How we verify answers

SmartPrescribe Ask AI retrieves and cites authoritative sources before generating educational content for healthcare professionals. This page summarises the quality controls that support our Medical Affairs governance.

1. Source ingestion

We maintain structured clinical knowledge from official and society sources, including:

• INFARMED RCM (Portugal) — dosing, contraindications, interactions
• DGS normas & guidelines (Portugal)
• EMA medicines overview and SmPC-derived content
• Curated drug–drug interaction datasets with severity, mechanism, and management fields
• Society guidelines — ESC, ACC/AHA, NICE, KDIGO (verified URL registry)

Chunks are tagged with source, language, publication or verification date, and content type for retrieval routing (not generic full-text search alone).

2. Citation verification

• Reference lists require verifiable identifiers (PMID and/or DOI) where applicable — not bare “Guidelines 2024” strings.
• Society guideline links are drawn from a curated allow-list, not model-generated URLs.
• SmPC citations use jurisdiction-appropriate labels (RCM INFARMED, EMA, FDA) shown in the Sources panel by clinician location.
• Low-evidence publication types (case reports, letters, editorials) are flagged when surfaced from literature search.

3. Pre-publication QA (interaction cards)

High-traffic drug–drug interaction outputs (e.g. antiplatelet therapy) pass automated checks before display, including:

• Complete citation format (author, title, journal, year)
• Guideline recency (e.g. ESC ACS 2023+, not superseded primary sources)
• Class and level of evidence on recommendations
• Numeric claims tied to named trials or meta-analyses
• SmPC link visibility rules by country
• Single educational disclaimer in UI (not duplicated in model prose)

4. Guideline drift monitoring

Ingested guideline content is periodically audited against source publication dates. Material flagged as potentially outdated triggers admin review and re-ingestion. This reduces the risk of citing superseded recommendations in educational outputs.

5. What we do not claim

• AI output is educational — not a substitute for clinical judgment or official prescribing information.
• We do not guarantee completeness of all drug–drug pairs or all global formularies.
• Generative text may still contain errors; clinicians must verify against local SmPC and protocols.
• In-game tutor chat (short post-case follow-ups) uses a lighter model path without full RAG; full Ask AI uses the complete verification stack.