How Kepler killed the circle in Prague
he trusted observation over eleganceWhat would make us revise the framework behind an answer?
Read the original LinkedIn postFrançais
AI in clinical practice
When AI enters the stroke decision, what changes—and how should we judge whether it helps?
The problem
Acute-stroke decisions already bring together clinical findings, imaging, evidence and the practical limits of a care network. An AI output adds another interpretation. What should the clinician check when it agrees with their judgment—and when it does not?
Reproducibility makes an answer easier to examine; it does not establish that the answer is valid. Useful scrutiny asks which evidence was selected, whether the patient resembles those studied, what remains uncertain and who takes responsibility for the decision.
Different judgments may reflect missing evidence, different circumstances or different goals. Disagreement alone does not establish decision noise or bias.
Examples: one referral, two decisions
Hypothetical educational scenario, adapted from the ALICE working case. These staged facts are discussion prompts, not a treatment recommendation, a model evaluation or study data.
23:00 — the first call
An 80-year-old man at a referring centre has NIHSS 4, a left M2 occlusion and ASPECTS 10, six hours after onset. Transfer takes two hours.
Would you consider immediate thrombectomy? Separately, where should reassessment and possible rescue take place? What fact would you need before answering either question?
The exact deficit and disabling status, M2 anatomy, thrombolysis status, baseline function and local monitoring resources are not specified. An AI answer must not silently fill those gaps.
02:00 — the information changes
In the constructed scenario, the deficit worsens to NIHSS 20 with persistent M2 occlusion. Transfer is accepted.
Which new clinical facts justify revisiting the earlier decision? What should be carried forward from the original reasoning? The trajectory does not prove that earlier transfer or treatment would have prevented deterioration.
What would you ask of an AI answer?
- Identify the clinical question and keep treatment and transfer distinct.
- Show the sources, their dates and the populations studied.
- Separate evidence from assumptions and missing information.
- Explain what would change the interpretation.
This is a discussion checklist. No AI output or comparison result is presented here.
Published posts in the author’s English; the selected labels and discussion questions are editorial drafts.
AI and the difficult referral call
the question becomes what you hand the physician who isWhat assistance would be useful when specialist advice is not immediately available?
Read the original LinkedIn postFrom imaging output to clinical reasoning
How do units standardize the human-AI handoffHow should an AI output be checked and acted on across different stroke services?
Read the original LinkedIn postNEuroAIS — AI use and views before ALICE
NEuroAIS is an invitation-based descriptive survey of ALICE attendees who make acute stroke-care decisions without supervision, at least sometimes. This includes physicians in training who sometimes decide independently.
It asks about overall views of AI, reported clinical use, resources consulted for difficult acute decisions, source checking, trust, perceived risks and training needs. Responses describe practices and attitudes in a conference-based convenience sample; they do not measure actual decision noise, bias reduction, clinical effectiveness or patient benefit.
Study status checked 2 November 2026: the first invitation round is complete. No findings are presented here. Any future summary requires completed analysis and review.
Read the public registration and protocol · Protocol v0.11; questionnaire v0.6.
This page provides study information. It does not enrol participants or collect survey responses.
Share your thoughts
What would make an AI answer more useful—or easier to challenge—in a difficult stroke decision?
The intended route is private editorial review, with possible follow-up. A website comment is not a NEuroAIS response, is not automatically published and does not request clinical advice. Please omit patient details.
Share a private thought