Healthcare AI — Quick Reference#

Last Updated: 2026-06-30 Full Chapters: docs/07-Healthcare-AI/

Healthcare AI Categories#

FDA Regulatory Classification (Quick Reference)#

Non-device CDS (NOT SaMD): Clinician can independently verify the basis for the AI recommendation by reviewing the same information the AI used. AI organizes and displays information; clinician applies judgment.

Device CDS (SaMD): Clinician cannot independently verify the basis — they are relying on the software's analysis. Model produces a risk score or recommendation from complex inputs the clinician cannot re-derive manually.

Key question: "Can the clinician look at the same data and derive the same conclusion without the AI's model?"

• Yes → Likely non-device CDS
• No → Likely SaMD requiring FDA regulatory engagement

For LLM-based SaMD: File a Predetermined Change Control Plan (PCCP) at 510(k) submission to enable model version updates without repeated submissions.

HIPAA Quick Reference for Clinical AI#

18 HIPAA identifiers to remove for Safe Harbor de-identification: Names, geographic data (<state level), dates (except year), phone, fax, email, SSN, MRN, health plan ID, account number, certificate/license, vehicle ID, device ID, URLs, IPs, biometrics, photos, any other unique identifier.

FHIR R4 — Key Resources for Clinical AI#

SMART on FHIR scope pattern: {context}/{resource}.{permission} Example: patient/Condition.read patient/DocumentReference.write

CDS Hooks — Standard Hooks#

CDS response time requirement: < 5 seconds (EHR timeout). LLM inference only if necessary.

Circuit breaker rule: If the CDS service is unavailable, the EHR workflow MUST proceed. Clinical workflow cannot depend on AI availability.

Clinical RAG — Key Design Decisions#

Medical ontologies for query expansion:

• SNOMED CT: clinical concepts, diagnoses
• RxNorm: medication standardization
• LOINC: lab tests and clinical observations
• ICD-10-CM: diagnosis codes

Clinical Documentation AI — Quality Metrics#

Mandatory rule: AI-generated content must NEVER be written to the medical record without explicit physician review and approval. The physician's signature attests to accuracy and completeness.

Clinical Decision Support — Alert Design#

Alert fatigue thresholds:

• Override rate > 70% → Alert specificity problem
• Override rate > 85% → Clinical staff has learned to dismiss without reading
• Alert volume > 10 per provider per shift → Fatigue risk

Medical Imaging AI — Quick Reference#

Patient Engagement AI — Design Principles#

Clinical AI Safety — Pre-Deployment Checklist#

• [ ] Risk tier classified (Tier 1 / 2 / 3)
• [ ] FDA SaMD status determined; regulatory counsel engaged if applicable
• [ ] HIPAA BAA signed with all AI vendors
• [ ] Demographic bias evaluation: no subgroup > 10 points below overall accuracy
• [ ] Golden dataset evaluation: ≥ 50 cases, meets quality threshold
• [ ] Adversarial testing: prompt injection, edge cases
• [ ] Clinical panel review: qualified reviewers approved output quality
• [ ] FMEA completed: failure modes documented with mitigations
• [ ] Rollback procedure defined and tested
• [ ] Safety event reporting channel established (low-friction)
• [ ] Model Review Board approval obtained

HMS AI Platform — Build Sequence#

Phase A — Foundation (before use case 1):
  AI Gateway → Prompt Registry → Model Registry → Audit Log → BAAs

Phase B — First Use Case (Discharge Summary AI):
  SMART on FHIR registration → FHIR client → SMART app → Eval pipeline → MRB

Phase C — Platform Expansion (before use case 3):
  Embedding Service → Clinical Vector Store → Integration Engine →
  CDS Hooks → Observability stack

Phase D — Scale (use cases 4–7):
  Prior Auth agent → Medical Coding batch → Care Gap → Patient Chatbot →
  Full champion network

Interview Quick Reference#

Healthcare AI Landscape:

• FDA SaMD classification: key question is whether clinician can independently verify the AI's reasoning basis
• PCCP at 510(k) submission = mechanism to update LLM model versions without repeated submissions

HIPAA and AI:

• BAA required before any PHI is transmitted to any external AI vendor
• Observability traces must scrub PHI — metadata only (hashed IDs, token counts, latency)
• Vector indexes of clinical notes = PHI data stores requiring same controls as source EHR data

EHR Integration:

• FHIR R4 for data retrieval; HL7 v2 ADT for event notification
• SMART on FHIR: minimum necessary scopes; EHR launch passes patient context automatically
• CDS Hooks: must respond < 5 seconds; circuit breaker — clinical workflow cannot depend on AI

Clinical RAG:

• Chunk at recommendation boundaries, not character counts
• Index metadata (organization, date, evidence grade) is required for citations and source weighting
• Index must be updated on a defined schedule — stale clinical knowledge is a patient safety risk

Clinical Documentation AI:

• AI draft must be reviewed and approved by physician before entering medical record
• Physician edit rate 10–25% = healthy engagement; < 5% = investigate rubber-stamping

Medical Imaging AI:

• 510(k) clearance verification: patient population + modality + indication must match deployment
• Deliver AI overlay in radiologist's standard PACS view without extra workflow steps

Patient Engagement AI:

• Escalation router is the safety-critical component: fail-safe, 24/7 human coverage
• SMS channel is most equitable: works on basic phone, no app required

Clinical AI Safety:

• Demographic bias evaluation: no subgroup > 10 points below overall accuracy
• FMEA before deployment: identify failure modes and mitigations
• Override rate > 70%: alert fatigue; < 5%: rubber-stamping risk

See Also#

• Healthcare AI Landscape
• HIPAA and AI
• EHR Integration
• Clinical RAG
• Clinical Decision Support
• HMS Reference Architecture
• Medical Imaging AI
• Patient Engagement AI
• Clinical Documentation AI
• AI Safety in Clinical Settings