Columbia Open Health Data (COHD) Reference Ingest Guide¶
Source Information¶
InfoRes ID: infores:cohd
Description: The COHD provides access to counts and patient prevalence (i.e., prevalence from electronic health records) of conditions, procedures, drug exposures, and patient demographics, and the co-occurrence frequencies between them. Count and frequency data were derived from the Columbia University Irving Medical Center's OHDSI database including inpatient and outpatient data. Counts are the number of patients with the concept, e.g., diagnosed with a condition, exposed to a drug, or who had a procedure. Frequencies are the number of patients with the concept divided by the total number of patients in the dataset. Clinical concepts (e.g., conditions, procedures, drugs) are coded by their standard concept ID in the OMOP Common Data Model. To protect patient privacy, all concepts and pairs of concepts where the count <= 10 were excluded, and counts were randomized by the Poisson distribution. This ingest consumes the Phase 2 COHD clinical knowledge provider KGX 'nodes' and 'edges' JSONL files, which report positively/negatively correlated concept pairs.
Citations: - https://www.nature.com/articles/sdata2018273 - https://github.com/NCATSTranslator/NCATSTranslator.github.io/raw/master/presentations/Translator_2020_Kick-Off_Presentation-Clinical_Data_Services.pdf - https://github.com/NCATSTranslator/Translator-All/wiki/COHD-KP
Data Access Locations: - https://stars.renci.org/var/data_services/cohd_2/cohd_nodes.jsonl - https://stars.renci.org/var/data_services/cohd_2/cohd_edges.jsonl
Data Provision Mechanisms: file_download
Data Formats: kgx
Data Versioning and Releases: 2024-11-25
Additional Notes: ['Data are consumed as pre-computed Phase 2 KGX JSONL files hosted on the RENCI stars server, rather than fetched directly from a live COHD API.']
Ingest Information¶
Ingest Categories: translator_knowledge_creator
Utility: COHD provides statistical co-occurrence associations between clinical concepts (conditions, drugs, procedures) computed from real-world patient electronic health records. This real-world-evidence signal complements curated/mechanistic sources and is useful for Translator hypothesis generation and prioritization (e.g. drug-disease and disease-disease associations grounded in observed clinical prevalence).
Scope: Positively and negatively correlated concept pairs (diseases, drugs, chemicals) derived from Columbia University EHR data via COHD. Demographic concepts and second-level study-result ("EPC") attributes are not yet ingested.
Relevant Files¶
| File Name | Location | Description |
|---|---|---|
| cohd_nodes.jsonl | https://stars.renci.org/var/data_services/cohd_2/cohd_nodes.jsonl | KGX jsonl file of KGX 'nodes' data from Phase 2 COHD clinical knowledge provider activities. |
| cohd_edges.jsonl | https://stars.renci.org/var/data_services/cohd_2/cohd_edges.jsonl | KGX jsonl file of KGX 'edges' data from Phase 2 COHD clinical knowledge provider activities. |
Included Content¶
| File Name | Included Records | Fields Used |
|---|---|---|
| cohd_nodes.jsonl | All node records. | id, name, categories, attributes |
| cohd_edges.jsonl | All edge records. | subject, predicate, object, score, sources, attributes |
Future Content Considerations¶
node_property_content: The contents of the COHD 'attributes' field are not yet fully processed by the ingestion pipeline. For example, these attributes cross-correlate concepts with related resources. - Relevant files: cohd_nodes.jsonl
edge_property_content: The contents of the COHD 'attributes' field are not yet fully processed by the ingestion pipeline. For example, these attributes describe supporting studies, thus providing valuable evidence annotation which should be added to the COHD knowledge graph. - Relevant files: cohd_edges.jsonl
Additional Notes: None
Target Information¶
Edge Types¶
| Subject Categories | Predicate | Object Categories | Knowledge Level | Agent Type | UI Explanation |
|---|---|---|---|---|---|
| biolink:Disease, biolink:Drug, biolink:ChemicalEntity, biolink:SmallMolecule, biolink:MolecularMixture | biolink:Disease, biolink:Drug, biolink:ChemicalEntity, biolink:SmallMolecule, biolink:MolecularMixture | statistical_association | data_analysis_pipeline | COHD observed these two clinical concepts (e.g. a condition, drug exposure, or procedure) co-occurring in patient electronic health records more often than expected. Counts and patient prevalence (frequency) were computed by an automated data-analysis pipeline over the Columbia University Irving Medical Center OMOP/OHDSI clinical database (inpatient and outpatient). This positive statistical association is represented with the Biolink 'positively_correlated_with' predicate. Because the edge reflects an observed statistical correlation in real-world clinical data - not a curated or mechanistic claim - it is assigned a 'statistical_association' knowledge level and 'data_analysis_pipeline' agent type. To protect patient privacy, concepts and pairs with count <= 10 were excluded and counts were randomized via a Poisson distribution. | |
| biolink:Disease, biolink:Drug, biolink:ChemicalEntity, biolink:SmallMolecule, biolink:MolecularMixture | biolink:Disease, biolink:Drug, biolink:ChemicalEntity, biolink:SmallMolecule, biolink:MolecularMixture | statistical_association | data_analysis_pipeline | COHD observed these two clinical concepts (e.g. a condition, drug exposure, or procedure) co-occurring in patient electronic health records less often than expected. Counts and patient prevalence (frequency) were computed by an automated data-analysis pipeline over the Columbia University Irving Medical Center OMOP/OHDSI clinical database (inpatient and outpatient). This negative statistical association is represented with the Biolink 'negatively_correlated_with' predicate. Because the edge reflects an observed statistical correlation in real-world clinical data - not a curated or mechanistic claim - it is assigned a 'statistical_association' knowledge level and 'data_analysis_pipeline' agent type. To protect patient privacy, concepts and pairs with count <= 10 were excluded and counts were randomized via a Poisson distribution. |
Node Types¶
| Node Category | Source Identifier Types | Additional Notes |
|---|---|---|
| biolink:Disease | DOID, UMLS | |
| biolink:SmallMolecule | UNII, CHEBI | |
| biolink:Drug | RXCUI | |
| biolink:MolecularMixture | CHEBI | |
| biolink:ChemicalEntity | MESH, UNII, CHEBI |
Future Modeling Considerations¶
edge_properties: Second level ("EPC") study result attributes from the Phase 2 source edge jsonl file are not yet being ingested (e.g. supporting study statistics carried in the COHD 'attributes' field). Consider ingesting these as StudyResult evidence on the edges.
Provenance Information¶
Contributors: - Richard Bruskiewich - data modeling, code author - Kara Fecho - domain expertise - Matthew Brush - data modeling, domain expertise