Skip to main content
HEA-01 · Healthcare Providers

Clinical Documentation Copilot

Outcome​

Reduce clinician documentation burden by turning an ambient or dictated encounter into a structured, coded clinical note grounded in the patient chart, with clinician attestation required before anything enters the medical record.

Portfolio attributeValue
Primary personaClinician
Primary aggregateEncounterDocumentation
Cosmos DB containers1
Grounding dataencounters, EHR context, templates, terminology
Development estimateUSD 60-220 monthly dev estimate
Runtime optionsPython FastAPI and C# ASP.NET Core

Productivity hypothesis​

These measures reflect the workflow, evidence, decision boundary, and industry risk for this use case.

Impact typeMeasureConservative internal assumptionCustomer measurement
ProductivityEncounter-to-note approval time7-12% reduction(baseline median encounter-to-note minutes - assisted median minutes) / baseline median minutes * 100
Quality or riskCorrected note sections2-5% reduction(baseline corrected note-section rate - assisted rate) / baseline rate * 100

Important: Conservative internal pilot assumptions only; not public benchmarks, commitments, or expected customer results. Validate against a comparable baseline or matched control cohort with mandatory human review.

Rationale: Encounter context, templates, and terminology can modestly reduce documentation preparation in a safety-sensitive workflow.

Pilot guardrail: Clinicians remain responsible for clinical accuracy, documentation approval, and all care decisions.

Customer validation plan​

  1. Baseline the current workflow with representative users and tasks.
  2. Run the accelerator on the same task mix with human review retained.
  3. Measure encounter-to-note approval time and corrected note sections with the documented formulas.
  4. Replace internal assumptions with customer-observed results before a business case.

Why Azure Cosmos DB​

This AI solution needs an operational data plane, not only a vector index:

  • Hybrid vector and full-text retrieval combines semantic recall with exact identifiers, codes, names, and domain language.
  • Hierarchical partition keys and metadata predicates keep tenant and domain boundaries in the normal query path.
  • Operational JSON, source chunks, workflow and agent state, feedback, and audit records remain together for low-latency reads and governed writes.
  • Change feed reacts to source and operational updates without polling.
  • Autoscale throughput, configurable consistency, and multi-region distribution support variable demand and enterprise availability.

Domain workflow​

Domain workflow

StageActorInput → outputApproval
1. Capture the encounter conversationClinicianAmbient audio or dictation with patient consent → Diarized TranscriptSegment streamPolicy controlled
2. Assemble relevant chart contextChart context servicePatient identifier and encounter reason → ChartContextSnapshot of problems, meds, allergies, resultsPolicy controlled
3. Draft the structured clinical noteDocumentation agentTranscript segments and chart snapshot → NoteDraft with per-statement evidence linksPolicy controlled
4. Suggest diagnosis and procedure codingCoding agentNote assessment and plan sections → CodingSuggestion set with specificity warningsPolicy controlled
5. Attest and file to the medical recordClinicianEdited note and accepted codes → Signed note written to the EHR encounterHuman required

Technical architecture​

Technical architecture

Cosmos DB data model​

ContainerHierarchical partition keyDocumentsSearch paths
encounterDocumentations/tenantId, /encounterIdEncounterDocumentation, TranscriptSegment, ChartContextSnapshot, NoteDraft, CodingSuggestion/segmentVectorEmbedding, /noteVectorEmbedding, /segmentVector, /noteVector

Query contract​

QueryContainerScopePurpose
Encounter documentation assemblyencounterDocumentationssingle-partitionLoad transcript, chart snapshot, and current draft for the note editor in one partition read.
Unattested note backlogencounterDocumentationscross-partition-boundedBounded cross-partition control-plane query: show the clinician every draft still awaiting signature within the compliance window. It cannot carry the second partition key because a clinician's backlog spans all of the encounters they saw, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by one clinicianId and the 24-hour attestation compliance window, executed with MaxItemCount paging and continuation tokens.
Unsupported statement reviewencounterDocumentationssingle-partitionHighlight generated sentences that lack a transcript or chart evidence link.
Coding specificity queueencounterDocumentationscross-partition-boundedBounded cross-partition control-plane query: list suggested codes flagged for insufficient specificity before claim submission. It cannot carry the second partition key because the coding integrity queue is owned by the health information management team across encounters, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by proposed codes on encounters inside the 72-hour pre-bill window, executed with MaxItemCount paging and continuation tokens.

Python and C# services route each document type to its declared container and hierarchical partition key. ETag-based writes protect mutable aggregates, while append-only records and bounded cross-partition queries preserve scale.

Azure deployment​

Azure deployment

Clone the source repository, authenticate with Azure, and run:

Set-Location accelerators/healthcare-providers/clinical-documentation-copilot
az login
./Deploy.ps1 -Runtime python -Location eastus2

Change python to csharp to deploy the .NET implementation.

Explore the implementation​

Return to all accelerators