Glossary¶

Every term used across the site, one line each.

Architecture¶

Term	Meaning
Medallion	A design with three quality tiers — bronze (raw), silver (cleaned/conformed), gold (consumption-ready). A lifecycle, not a schema.
Bronze	Raw ingested data, as-is; schema drift tolerated; PHI present & access-gated.
Silver	De-identified, type-clean, standardized. The PHI boundary — everything downstream reads only from here.
Gold	Consumption-shaped outputs: analytics marts, feature store, vector index.
Delta / Iceberg	Open table formats giving object storage ACID transactions, schema evolution, time travel.
Unity Catalog	Databricks governance layer: catalogs/schemas, table/column access, automatic lineage.
Lineage	The dependency graph from sources → models → outputs (auto-generated from dbt here).

Term	Meaning
FHIR	HL7's standard for clinical data as JSON Resources (Patient, Encounter, Condition, Observation…). Often exported as NDJSON.
OMOP CDM	OHDSI's Common Data Model — one standard format + vocabulary so disparate sources analyze the same way.
Concept id	OMOP's standard integer code for a clinical concept (e.g. gender 8507 = male); loaded from OHDSI Athena.
ICD-10 / LOINC / SNOMED / RxNorm	Vocabularies for diagnoses / labs & observations / clinical findings / medications.
PHI	Protected Health Information — names, SSN, address, full dates, etc.
Safe Harbor	HIPAA de-identification route: remove 18 specified identifier types.
Date-shift	Shift each patient's dates by a fixed per-patient offset — preserves intervals while de-identifying.
patient_key	Salted-hash surrogate replacing the real patient id.

Term	Meaning
HL7	Health Level Seven International — the standards body that publishes FHIR (and older HL7 v2 messaging).
HIPAA	US health-privacy law. Defines PHI and the two de-identification routes.
Expert Determination	HIPAA de-id route #2: a qualified statistician certifies re-identification risk is "very small".
De-identification	Removing PHI so data can be used for analytics/ML. In Vitals: done at the silver boundary.
OHDSI ("Odyssey")	Observational Health Data Sciences and Informatics — the open community maintaining OMOP, its tools, and vocabularies.
Athena	OHDSI's vocabulary repository — where standard OMOP concept_ids are downloaded from.
Interoperability	Different systems exchanging data they can both use — FHIR's purpose.
NDJSON	Newline-delimited JSON — one object per line; the FHIR bulk-export format.
EHR	Electronic Health Record — the clinical source system.
Claims (837 / 835)	Billing data: 837 = claim submitted to the payer; 835 = the payer's remittance/payment.
CPT	Current Procedural Terminology — procedure codes in claims (MRI, physical therapy, injection…).
PRO	Patient-Reported Outcome — a survey the patient fills in (e.g. pain, disability).
ODI	Oswestry Disability Index — the PRO instrument used here (0–100; higher = worse).
Synthea	Open-source synthetic FHIR patient generator — PHI-free data the project can publish.
Value-based care	Paying for outcomes, not volume — the cost mart reflects it.

Term	Meaning
dbt	SQL transformation framework: versioned models, tests, docs, lineage. Builds silver→gold here.
Dimensional model	Kimball star schema: `dim_` (entities) + `fct_` (events/measurements).
Semantic layer / metric	Governed, reusable metric definitions analysts consume (e.g. surgery rate).
Feature store	Serves ML features consistently for training (offline) and scoring (online). Here: Feast.
Vector index	Embeddings of text for similarity search / RAG. Here: TF-IDF (pgvector in prod).
RAG	Retrieval-Augmented Generation: fetch relevant text, feed it to an LLM.
MLflow	Experiment tracking + model registry.

Term	Meaning
ROC-AUC	Ranking quality of a classifier (0.5 = random, 1.0 = perfect).
Label leakage	A feature that encodes the outcome — inflates offline scores, fails in production. Avoided by using only conservative-care claims.
Feature selection	Using a curated subset of available features in a model (store stays rich, model stays clean).
PSI	Population Stability Index — drift metric per feature: <0.1 stable, 0.1–0.2 moderate, >0.2 significant.
Structured Streaming	Spark's streaming engine; same API as batch. Uses trigger, checkpoint, watermark.
Window function	Per-partition ordered computation (e.g. 7-row rolling mean per patient).

Term	Meaning
DuckDB	In-process single-file analytical DB. Runs the MVP with no cloud (Databricks is the deploy target).
mise / uv	Pin tool versions / manage Python deps fast.