How Data Transformation Agent works
The following terms and concepts are essential for working with the AWS HealthLake Data Transformation Agent.
Transformation profile
A transformation profile is the reusable, versioned definition of how source data converts to FHIR R4. You create a profile once and reuse it across all datastores and jobs in your account.
-
For C-CDA, a profile contains Velocity templates.
-
For CSV, a profile contains a YAML mapping configuration.
-
A draft (version 0) is a mutable working copy you can edit freely.
-
Publishing creates an immutable, numbered version (v1, v2, ...).
-
Transformation jobs always use the profile's latest published version, so in-progress draft edits never affect in-progress transformations.
-
Rollback returns to any previous version by creating a new version that preserves the full audit trail.
-
Cloning creates a new profile from any existing version, giving you a separate profile to modify without affecting the original.
Transformation job
A transformation job is an execution that applies a published profile to your data.
Data Transformation Agent offers two run modes:
-
Sync (real-time): convert a single input and get FHIR back immediately. Used for testing a profile and for real-time, per-request conversions.
-
Bulk (asynchronous): convert large datasets from Amazon Amazon S3 at scale.
Sync conversions return the converted FHIR resources as a FHIR Bundle in the API response. Bulk transformation jobs return FHIR resources in one of two locations:
-
Amazon S3: write FHIR NDJSON to an Amazon Amazon S3 location.
-
HealthLake: convert and ingest FHIR resources directly into a HealthLake datastore.
Data Transformation AI agent
Data Transformation AI agent authors and edits a profile's conversion logic.
-
For C-CDA, it edits Velocity templates from natural language instructions.
-
For CSV, it analyzes sample files to produce a YAML configuration with column-to-FHIR field mappings, date and value translations, primary/foreign-key relationships, and aggregation rules, flagging anything that needs review.
You can interact with the agent through natural language including any of the following combinations: instructions, FHIR validation errors, schema documentation, and sample data. You can also edit profiles manually.
