Add schema.sql

schema.sql

@@ -0,0 +1,231 @@
+-- FHIR-SQL fine-tuning study: frozen core clinical schema.
+--
+-- This is the benchmark-facing schema shown to models in every prompt (benchmark
+-- authoring, SFT prompt format, RL reward execution). It is a deliberately curated
+-- subset of the full flattened data -- see METHODOLOGY_LOG.md for the two-layer
+-- rationale (full fidelity in the database, curated scope in what models see) and
+-- the token-cost/scope reasoning.
+--
+-- Generated from the actual column types DuckDB inferred when loading
+-- data/train.duckdb, not hand-assumed -- see scripts/flatten_to_duckdb.py for
+-- the extraction logic that produces this shape.
+--
+-- Version stamp:
+--   Date frozen:        2026-08-02
+--   Synthea build:      v3.4.0-18-ga07a65555 (git-describe string embedded in
+--                        generated Patient resources; downloaded from the
+--                        GitHub v4.0.0 release page -- see methodology log)
+--   Train population:   18,999 patients (target was ~15,000; see log for
+--                        per-batch seed/state/age-bracket design)
+--   Held-out population: 6,383 patients (target was ~5,000)
+--   Populations verified disjoint: 0 patient_id overlap between train and held-out
+--
+-- Every number produced downstream (benchmark accuracy, cost tables, etc.) is
+-- relative to this artifact. Do not modify this file without a note on what
+-- changed and why.
+--
+-- Provider attribution: Synthea generates exactly one participant per encounter,
+-- typed "primary performer" only -- it does not distinguish admitting/attending/
+-- consulting roles, so those remain unanswerable regardless of schema design.
+-- Procedure.performer is never populated by this Synthea version (0/29,947 in a
+-- full batch) -- procedure-level provider attribution is not available at all.
+--
+-- Naming convention: columns are named to match FHIR element names directly,
+-- so a model's clinical-language understanding maps onto the schema with as
+-- little translation as possible:
+--   - Primary key: `id` (matches every FHIR resource's own `id` element).
+--   - Foreign keys: `patient_id`, `encounter_id` (SQL join-key convention;
+--     not itself a literal FHIR field name, since FHIR expresses this via
+--     subject/patient/encounter *reference* elements, but resolving those
+--     references to a flat join key needs a name, and `<type>_id` is the
+--     clearest SQL-side compromise).
+--   - Primary coding triple on each table: `code`, `system`, `display`
+--     (matches FHIR Coding.code/.system/.display exactly).
+--   - Where a resource's own field name differs from the generic "code"
+--     (Encounter.type, Encounter.class, Immunization.vaccineCode,
+--     CarePlan.category), the coding triple is prefixed with that field name
+--     instead: `type_code/type_system/type_display`, `class_code`,
+--     `vaccineCode/vaccineCode_system/vaccineCode_display`,
+--     `category_code/category_system/category_display`.
+--   - Status/descriptive fields: exact camelCase FHIR element names
+--     (clinicalStatus, verificationStatus, intent, criticality).
+--   - Dates: exact FHIR element names (birthDate, deceasedDateTime,
+--     onsetDateTime, abatementDateTime, recordedDate, effectiveDateTime,
+--     authoredOn, performedDateTime, occurrenceDateTime); Period-typed
+--     start/end kept as `period_start`/`period_end` (Period.start/.end).
+--   - value[x]: `valueQuantity`, `unit` (Quantity.unit), `valueCodeableConcept`
+--     (+ `valueCodeableConcept_system`), `valueString`.
+--   - Provider-reference columns match the FHIR field they were extracted
+--     from: `requester_npi`/`requester_name` on medication_request (from
+--     MedicationRequest.requester), `participant_npi`/`participant_name` on
+--     encounter (from Encounter.participant).
+--   - `race`/`ethnicity` on patient (US-Core extensions -- see
+--     scripts/flatten_to_duckdb.py's us_core_ext_text macro).
+--   - `imaging_study` (ImagingStudy resource) makes radiology-volume questions
+--     answerable (department-level radiology questions remain unanswerable --
+--     no department/service-line concept exists anywhere in Synthea's FHIR
+--     output).
+--
+-- Indexes: secondary (ART) indexes are added directly to train.duckdb and
+-- heldout.duckdb (not a change to this file -- no column/table/logical change,
+-- only a physical one) on the coding-triple columns (condition.code,
+-- observation.code, medication_request.code, encounter.class_code,
+-- encounter.type_code, procedure.code, immunization.vaccineCode, allergy.code,
+-- careplan.category_code, diagnostic_report.code, imaging_study.procedureCode,
+-- imaging_study.modality), to support the RL execution-efficiency reward term.
+-- patient_id/encounter_id deliberately NOT indexed -- DuckDB's ART index isn't
+-- used by the optimizer to accelerate joins, only point/highly-selective
+-- (<0.1% of rows) filters. See METHODOLOGY_LOG.md for the full verification
+-- history of this schema and its indexes.
+
+CREATE TABLE patient (
+    id                    VARCHAR PRIMARY KEY,
+    gender                VARCHAR,
+    birthDate             DATE,
+    deceasedDateTime      TIMESTAMP,
+    maritalStatus         VARCHAR,
+    state                 VARCHAR,     -- address[0].state
+    city                  VARCHAR,     -- address[0].city
+    postalCode            VARCHAR,     -- address[0].postalCode
+    race                  VARCHAR,     -- US-Core race extension, ombCategory text
+    ethnicity             VARCHAR      -- US-Core ethnicity extension, ombCategory text
+);
+
+CREATE TABLE condition (
+    id                    VARCHAR PRIMARY KEY,
+    patient_id            VARCHAR,     -- join key -> patient.id
+    encounter_id          VARCHAR,     -- join key -> encounter.id
+    code                  VARCHAR,
+    system                VARCHAR,     -- kept alongside code deliberately: code-system confusion (SNOMED vs ICD-10 vs LOINC) is a failure mode to observe
+    display               VARCHAR,
+    clinicalStatus        VARCHAR,
+    verificationStatus    VARCHAR,
+    onsetDateTime         TIMESTAMP,
+    abatementDateTime     TIMESTAMP,
+    recordedDate          TIMESTAMP
+);
+
+CREATE TABLE observation (
+    id                    VARCHAR PRIMARY KEY,
+    patient_id            VARCHAR,     -- join key -> patient.id
+    encounter_id          VARCHAR,     -- join key -> encounter.id
+    code                  VARCHAR,
+    system                VARCHAR,
+    display               VARCHAR,
+    category              VARCHAR,
+    status                VARCHAR,
+    effectiveDateTime     TIMESTAMP,
+    valueQuantity         DOUBLE,      -- value[x] flattened per plan design rule
+    unit                  VARCHAR,
+    valueCodeableConcept  VARCHAR,
+    valueCodeableConcept_system VARCHAR,  -- code-system pairing for the value itself, when value[x] is coded
+    valueString           VARCHAR
+);
+
+CREATE TABLE medication_request (
+    id                    VARCHAR PRIMARY KEY,
+    patient_id            VARCHAR,     -- join key -> patient.id
+    encounter_id          VARCHAR,     -- join key -> encounter.id
+    code                  VARCHAR,
+    system                VARCHAR,
+    display               VARCHAR,
+    status                VARCHAR,
+    intent                VARCHAR,
+    authoredOn            TIMESTAMP,
+    requester_npi         VARCHAR,     -- prescribing physician's NPI (from MedicationRequest.requester)
+    requester_name        VARCHAR
+);
+
+CREATE TABLE encounter (
+    id                    VARCHAR PRIMARY KEY,
+    patient_id            VARCHAR,     -- join key -> patient.id
+    class_code            VARCHAR,     -- Encounter.class.code (AMB/EMER/IMP/HH/VR)
+    type_code             VARCHAR,     -- Encounter.type[0].coding[0]
+    type_system           VARCHAR,
+    type_display          VARCHAR,
+    status                VARCHAR,
+    period_start          TIMESTAMP,
+    period_end            TIMESTAMP,
+    reasonCode            VARCHAR,
+    participant_npi       VARCHAR,     -- primary-performer physician's NPI (Synthea models only one role per encounter, not admitting/attending/etc separately)
+    participant_name      VARCHAR
+);
+
+CREATE TABLE procedure (
+    id                    VARCHAR PRIMARY KEY,
+    patient_id            VARCHAR,     -- join key -> patient.id
+    encounter_id          VARCHAR,     -- join key -> encounter.id
+    code                  VARCHAR,
+    system                VARCHAR,
+    display               VARCHAR,
+    status                VARCHAR,
+    performedDateTime     TIMESTAMP
+    -- Note: Procedure.performer (physician who performed it) is never populated
+    -- by this Synthea version (confirmed 0/29,947) -- not available at all.
+);
+
+CREATE TABLE immunization (
+    id                    VARCHAR PRIMARY KEY,
+    patient_id            VARCHAR,     -- join key -> patient.id
+    encounter_id          VARCHAR,     -- join key -> encounter.id
+    vaccineCode           VARCHAR,
+    vaccineCode_system    VARCHAR,
+    vaccineCode_display   VARCHAR,
+    status                VARCHAR,
+    occurrenceDateTime    TIMESTAMP
+);
+
+CREATE TABLE allergy (
+    id                    VARCHAR PRIMARY KEY,
+    patient_id            VARCHAR,     -- join key -> patient.id
+    code                  VARCHAR,
+    system                VARCHAR,
+    display               VARCHAR,
+    clinicalStatus        VARCHAR,
+    verificationStatus    VARCHAR,
+    category              VARCHAR,
+    criticality           VARCHAR,
+    recordedDate          TIMESTAMP
+);
+
+CREATE TABLE careplan (
+    id                    VARCHAR PRIMARY KEY,
+    patient_id            VARCHAR,     -- join key -> patient.id
+    encounter_id          VARCHAR,     -- join key -> encounter.id
+    category_code         VARCHAR,     -- CarePlan.category[0].coding[0]
+    category_system       VARCHAR,
+    category_display      VARCHAR,
+    status                VARCHAR,
+    period_start          TIMESTAMP,
+    period_end            TIMESTAMP
+);
+
+CREATE TABLE diagnostic_report (
+    id                    VARCHAR PRIMARY KEY,
+    patient_id            VARCHAR,     -- join key -> patient.id
+    encounter_id          VARCHAR,     -- join key -> encounter.id
+    code                  VARCHAR,
+    system                VARCHAR,
+    display               VARCHAR,
+    category              VARCHAR,
+    status                VARCHAR,
+    effectiveDateTime     TIMESTAMP
+);
+
+CREATE TABLE imaging_study (
+    id                    VARCHAR PRIMARY KEY,
+    patient_id            VARCHAR,     -- join key -> patient.id
+    encounter_id          VARCHAR,     -- join key -> encounter.id
+    status                VARCHAR,
+    started               TIMESTAMP,
+    numberOfSeries        INTEGER,
+    numberOfInstances     INTEGER,
+    procedureCode         VARCHAR,     -- the imaging procedure performed, e.g. "Plain X-ray of ankle region"
+    procedureCode_system  VARCHAR,
+    procedureCode_display VARCHAR,
+    modality               VARCHAR,    -- DICOM modality code, e.g. "DX" = Digital Radiography (from series[0])
+    modality_system        VARCHAR,
+    modality_display       VARCHAR,
+    bodySite               VARCHAR,    -- from series[0]
+    bodySite_display       VARCHAR
+);