Ein synthetischer Anamnesebogen wird in eine strukturierte Zusammenfassung überführt: Befundstruktur, Kontraindikationen, offene Punkte, jede Aussage mit sichtbarer Quelle im Bogen. Erzeugen Sie neue synthetische Fälle und prüfen Sie die Nachvollziehbarkeit. Keine Patientendaten, keine Diagnose.
A synthetic intake questionnaire is turned into a structured summary: findings structure, contraindications, open points, every statement with a visible source in the form. Generate new synthetic cases and check traceability. No patient data, no diagnosis.
Aesthetic dermatology · outpatient intake questionnaire · structured extraction with a visible source for every statement · maturity phase 2 (assisted)
Clipboard access was not available. The summary is selected below — copy it manually (Ctrl/Cmd + C).
The form is read field by field. Explicit answers — yes / no / unsure, checkboxes, dates — become statements marked certain · form field.
Free text (reason for visit, medications, allergies, expectations) runs through a small lexicon of regular expressions: drug names and brands,
concern vocabulary, expectation cues. Every hit keeps its character offsets, which is what makes the highlight on the left possible, and is labelled
inferred from text together with the phrase that fired. A rule table (– rules, excerpt below) combines field answers and text hits
into contraindications with a severity — info, caution, defer — and notes when a rule touches the treatment the patient actually asked for.
A consistency layer compares the two channels: an anticoagulant named in the medication list but denied in the yes/no question becomes an open question,
not a silent assumption. Nothing here is a diagnosis; the output is a reading aid for the practitioner before the consultation.
| Trigger | Severity | Statement |
|---|
Free-text normalisation is the weak point of a lexicon: misspellings (“Apixiban”), brand names from other countries, mixed languages, doses written as sentences, negations (“stopped Marcumar in 2023”). A language model is good at mapping such text onto a controlled vocabulary (ATC code, SNOMED concept) — as a normaliser that proposes candidates with their source span, which the same deterministic rule table then evaluates. Rules, severities and the consent logic stay auditable and versioned; the model does not decide, and a candidate without a traceable span is discarded.
In a clinical context a summary is only useful if the practitioner can verify each line in seconds. A statement without a pointer to the form field or the phrase it came from has to be re-read in full — which removes the time saving and adds a new risk: trusting a line that was inferred rather than stated. Hence every statement carries its origin and a certainty label, and anything ambiguous is turned into a question instead of a conclusion. It also makes the system reviewable: a wrong extraction can be traced to a rule or a lexicon entry and fixed, rather than re-prompted.
Not a medical device, no medical advice — demo of structured extraction with source tracing. Everything is computed locally in your browser; nothing is sent anywhere and nothing is stored. Patients are generated from templates with a seeded random generator; the seed is kept in the URL so a shared link shows the same case. Rule thresholds (6–12 months after isotretinoin, 2–4 weeks after sun exposure, etc.) reflect common practice conventions, not a clinical guideline.