TAXIS (Transparent Analytic Knowledge Graph for Interoperable Science) is an OHDSI-led methodological infrastructure project that derives, evaluates, and validates clinically meaningful relationships between coded health concepts using retrospective, de-identified data in the OMOP Common Data Model (CDM). The project combines large-scale co-occurrence statistics with large language model (LLM) classification to characterize semantic relationships (e.g., causal, confusable, hierarchical). Human reviewers validate discordant or uncertain classifications.
The name TAXIS comes from the Greek word τάξις (táxis), meaning order, arrangement, or structured framework. The project reflects this sense of systematic organization, aiming to bring order and transparency to the complex web of clinical relationships.
The resulting knowledge graph provides a foundational analytic resource that can be used to:
- Simplify phenotype and cohort development
- Accelerate quality measurement
- Support clinical decision support rules
- Enable hypothesis generation at scale
- Advance applied clinical analytics in learning health systems
Importantly, no personally identifiable (PII) or protected health information (PHI) is used at any stage.
We invite OHDSI collaborators and external researchers to participate directly:
-
Run the Study Package
- Clone and execute the TAXIS study package against your OMOP CDM instance.
- SQL scripts (e.g.,
concept_ab.sql) compute co-occurrence statistics, directionality ratios, odds ratios, and related measures.
-
Upload Step 5 Output
- Once the study package produces a step_5 CSV file, upload it to the TAXIS web portal.
- The portal will process the data, apply LLM-based classification of relationships, and provide validated outputs.
-
Download Validated Results
- Retrieve enriched files containing both statistical evidence and semantic validation.
- Sample outputs are provided in the examples directory.
By contributing, you:
- Help build the world’s most comprehensive, data-driven clinical knowledge graph.
- Gain early access to validated relationship datasets for local analytic use.
- Contribute to open science infrastructure that supports reproducible, scalable analytics.
The first deliverable of TAXIS is the Concept AB (CONCEPT_CF_AB) package, which creates a curated, clinically meaningful, and analytically useful library of relationships between pairs of OMOP “clinical finding” concepts. This supports everyday analytic tasks such as cohort definition and outcome specification.
- Input: The package reads from OMOP CDM tables (
condition_occurrence,concept,concept_relationship,concept_ancestor). - Processing: It computes frequent, statistically significant person-level co-occurrences within a configurable ±30-day window. It then materializes a final output table,
concept_ab_step_5, with co-occurrence counts and statistical measures (lift, odds ratio, directionality ratio, etc.). - Classification: Concept pairs are categorized using an LLM into relationship types (e.g., causal, subset, similar presentation).
- Portability: Parameterized, SqlRender-compatible SQL is orchestrated in R, enabling execution across supported DBMSs.
- OS: Windows, macOS, or Linux
- R ≥ 4.2; Java ≥ 8 (17 LTS recommended)
- Required R packages: SqlRender, DatabaseConnector
- Database: OMOP CDM with vocabulary loaded
- A DBA-provisioned least-privilege account with:
- SELECT on required CDM and curated lookup tables
- DDL/DML rights only in a designated results schema
- Connection fields: DBMS, DB_SERVER, DB_PORT, DB_USER, DB_AUTH_MODE, JDBC folder
- Schemas: CDM schema, lookup schema, results schema
- Runtime controls: co-occurrence window (days), minimum persons, batch counts, index settings
- Edit
.Renvironwith site-specific values. - Install dependencies; run
concept_cf_ab_precheck.Rto validate environment. - Run
concept_cf_ab_run.Rto execute init → batch loop → finalize. - Review outputs in the designated results schema.
- Ensure
.Renvironsettings match CAB_SQL_* entries. - Confirm JDBC drivers are installed.
- Check database permissions if create/drop/index fails.
- Load the lookup CSV (
concept_ab_cf_rollup) into the project schema.
- TAXIS does not modify system-level security.
- Connections use OHDSI’s DatabaseConnector with least-privilege credentials.
OHDSI references: DatabaseConnector Docs, SqlRender Docs
2024: Foundational work on diagnostic hierarchies and early relationship table Sep 2025: Launch of TAXIS OHDSI Network Study, protocol published Oct 2025: First large-scale outputs: v1 Diagnosis–Diagnosis Relationship Tables and network engagement launch at OHDSI Symposium Oct-Nov 2025: Participating site run study package for disagnosis-diagnosis relationships locally and upload results thorught TAXIS website for validation and aggregation, Public domain Medication relationships aggregated and published Dec 2025: Diagnosis-diagnosis relationships based on Network data review and cleanup Jan 2026: Add Service Relationships to study package and Diagnosis-Diagnosis relationships V1 availablel Feb-Mar 2026: Participating site run updated study package for service relationships locally and upload results thorught TAXIS website for validation and aggregation Mar 2026: Service relationships based on Network data review and cleanup Apr 2026: Diagnosis, Service and Medication relationships V1 available 2026: Work with OMOP vocabulary team to incorporate results where appropriate into standard and evaluate how to leverage the relatinships fully in ATLAS.
- docs/
- Study protocol, concept papers, symposium drafts, technical documentation, timeline.
- study_package/
concept_ab.sql– Core OMOP condition co-occurrence analysis pipeline.
- examples/
sample_step5.csv– Example OMOP site output for upload.sample_validated_output.csv– Example download after LLM validation.
- publications/
- OHDSI Symposium brief report and related materials.
- Expand from diagnosis–diagnosis pairs to multi-domain relationships (diagnosis–service, diagnosis–medication).
- Improve handling of large file uploads and watchdog functions for failed jobs.
- Extend schema to capture per-upload site statistics in addition to aggregate master record counts.
- Publish reproducible pipelines for public use via OHDSI GitHub.
If you use TAXIS outputs in your work, please cite:
Overhage JM, Bandeian S, et al. TAXIS: Enhancing OMOP Capabilities Through Knowledge Graphs. OHDSI Symposium 2025 (forthcoming).
-
Study Leads:
- J. Marc Overhage, MD, PhD – Indiana University / Elevance Health
- Stephen H. Bandeian, MD, JD – Johns Hopkins University
-
OHDSI Forums Tag: TAXIS