# Research: FHIR Radiology Integration
**Feature**: 007-fhir-radiology-integration
**Date**: 2025-12-15
**Status**: Complete
## Research Questions
### 1. Patient Matching Strategy
**Question**: How should MIMIC-CXR subject_ids be mapped to FHIR Patient resources?
**Decision**: Create new synthetic FHIR patients using Synthea for unmatched MIMIC subject_ids
**Rationale**:
- Synthea generates internally consistent patient records (demographics, conditions, encounters)
- Existing synthetic FHIR data can be safely modified for demo purposes
- Ensures referential integrity - every image links to a valid Patient resource
**Alternatives Considered**:
- Manual mapping file: Rejected - doesn't scale, requires maintenance
- Random assignment: Rejected - creates incoherent patient stories
- Leave unlinked: Rejected - defeats the purpose of the feature
### 2. FHIR Server Implementation
**Question**: What FHIR server is currently deployed?
**Decision**: InterSystems IRIS for Health FHIR repository
**Rationale**:
- Already deployed on EC2 with existing FHIR Patient resources
- Native vector support in same database as image embeddings
- REST API access for FHIR resource queries
**Implementation Notes**:
- FHIR R4 standard
- REST endpoint pattern: `GET /fhir/r4/[ResourceType]?[params]`
- Supports `_include` and `_revinclude` for linked resources
### 3. FHIR ImagingStudy Resource Structure
**Decision**: Use FHIR R4 ImagingStudy with standard fields for MIMIC-CXR integration
**Key Fields for Patient Linking**:
```json
{
"resourceType": "ImagingStudy",
"id": "study-s50414267",
"status": "available",
"subject": { "reference": "Patient/p10002428" },
"encounter": { "reference": "Encounter/enc-123" },
"started": "2024-01-15T10:30:00Z",
"identifier": [
{
"system": "urn:mimic-cxr:study",
"value": "s50414267"
}
],
"modality": [
{ "system": "http://dicom.nema.org/resources/ontology/DCM", "code": "CR" }
],
"numberOfSeries": 1,
"numberOfInstances": 1,
"series": [
{
"uid": "1.2.3.4.5",
"modality": { "code": "CR" },
"instance": [
{
"uid": "1.2.3.4.5.1",
"sopClass": { "code": "1.2.840.10008.5.1.4.1.1.1.1" },
"number": 1
}
]
}
]
}
```
**Rationale**:
- `subject` links to FHIR Patient via reference
- `encounter` enables temporal context
- `identifier` preserves MIMIC study_id for traceability
- Standard DICOM modality codes (CR = Computed Radiography)
### 4. FHIR DiagnosticReport for Radiology Reports
**Decision**: Store MIMIC-CXR report text as FHIR DiagnosticReport with presentedForm
**Key Fields**:
```json
{
"resourceType": "DiagnosticReport",
"id": "report-s50414267",
"status": "final",
"code": {
"coding": [
{
"system": "http://loinc.org",
"code": "18748-4",
"display": "Diagnostic imaging study"
}
]
},
"subject": { "reference": "Patient/p10002428" },
"encounter": { "reference": "Encounter/enc-123" },
"effectiveDateTime": "2024-01-15T11:00:00Z",
"imagingStudy": [
{ "reference": "ImagingStudy/study-s50414267" }
],
"conclusion": "No acute cardiopulmonary process.",
"conclusionCode": [
{
"coding": [
{
"system": "http://snomed.info/sct",
"code": "373067005",
"display": "No abnormality detected"
}
]
}
],
"presentedForm": [
{
"contentType": "text/plain",
"data": "RklORElOR1M6IE5vIGFjdXRlIGNhcmRpb3B1bG1vbmFyeSBwcm9jZXNzLg=="
}
]
}
```
**Rationale**:
- `presentedForm` stores full report text (base64 encoded)
- `conclusion` provides human-readable summary
- `imagingStudy` links report to ImagingStudy resource
- Standard LOINC codes for diagnostic imaging
### 5. MCP Tool Architecture
**Decision**: Follow existing MCP server patterns with 6 new radiology-specific tools
**Tools to Implement**:
| Tool Name | Description | FHIR Query |
|-----------|-------------|------------|
| `get_patient_imaging_studies` | All ImagingStudy for a patient | `ImagingStudy?subject=Patient/{id}` |
| `get_imaging_study_details` | Single study with full metadata | `ImagingStudy/{id}?_include=*` |
| `get_radiology_reports` | DiagnosticReport for imaging | `DiagnosticReport?imaging-study={id}` |
| `search_patients_with_imaging` | Patients with imaging by criteria | `ImagingStudy?_include=ImagingStudy:subject` |
| `get_encounter_imaging` | Studies for an encounter | `ImagingStudy?encounter=Encounter/{id}` |
| `list_radiology_queries` | Available query templates | N/A (metadata tool) |
**Rationale**:
- Follows existing MCP tool patterns in `fhir_graphrag_mcp_server.py`
- Uses IRIS for Health REST API for FHIR queries
- Returns JSON for LLM consumption
### 6. Patient-Image Mapping Table
**Decision**: Create new IRIS table `VectorSearch.PatientImageMapping`
**Schema**:
```sql
CREATE TABLE VectorSearch.PatientImageMapping (
MIMICSubjectID VARCHAR(255) PRIMARY KEY,
FHIRPatientID VARCHAR(255) NOT NULL,
FHIRPatientName VARCHAR(500),
MatchConfidence DECIMAL(3,2),
MatchType VARCHAR(50), -- 'exact', 'synthea_generated', 'manual'
CreatedAt TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);
```
**Rationale**:
- Enables fast lookups during image search
- Preserves match metadata for auditing
- Supports idempotent import operations
### 7. UI Enhancement for Image Search
**Decision**: Modify `search_medical_images` response to include patient context
**Current Response** (showing "Unknown Patient"):
```json
{
"images": [
{
"image_id": "p10002428_s50414267_view1",
"study_type": "Unknown Study",
"patient_id": "Unknown Patient"
}
]
}
```
**Enhanced Response**:
```json
{
"images": [
{
"image_id": "p10002428_s50414267_view1",
"study_id": "s50414267",
"patient_name": "John Smith",
"patient_mrn": "MRN-12345",
"fhir_patient_id": "Patient/p10002428",
"view_position": "PA",
"similarity_score": 0.87
}
]
}
```
**Rationale**:
- Addresses FR-003: Display patient name in search results
- Enables FR-009: Navigation to patient record
## Dependencies
| Dependency | Purpose | Status |
|------------|---------|--------|
| InterSystems IRIS for Health | FHIR repository + vector storage | Deployed |
| Synthea | Generate synthetic patients | Available (external) |
| MCP SDK | Tool exposure | Installed |
| NV-CLIP | Image embeddings | Deployed |
## Risks and Mitigations
| Risk | Impact | Mitigation |
|------|--------|------------|
| MIMIC subject_ids have no FHIR match | Images show "Unlinked" | Generate Synthea patients for unmatched |
| FHIR API rate limits | Slow queries | Batch operations, caching |
| Data inconsistency | Orphaned resources | Idempotent import with validation |
## Next Steps
1. Create `VectorSearch.PatientImageMapping` table
2. Implement FHIR resource generation scripts
3. Add 6 new MCP tools to server
4. Update UI to display patient context
5. Write tests (contract, integration, UX)
