a language model recommend stroke treatment and show its working?

Two vision models read a CT scan. A language model reads their numbers and a triage note, then decides on clot-busting drugs and clot removal against a rulebook written from the 2026 guideline. A verifier then checks every claim it makes against the scan's numbers, the note and the rulebook.

How it works, step by step

    Case explorer

    Every answer here was produced earlier with the same prompt and verifier. Click any claim to light up what it rests on: a number in the imaging contract, a line of the note, or a rulebook passage. Claims the verifier flagged are marked ✕ flagged with the reason.

    Pick a case on the left.

    What we found

    Hallucination rates are the verifier's, and the human check below shows they are upper bounds: people confirmed 18 of its 54 flags. The comparison between models holds.

    Try it: build a case

    Set the patient and the scan. The project's own note generator writes the triage note, then Claude Sonnet 5.5 answers live and the verifier checks it, exactly as in the evaluation. There is no free text: every value stays inside the ranges the system was tested on. If a live run cannot be served, you get the closest precomputed case and the reason.

    Timing
    Patient
    Imaging (what the vision models would report)
    Medicines and history

      About this project

      What it is not

      A research prototype built for a university course. It is not a medical device, has not been validated clinically, and must never be used to decide anyone's treatment. The rulebook is the team's reading of the guideline and has not been reviewed by a clinician.

      Limits we report

      • The "core" is a model trained on final infarct, not acute core, in a cohort that was reperfused.
      • Real cases have no ASPECTS and mostly an unknown occlusion site, so the rules leave many decisions undetermined.
      • The bleed detector flags about half of ISLES'24's presumed bleed-free scans: a domain shift, not an RSNA failure.
      • Hallucination rates are the verifier's; humans confirmed about a third of its flags.

      Data and credit

      • ISLES'24 (Zenodo 17652035), CC BY-NC-SA 4.0: Riedel et al., ISLES'24 – A Real-World Longitudinal Multimodal Stroke Dataset, arXiv:2408.11142; de la Rosa et al., arXiv:2408.10966. Cases are shown by alias only. This page is shared under the same licence.
      • RSNA Intracranial Haemorrhage (Kaggle): aggregate numbers only, no images. Flanders et al., Radiol Artif Intell 2020;2:3.
      • Guideline: 2026 AHA/ASA acute ischaemic stroke guideline, DAWN and DEFUSE-3. The rulebook is in our own words with pointers to the sources; no guideline text is reproduced.
      • Models: Claude Sonnet 5.5 and Haiku 5.5 (Anthropic), Qwen3.5-27B, Gemma 3 27B and 4B, MedGemma 27B; nnU-Net v2; ResNet-50.