Build safe, effective clinical AI with evidence, governance and interoperable code.
Scope: Prototypes → pilots → regulated products.
Non-negotiables: No patient-identifiable data; no clinical claims without evidence; respect regional law.
Tag every thread:
Family = [LLM | GPT-based App | GenAI | Agency] · Evidence = E0–E5 · Governance = G0–G4 · Region = [NA | SA | EU | KSA | JP | AUS]
Lifecycle:
Idea → Research → Medical Guidelines → DPIA & Patient rights → Big Data & APIs (ORCID for provenance) → Training →Clinic/Praxis Test (IRB) → Controlling→ Publishing → Post-market & Updates.
House Rules (must follow)
1. No PHI/PII. Use de-identified or synthetic data only.
2. Regulatory truth-in-labeling. Tag your thread stage: Idea, Prototype — Not for clinical use, IRB Pilot, Regulatory-cleared/approved.
3. Intended Use first. One sentence: user · condition · decision · setting.
4. Evidence or it didn’t happen. Datasets, external validation, confidence intervals, and documented failure modes required for E2+.
5. Human-in-the-Loop. Define who can override and the alert workflow.
6. Security. Prefer on-prem/VPC; include diagrams; log/audit every inference used in care.
7. Fairness. Provide subgroup metrics and mitigation strategies.
8. Change Control. Model registry, rollback plan, drift monitoring (PCCP / AI-Act style).
9. No promo spam. Disclose funding and conflicts of interest.
10. Follow local law & platform policy. Posts risking patient harm will be removed.
Standard Post Template (copy & fill for any project)
Title
Family / Evidence / Governance / Region
Intended Use & clinical risk (why it matters
Regulatory path (MDR/FDA/SFDA/Japan PMDA/AUS TGA) & current stage
Model family & components (LLM, vision, rules, retrieval)
Data (source, size, labeling, de-ID method)
Validation (internal, external, prospective; metrics & calibration)
Workflow integration (FHIR/DICOM, EHR/PACS, roles, alerting)
Privacy & lawful basis (HIPAA/GDPR/PDPL/JP law; DPIA/IRB)
Safety & oversight (guardrails, escalation, audit trails)
Change plan (PCCP-style updates, drift monitoring, rollback)
Limitations & open risks
Repo/diagram/demo (no real patient data)