Generative AI in Healthcare Training Course

Participants can design controlled prompts and output schemas for bounded healthcare tasks.

• System, user and developer prompts separated for controlled clinical tasks
• Few-shot prompting with approved exemplars and explicit output constraints
• Chain-of-verification prompts for checking claims, omissions and unsupported citations
• Structured output schemas for referrals, discharge summaries and coding queries
• Prompt injection defences using delimiters, instructions hierarchy and input validation
• Prompt libraries governed through versioning, ownership and clinical approval

Participants can specify a retrieval-augmented generation design using controlled healthcare data sources.

• FHIR resource mapping for patient, encounter, observation and medication data
• Retrieval-augmented generation architecture with source ranking and citation display
• Chunking strategies balancing clinical context, retrieval precision and update frequency
• Embedding evaluation using recall@k, precision@k and clinically relevant documents
• De-identification workflows applying HIPAA Safe Harbor and expert determination
• Access controls enforcing minimum necessary data and tenant separation

Participants can create an evaluation plan for safety, accuracy and clinical usefulness.

• Reference answer sets built from guidelines, policies and verified records
• Test cases stratified by acuity, demographic subgroup and clinical specialty
• Hallucination scoring for fabricated facts, citations and clinical recommendations
• Clinician adjudication protocols defining acceptable, reviewable and unsafe responses
• Calibration review comparing model confidence, evidence strength and reviewer certainty
• Release thresholds documented through acceptance criteria and residual-risk decisions

Participants can apply structured governance controls to healthcare generative AI risks.

• NIST AI RMF functions mapped to healthcare generative AI controls
• Risk registers recording harm scenarios, owners, mitigations and review dates
• Bias assessment across language, disability, race, sex and socioeconomic proxies
• Human-in-the-loop escalation rules for clinical, administrative and patient-facing outputs
• Incident response playbooks covering harmful output, data exposure and service failure
• Model cards documenting intended use, limitations, evaluation and monitoring

Participants can define privacy, regulatory and contractual requirements for AI deployments.

• Privacy impact assessments identifying purpose, data flows, retention and disclosures
• HIPAA Privacy Rule requirements for permitted uses and minimum necessary access
• GDPR roles, lawful bases and automated-decision safeguards for health data
• FDA software-function policy distinctions for clinical decision support
• Good Machine Learning Practice principles applied to intended-use documentation
• Data processing agreements specifying vendors, subprocessors, breach duties and audits

Participants can redesign healthcare workflows with appropriate human review and escalation points.

• Workflow mapping identifies handoffs, exceptions, documentation burden and decision rights
• Ambient documentation design sets clinician review, attestation and correction steps
• Prior-authorisation drafting boundaries separate evidence retrieval from payer determinations
• Patient-message triage routes symptoms, urgency and language needs safely
• Contact-centre assistants use approved knowledge, escalation and disclosure scripts
• Failure-mode-and-effects analysis prioritises workflow controls before release into production

Participants can define security and permission boundaries for agentic healthcare AI workflows.

• Agentic workflow boundaries restrict actions, tools, permissions and spending authority
• Tool calling schemas validate parameters before scheduling or record updates
• Least privilege design separates retrieval, drafting, ordering and approval permissions
• OWASP LLM Top 10 threats assessed for healthcare assistants
• Red-team scenarios probe prompt injection, data exfiltration and unsafe actions
• Audit logs capture prompts, sources, tool calls, reviewers and outcomes

Participants can assess vendor options and establish ongoing model monitoring responsibilities.

• Build-versus-buy analysis compares model access, data control, support and cost
• Vendor due diligence reviews security evidence, data terms and subcontractors
• Service-level objectives define latency, availability, recovery and escalation targets
• Cost models estimate tokens, retrieval, integration, review and monitoring
• Model monitoring tracks quality drift, refusal rates and harmful-output signals
• Change control governs prompt updates, model versions and knowledge refreshes

Participants can prepare an accountable implementation plan for a prioritised generative AI use case.

• Use-case prioritisation matrix scores value, feasibility, risk and readiness
• Stakeholder maps assign clinical, privacy, security and operational accountabilities
• Pilot charters define population, workflow, success measures and stop criteria
• Implementation roadmaps sequence data, integration, assurance and adoption work
• Training plans prepare clinicians for verification, escalation and appropriate reliance
• Personal application plans specify first use case, sponsors and next actions