Appendix F — Known Limitations
This appendix documents known limitations and constraints in the current release of Parthenon. Items are organized by module and include workarounds where available.
Analysis Capabilities
Treatment Pathway Visualization
The Sankey/Sunburst interactive diagram for treatment pathways is not yet implemented. Pathway results are currently shown in tabular form. An interactive Sankey visualization is planned for a future release.
Network Studies
Multi-database federated execution (sending analysis packages to distributed sites) is not yet supported. Each data source must be directly accessible from the Parthenon server.
Negative Controls
Automated negative control outcome generation (used in PLE and SCCS calibration) is not yet integrated. Negative control concept sets must be manually curated and added to analysis configurations.
R Runtime Dependency
PLE (CohortMethod), PLP (PatientLevelPrediction), SCCS, and Evidence Synthesis analyses require the R runtime container to be running and connected. Configuration UIs for these analysis types are fully implemented, but execution depends on the R/HADES packages being properly installed in the R container.
Genomics
ClinVar Annotation Requires Local Database
ClinVar variant annotation requires downloading and loading the ClinVar database locally via POST /api/v1/genomics/clinvar/sync. The initial download is approximately 1.5 GB and indexing can take 15--30 minutes. Without a local ClinVar database, variant clinical significance annotations will not be available.
Run the ClinVar sync during off-hours to avoid impacting other operations. The sync is incremental after the initial load.
VCF File Size
Large VCF files (>500 MB) may time out during upload via the web interface. For files exceeding this size, use the CLI bulk loader or split the VCF by chromosome before uploading.
Multi-Sample VCF
Multi-sample VCF files are supported but each sample must be matched to an OMOP person_id individually via the person matching step. If sample identifiers in the VCF do not match patient identifiers in the CDM, manual mapping is required.
Imaging
Single-Frame DICOM Only
The current DICOM viewer (Cornerstone3D) supports single-frame DICOM instances. Multi-frame DICOM objects (e.g., enhanced MR, enhanced CT, ultrasound cine loops) are planned for a future release. Multi-frame files can be indexed and stored but will not render in the built-in viewer.
DICOMweb Server Required for Remote Studies
Retrieving imaging studies from a remote PACS requires a DICOMweb-compatible server (Orthanc, dcm4chee, Google Cloud Healthcare API, etc.). Legacy DICOM C-FIND/C-MOVE protocols are not supported. Local DICOM file import is available as an alternative.
Radiomics Feature Extraction
Radiomics feature extraction (texture, shape, intensity features) is available but relies on pre-computed features. Real-time feature extraction from raw DICOM pixel data is not yet supported within Parthenon; use external tools (PyRadiomics) and import results.
NLP Report Extraction
Imaging report NLP extraction uses the configured AI provider (Ollama/MedGemma by default). Extraction quality depends on the model's capabilities and the structure of the radiology report text. Complex or highly abbreviated reports may produce incomplete structured findings.
HEOR
Cost Models Depend on CDM Cost Data
HEOR cost-effectiveness and budget impact analyses require cost data in the OMOP CDM cost table. If your CDM does not include cost data (common in EHR-only installations), HEOR economic analyses will produce incomplete or estimated results. The HEOR module can work with external cost assumptions configured via analysis parameters.
Rebate Simulation Accuracy
Payer contract rebate simulations use simplified models. Actual rebate calculations depend on complex contractual terms that may not be fully captured in the simulation parameters. Results should be used for directional analysis, not as definitive financial projections.
FHIR Integration
Tested EHR Vendors
FHIR R4 integration has been tested with Epic and Cerner (Oracle Health) EHR systems. Other FHIR R4-compliant vendors should work but may require custom mapping adjustments for non-standard resource profiles or extensions.
SMART Backend Services Authentication
FHIR connections use SMART Backend Services (system-to-system JWT auth). Your EHR vendor must support this authorization profile and provide a registered client ID and JWKS endpoint or public key. User-facing SMART on FHIR (interactive OAuth2 launch) is not supported.
Bulk Data Export Timeouts
FHIR Bulk Data Export operations on large patient populations may take hours. If the EHR server's export job times out or fails mid-export, Parthenon will retry from the last checkpoint. However, some EHR vendors impose export size limits or concurrent export restrictions that may require coordination with the EHR team.
Resource Type Coverage
Parthenon maps the following FHIR R4 resource types to OMOP CDM tables: Patient, Condition, MedicationRequest, Procedure, Observation, Encounter, DiagnosticReport, and Immunization. Other resource types (AllergyIntolerance, CarePlan, Goal, etc.) are not yet mapped and will be skipped during sync.
Data and Vocabulary
Single CDM Dialect
Parthenon currently supports PostgreSQL only. Support for SQL Server, BigQuery, Redshift, and Spark is on the roadmap.
CDM Versions
Only OMOP CDM v5.4 is fully supported. CDM v5.3 databases will work for most queries but may encounter column naming differences in newer CDM additions.
Performance
Large CDM Generation Times
Cohort generation on CDMs with >100 million patients may take 15--60 minutes for complex expressions. There is no progress streaming to the UI during generation --- only the terminal status (Running/Complete) is displayed.
Achilles Results Are Pre-Computed
The Data Explorer reflects Achilles pre-computed results, not real-time CDM queries. Results may be weeks or months old depending on when Achilles was last run against the source.
Authentication and Access
Mobile Responsiveness
The interface is optimized for desktop use (1280px or wider). Tablet layouts are partially supported; mobile (<768px) use is not recommended for clinical research workflows.
Screen Reader Support
The React application has partial ARIA labeling. Full screen reader compatibility (WCAG 2.1 AA) is targeted for a future release.
Search
Documentation Search
The Docusaurus documentation site (/docs/) uses Lunr.js for client-side search. This provides good results for exact and partial term matches but lacks the relevance ranking and typo tolerance of a server-side search engine. Algolia DocSearch integration is planned as a future enhancement.
File Size Limits
| Feature | Limit | Workaround |
|---|---|---|
| Web file upload (general) | 500 MB per file | Use CLI bulk loader |
| VCF upload | 500 MB per file | Split by chromosome |
| DICOM local import | No limit (async) | Use import-local/trigger for large batches |
| Athena vocabulary ZIP | 2 GB | Upload directly; extraction is streamed |
| API request body | 100 MB | Use streaming upload endpoints |
| Export file size | Unlimited | Streaming response |