pyomop: OMOP Swiss Army Knife 🔧¶
✨ Overview¶
pyomop is your OMOP Swiss Army Knife 🔧 for working with OHDSI OMOP Common Data Model (CDM) v5.4 or v6 compliant databases using SQLAlchemy as the ORM. It supports converting query results to pandas DataFrames for machine learning pipelines and provides utilities for working with OMOP vocabularies. Table definitions are based on the omop-cdm library. Pyomop is designed to be a lightweight, easy-to-use library for researchers and developers experimenting and testing with OMOP CDM databases. It can be used both as a commandline tool and as an imported library in your code.
- Supports SQLite, PostgreSQL, and MySQL. CDM and Vocab tables are created in the same schema. (See usage below for more details)
- LLM-based natural language queries via langchain. Usage.
- 🔥 FHIR to OMOP conversion utilities. (See usage below for more details)
- Execute QueryLibrary. (See usage below for more details)
Please ⭐️ If you find this project useful!
Installation¶
Stable release:
pip install pyomop
Development version:
git clone https://github.com/dermatologist/pyomop.git
cd pyomop
pip install -e .
LLM support:
pip install pyomop[llm]
✨ See this notebook or script for examples. 👇 MCP SERVER is recommended for advanced usage.¶
Docker¶
- A docker-compose is provided to quickly set up an environment with postgrs, webapi, atlas and a sql script to create a source in webapi. The script can be run using the
psqlcommand line tool or via the webapi UI. Please refresh after running the script by sending a request to /WebAPI/source/refresh.
🔧 Usage¶
import asyncio
import datetime
from sqlalchemy import select
from pyomop import CdmEngineFactory, CdmVector, CdmVocabulary
# cdm6 and cdm54 are supported
from pyomop.cdm54 import Base, Cohort, Person, Vocabulary
async def main():
cdm = CdmEngineFactory() # Creates SQLite database by default for fast testing
# cdm = CdmEngineFactory(db='pgsql', host='', port=5432,
# user='', pw='',
# name='', schema='')
# cdm = CdmEngineFactory(db='mysql', host='', port=3306,
# user='', pw='',
# name='')
engine = cdm.engine
# Comment the following line if using an existing database. Both cdm6 and cdm54 are supported, see the import statements above
await cdm.init_models(Base.metadata) # Initializes the database with the OMOP CDM tables
vocab = CdmVocabulary(cdm, version='cdm54') # or 'cdm6' for v6
# Uncomment the following line to create a new vocabulary from CSV files
# vocab.create_vocab('/path/to/csv/files')
async with cdm.session() as session: # type: ignore
# Add Persons
async with session.begin():
session.add(
Person(
person_id=100,
gender_concept_id=8532,
gender_source_concept_id=8512,
year_of_birth=1980,
month_of_birth=1,
day_of_birth=1,
birth_datetime=datetime.datetime(1980, 1, 1),
race_concept_id=8552,
race_source_concept_id=8552,
ethnicity_concept_id=38003564,
ethnicity_source_concept_id=38003564,
)
)
session.add(
Person(
person_id=101,
gender_concept_id=8532,
gender_source_concept_id=8512,
year_of_birth=1980,
month_of_birth=1,
day_of_birth=1,
birth_datetime=datetime.datetime(1980, 1, 1),
race_concept_id=8552,
race_source_concept_id=8552,
ethnicity_concept_id=38003564,
ethnicity_source_concept_id=38003564,
)
)
# Query the Person
stmt = select(Person).where(Person.person_id == 100)
result = await session.execute(stmt)
for row in result.scalars():
print(row)
assert row.person_id == 100
# Query the person pattern 2
person = await session.get(Person, 100)
print(person)
assert person is not None
assert person.person_id == 100
# Convert result to a pandas dataframe
vec = CdmVector()
# https://github.com/OHDSI/QueryLibrary/blob/master/inst/shinyApps/QueryLibrary/queries/person/PE02.md
result = await vec.query_library(cdm, resource='person', query_name='PE02')
df = vec.result_to_df(result)
print("DataFrame from result:")
print(df.head())
result = await vec.execute(cdm, query='SELECT * from person;')
print("Executing custom query:")
df = vec.result_to_df(result)
print("DataFrame from result:")
print(df.head())
# Close engine
await engine.dispose() # type: ignore
# Run the main function
asyncio.run(main())
🔥 FHIR to OMOP mapping¶
pyomop can load FHIR Bulk Export (NDJSON) files into an OMOP CDM database.
- Sample datasets: https://github.com/smart-on-fhir/sample-bulk-fhir-datasets
- Remove any non-FHIR files (for example,
log.ndjson) from the input folder. - Download OMOP vocabulary CSV files (for example from OHDSI Athena) and place them in a folder.
Run:
pyomop --create --vocab ~/Downloads/omop-vocab/ --input ~/Downloads/fhir/
This will create an OMOP CDM in SQLite, load the vocabulary files, and import the FHIR data from the input folder and reconcile vocabulary, mapping source_value to concept_id. The mapping is defined in the mapping.example.json file. The default mapping is here. Mapping happens in 5 steps as implemented here.
- Example using postgres (Docker)
pyomop --dbtype pgsql --host localhost --user postgres --pw mypass --create --vocab ~/Downloads/omop-vocab/ --input ~/Downloads/fhir/
- FHIR to data frame mapping is done with FHIRy
- Most of the code for this functionality was written by an LLM agent. The prompts used are here
Command-line¶
-c, --create Create CDM tables (see --version).
-t, --dbtype TEXT Database Type for creating CDM (sqlite, mysql or
pgsql)
-h, --host TEXT Database host
-p, --port TEXT Database port
-u, --user TEXT Database user
-w, --pw TEXT Database password
-v, --version TEXT CDM version (cdm54 (default) or cdm6)
-n, --name TEXT Database name
-s, --schema TEXT Database schema (for pgsql)
-i, --vocab TEXT Folder with vocabulary files (csv) to import
-f, --input DIRECTORY Input folder with FHIR bundles or ndjson files.
-e, --eunomia-dataset TEXT Download and load Eunomia dataset (e.g.,
'GiBleed', 'Synthea')
--eunomia-path TEXT Path to store/find Eunomia datasets (uses
EUNOMIA_DATA_FOLDER env var if not specified)
--connection-info Display connection information for the database (For R package compatibility)
--mcp-server Start MCP server for stdio interaction
--pyhealth-path TEXT Path to export PyHealth compatible CSV files
--help Show this message and exit.
MCP Server¶
pyomop includes an MCP (Model Context Protocol) server that exposes tools for interacting with OMOP CDM databases. This allows MCP clients to create databases, load data, and execute SQL statements.
Usage with MCP Clients¶
The MCP server can be used with any MCP-compatible client such as Claude desktop. Example configuration for VSCODE as below is already provided in the repository. So if you are viewing this in VSCODE, you can start server and enable tools directly in Copilot.
{
"servers": {
"pyomop": {
"command": "uv",
"args": ["run", "pyomop", "--mcp-server"]
}
}
}
- If the vocabulary is not installed locally or advanced vocabulary support is required from Athena, it is recommended to combine omop_mcp with PyOMOP.
Available MCP Tools¶
- create_cdm: Create an empty CDM database
- create_eunomia: Add Eunomia sample dataset
- get_table_columns: Get column names for a specific table
- get_single_table_info: Get detailed table information, including foreign keys
- get_usable_table_names: Get a list of all available table names
- run_sql: Execute SQL statements with error handling
- example_query: Get example queries for specific OMOP CDM tables from OHDSI QueryLibrary
-
check_sql: Validate SQL query syntax before execution
-
create_cdm and create_eunomia support only local sqlite databases to avoid inadvertent data loss in production databases.
HTTP Transport Support¶
The MCP server now supports both stdio (default) and HTTP transports:
Stdio transport (default):
pyomop --mcp-server
# or
pyomop-mcp-server
HTTP transport:
pyomop-mcp-server-http
# or with custom host/port
pyomop-mcp-server-http --host 0.0.0.0 --port 8000
# or via Python module
python -m pyomop.mcp.server --http --host 0.0.0.0 --port 8000
To use HTTP transport, install additional dependencies:
pip install pyomop[http]
# or for both LLM and HTTP features
pip install pyomop[llm,http]
Available Prompts¶
- query_execution_steps: Provides step-by-step guidance for executing database queries based on free text instructions
Eunomia import and cohort creation¶
pyomop -e Synthea27Nj -v 5.4 --connection-info
pyomop -e GiBleed -v 5.3 --connection-info
PyHealth and PLP Compatibility (For Machine Learning pipelines)¶
pyomop supports exporting OMOP CDM data (to --pyhealth-path) in a format compatible with PyHealth, a machine learning library for healthcare data analysis (See Notebook and usage below). Additionally, you can export the connection information for use with the various R packages such as PatientLevelPrediction using the --connection-info option.
pyomop -e GiBleed -v 5.3 --connection-info --pyhealth-path ~/pyhealth
Additional Tools¶
Supported Databases¶
- PostgreSQL
- MySQL
- SQLite
Environment Variables for Database Connection¶
You can configure database connection parameters using environment variables. These will be used as defaults by pyomop and the MCP server:
PYOMOP_DB: Database type (sqlite,mysql,pgsql)PYOMOP_HOST: Database hostPYOMOP_PORT: Database portPYOMOP_USER: Database userPYOMOP_PW: Database passwordPYOMOP_SCHEMA: Database schema (for PostgreSQL)
Example usage:
export PYOMOP_DB=pgsql
export PYOMOP_HOST=localhost
export PYOMOP_PORT=5432
export PYOMOP_USER=postgres
export PYOMOP_PW=mypass
export PYOMOP_SCHEMA=omop
These environment variables will be checked before assigning default values for database connection in pyomop and MCP server tools.
Contributing¶
Pull requests are welcome! See CONTRIBUTING.md.
