data-dictionary-cui-mapping

data-dictionary-cui-mapping

This package assists with mapping a user's data dictionary fields to UMLS concepts. It is designed to be modular and flexible to allow for different configurations and use cases.

Roughly, the high-level steps are as follows:

• Configure yaml files
• Load in data dictionary
• Preprocess desired columns
• Query for UMLS concepts using any or all of the following pipeline modules:
  • umls (UMLS API)
  • metamap (MetaMap API)
  • semantic_search (relies on access to a custom Pinecone vector database)
  • hydra_search (combines any combination of the above three modules)
• Manually curate/select concepts in excel
• Create data dictionary file with new UMLS concept fields

Prerequisites

• For UMLS API and MetaMap API, you will need to have an account with the UMLS API and/or MetaMap API. You can sign up for an account here: https://www.nlm.nih.gov/research/umls/index.html
• For Semantic Search with Pinecone, you will need to have an account with Pinecone. You can sign up for an account here: https://www.pinecone.io/. Please reach out to me if you would like temporary access to my Pinecone index to explore these embeddings.

Installation

Use the package manager pip to install data-dictionary-cui-mapping from PyPI or pip install from the GitHub repo. The project uses poetry for packaging and dependency management.

pip install data-dictionary-cui-mapping
#pip install git+https://github.com/kevon217/data-dictionary-cui-mapping.git

Input: Data Dictionary

Below is a sample data dictionary format (.csv) that can be used as input for this package:

variable name	title	permissible value descriptions
AgeYrs	Age in years
CaseContrlInd	Case control indicator	Case;Control;Unknown

Configuration Files

In order to run and customize these pipelines, you will need to create/edit yaml configuration files located in configs. Run configurations are saved and can be reloaded.

├───ddcuimap
│   ├───configs
│   │   │   config.yaml
│   │   │   __init__.py
│   │   │
│   │   ├───apis
│   │   │       __init__.py
│   │   │       config_metamap_api.yaml
│   │   │       config_pinecone_api.yaml
│   │   │       config_umls_api.yaml
│   │   │
│   │   ├───custom
│   │   │       de.yaml
│   │   │       hydra_base.yaml
│   │   │       pvd.yaml
│   │   │       title_def.yaml
│   │   │
│   │   ├───semantic_search
│   │   │       embeddings.yaml

CUI Batch Query Pipelines

STEP-1A: RUN BATCH QUERY PIPELINE

IMPORT PACKAGES

# from ddcuimap.umls import batch_query_pipeline as umls_bqp
# from ddcuimap.metamap import batch_query_pipeline as mm_bqp
# from ddcuimap.semantic_search import batch_hybrid_query_pipeline as ss_bqp
from ddcuimap.hydra_search import batch_hydra_query_pipeline as hs_bqp

from ddcuimap.utils import helper
from omegaconf import OmegaConf

LOAD/EDIT CONFIGURATION FILES

cfg_hydra = helper.compose_config(overrides=["custom=hydra_base"])
# cfg_umls = helper.compose_config(overrides=["custom=de", "apis=config_umls_api"])
cfg_mm = helper.compose_config(overrides=["custom=de", "apis=config_metamap_api"])
cfg_ss = helper.compose_config(
    overrides=[
        "custom=title_def",
        "semantic_search=embeddings",
        "apis=config_pinecone_api",
    ]
)

# # UMLS API CREDENTIALS
# cfg_umls.apis.umls.user_info.apiKey = ''
# cfg_umls.apis.umls.user_info.email = ''

# # MetaMap API CREDENTIALS
# cfg_mm.apis.metamap.user_info.apiKey = ''
# cfg_mm.apis.metamap.user_info.email = ''
#
# # Pinecone API CREDENTIALS
# cfg_ss.apis.pinecone.index_info.apiKey = ''
# cfg_ss.apis.pinecone.index_info.environment = ''

print(OmegaConf.to_yaml(cfg_hydra))

RUN BATCH QUERY PIPELINE

# df_umls, cfg_umls = umls_bqp.run_umls_batch(cfg_umls)
# df_mm, cfg_mm = mm_bqp.run_mm_batch(cfg_mm)
# df_ss, cfg_ss = ss_bqp.run_hybrid_ss_batch(cfg_ss)
df_hydra, cfg_step1 = hs_bqp.run_hydra_batch(cfg_hydra, cfg_umls=None, cfg_mm=cfg_mm, cfg_ss=cfg_ss)

print(df_hydra.head())

STEP-1B: *MANUAL CURATION STEP IN EXCEL

CURATION/SELECTION

*see curation example in notebooks/examples_files/DE_Step-1_curation_keepCol.xlsx

STEP-2A: CREATE DATA DICTIONARY IMPORT FILE

IMPORT CURATION MODULES

from ddcuimap.curation import create_dictionary_import_file
from ddcuimap.curation import check_cuis
from ddcuimap.utils import helper

CREATE DATA DICTIONARY IMPORT FILE

cfg_step1 = helper.load_config(helper.choose_file("Load config file from Step 1"))
df_dd = create_dictionary_import_file.create_dd_file(cfg_step1)
print(df_dd.head())

STEP-2B: CHECK CUIS IN DATA DICTIONARY IMPORT FILE

CHECK CUIS

cfg_step2 = helper.load_config(helper.choose_file("Load config file from Step 2"))
df_check = check_cuis.check_cuis(cfg_step2)
print(df_check.head())

Output: Data Dictionary + CUIs

Below is a sample modified data dictionary with curated CUIs after:

• Running Steps 1-2 on title then taking the generated output dictionary file and;
• Running Steps 1-2 again on permissible value descriptions to get the final output dictionary file.

variable name	title	data element concept identifiers	data element concept names	data element terminology sources	permissible values	permissible value descriptions	permissible value output codes	permissible value concept identifiers	permissible value concept names	permissible value terminology sources
AgeYrs	Age in years	C1510829;C0001779	Age-Years;Age	UMLS;UMLS
CaseContrlInd	Case control indicator	C0007328	Case-Control Studies	UMLS	Case;Control;Unknown	Case;Control;Unknown	1;2;999	C1706256;C4553389;C0439673	Clinical Study Case;Study Control;Unknown	UMLS;UMLS;UMLS

Semantic Search with SentenceTransformers Batch Queries

More documentation to come... Basic pipeline is described below:

Subset/Embed/Upsert UMLS Metathesaurus for Pinecone vector database

Step 1: Subset local copy of UMLS Metathesaurus

Step 2: Embed UMLS CUI names and definitions and format metadata

Step 3: Upsert embeddings and metadata into Pinecone index

Query UMLS Metathesaurus vector database with data dictionary embeddings

Step 1: Embed data dictionary fields

Step 2: Batch Query data dictionary against CUI names and definitions in Pinecone index

Step 3: Evaluate/Curate Results

Step 4: Create data dictionary based on curation

Acknowledgements

The MetaMap API code included is from Will J Roger's repository --> https://github.com/lhncbc/skr_web_python_api

Special thanks to Olga Vovk, Henry Ogoe, and Sofia Syed for their guidance, feedback, and testing of this package.

License

MIT