omop-cdm-graph

A lightweight, persistent knowledge graph of the OMOP Common Data Model v5.4 schema, built on Ladybug — an embeddable graph database with Cypher query support.

The graph encodes all 37 CDM tables as nodes, their 438 columns as typed property nodes, and all 186 foreign-key relationships as directed edges. Once built, the database can be queried to enumerate tables, resolve column metadata, and compute join paths between arbitrary table pairs.

---

Motivation

Generating valid SQL over an OMOP CDM instance requires knowledge of the schema's relational structure: which tables exist, what columns they carry, and how they join. Encoding this as a property graph makes the structure machine-queryable in a way that flat documentation or ad-hoc string templates do not.

---

Graph model

Node type	Count	Properties
OmopDomain	7	name
OmopTable	37	name, domain, description
OmopColumn	438	id, table_name, column_name, data_type, is_pk, is_fk, is_nullable, description

Edge type	Count	Properties
BELONGS_TO_DOMAIN	37	—
HAS_COLUMN	438	—
FOREIGN_KEY	186	join_type, fk_desc
JOINS_TO	186	from_column, to_column, join_sql, join_type

JOINS_TO edges are derived from FOREIGN_KEY edges and provide a direct table-to-table traversal surface, including self-referential joins (e.g. visit_detail.parent_visit_detail_id).

---

Requirements

• Python ≥ 3.10
• uv

---

Installation

git clone <repo>
cd omop-cdm_graph
uv sync

---

Build

uv run python scripts/build_graph.py

Produces omop_cdm.lbug in the working directory. Re-running overwrites the existing file. To skip a rebuild if the file already exists:

uv run python scripts/build_graph.py --no-rebuild

---

Usage

from omop_graph.querier import OMOPGraphQuerier

q = OMOPGraphQuerier("./omop_cdm.lbug")

# All columns for a table
q.get_columns("condition_occurrence")

# Tables directly joinable from a given table
q.get_direct_joins("condition_occurrence")

# Both inbound and outbound joins
q.get_all_joins_for_table("person")

# Shortest join path between two tables (BFS, undirected)
q.get_join_path("condition_occurrence", "location")
# → ['condition_occurrence', 'person', 'location']

# SQL JOIN snippet for a direct relationship
q.get_join_sql("condition_occurrence", "visit_occurrence")

# Tables containing a given column name
q.find_tables_with_column("person_id")

# Tables in a domain
q.get_tables_in_domain("derived")

# Raw Cypher passthrough
q.run_cypher(
    "MATCH (t:OmopTable)-[:JOINS_TO]->(t2:OmopTable {name: $name}) RETURN t.name",
    parameters={"name": "concept"}
)

---

Structure

src/omop_graph/
  schema.py     — Pydantic models; all 37 tables, 438 columns, 186 joins (source of truth)
  builder.py    — Ladybug DDL and graph construction
  querier.py    — Query interface over the built graph
  cli.py        — Entry point for omop-graph-build
scripts/
  build_graph.py
tests/
  test_schema.py   — Schema completeness and integrity
  test_builder.py  — Graph construction and query correctness

---

Tests

uv run pytest tests/ -v

33 tests covering schema completeness, join referential integrity, graph construction, and all query methods.

---

Schema source

All table and column definitions follow the OMOP CDM v5.4 specification. The schema is defined statically in schema.py and does not require a live CDM instance.

Contact

k24118093@kcl.ac.uk