| Accessibility |
Lower barriers for clinicians and researchers without deep SQL/R/Python expertise |
Create dependence on proprietary models and expensive infrastructure |
Open, model-independent workflows |
| Phenotyping |
Accelerate phenotype definition, concept-set creation, and cohort design |
Generate inconsistent or clinically incorrect definitions |
Human review, validation, versioning |
| Analytics |
Automate coding, study setup, diagnostics, and interpretation |
Produce executable but epidemiologically inappropriate analyses |
Transparent methods and reproducible code |
| Reproducibility |
Automatically document workflows, parameters, and provenance |
Prompt/model changes may produce different results |
Full provenance of model, prompt, data, tools, and outputs |
| Scientific quality |
Help detect errors, missing analyses, confounding, and data-quality issues |
Make weak science look polished and convincing |
Diagnostics, sensitivity analyses, negative controls |
| Global equity |
Enable more institutions worldwide to participate in OHDSI studies |
Widen gaps between organizations with different AI resources |
Open tools, benchmarks, and shared infrastructure |
| Community knowledge |
Turn OHDSI documentation and experience into accessible AI assistants |
Replace community learning with opaque automated answers |
Keep humans and community expertise in the loop |
| Long-term opportunity |
Position OHDSI as a global platform for validating AI-generated real-world evidence |
Allow uncontrolled AI-generated evidence to erode trust |
AI should accelerate OHDSI’s scientific process, not replace it |