AI-native workflow and governance layer for healthcare verticals.
HDP takes clinical documentation from ambient capture through structured assessment and human-in-the-loop review to EMR write-back, with the consent, provenance, and audit primitives those workflows demand. Home health is the first vertical: OASIS-E assessments drafted by an LLM scribe, gated through confidence-aware review, and synced to EMRs — including the many that expose no API. It is a reference architecture under active development; no real patient data is processed anywhere in this repository, and all fixtures are synthetic.
HDP is not a FHIR clinical data repository, and doesn't compete with one. Pair it with a CDR such as Medplum or HAPI: HDP is the workflow that fills the CDR and the governance that controls what leaves it. The canonical record model treats FHIR R4 as a projection — one output among several (IPS, OMOP) — rather than the storage shape. The build-vs-pair evaluation is recorded in the FHIR surface research.
Three layers, dependencies pointing strictly downward:
+--------------------------------------------------------------+
| Apps composed deployables (home-health-scribe) |
+--------------------------------------------------------------+
| Verticals domain workflows: OASIS-E assessment, ambient |
| scribe, HITL review, EMR sync |
+--------------------------------------------------------------+
| Platform canonical records | consent | provenance | |
| primitives audit | identity | ingest | agent runtime |
+--------------------------------------------------------------+
FHIR R4 / IPS / OMOP are projections, not storage
The platform layer is the governance spine: canonical domain records carrying
trust level and provenance, grant-based consent with 42 CFR Part 2 fail-closed
category semantics, append-only audit, identity linkage, and an agent runtime
whose confidence scores feed the review queue. The vertical layer owns the
domain: OASIS-E modeled as first-class structured data, a
suggested → reviewed → accepted/edited/rejected review state machine with a
finalize gate, and a transport-agnostic EMR adapter contract — HTTP APIs where
they exist, browser-level automation where they don't. Schemas and semantics
are specified in the design corpus below; each package README labels its
implementation depth.
git clone https://github.com/peterkc/health-data-platform.git
cd health-data-platform
just up # Postgres 18 (pgvector) + Mongo 7; bootstraps .env on first run
just verify # lint + tests (CI mirror)just dev (the composed home-health-scribe app) lands with the
make-it-run epic.
The vault branch carries the design corpus — table designs (30-table DDL
plan, uuidv7 keys), the consent model, FHIR surface evaluation,
interoperability research, and ADRs:
browse the vault branch.