Skip to main content

Oncology Record Ingestion Status

To: Sumeet
From: Technical Assessment
Re: Distance from ingesting synthetic oncology record into FHIR and displaying in Lyfe
Date: December 2024

Executive Answer​

We are 2-4 weeks away from a basic demo of ingesting one synthetic oncology record into FHIR and displaying it in Lyfe, but 3-4 months away from a production-ready system that could handle real oncology data reliably.

Current Capabilities​

✅ What Works Today​

  1. Document Upload & Processing

    User uploads PDF → AI extracts data → Stores in database
    • Can accept oncology reports (PDF/image)
    • AI extracts basic patient information
    • Stores extracted data in Supabase

  2. FHIR Structure Understanding

    • System knows FHIR Patient, Condition, Observation resources
    • Can create basic FHIR bundles
    • Proper resource references

  3. UI Display

    • Patient demographics display
    • Conditions list
    • Medications view
    • Timeline visualization

❌ What's Missing for Oncology​

  1. Oncology-Specific FHIR Profiles

    • No mCODE (minimal Common Oncology Data Elements) implementation
    • Missing cancer-specific resources (TNM staging, biomarkers)
    • No genomic data handling

  2. Oncology Data Extraction

    • AI not trained on oncology reports
    • Can't extract: staging, grade, histology, biomarkers
    • No ICD-O-3 coding support

  3. Clinical Display Requirements

    • No tumor board view
    • Missing treatment timeline
    • No chemotherapy regimen display
    • No imaging integration

Path to Demo (2-4 weeks)​

Week 1: Oncology FHIR Mapping​

// Define oncology-specific extraction
interface OncologyExtraction {
  diagnosis: {
    primarySite: string;      // e.g., "Lung"
    histology: string;        // e.g., "Adenocarcinoma"
    stage: {
      clinical: string;       // e.g., "cT2N1M0"
      pathologic?: string;    // e.g., "pT2N1M0"
    };
    grade?: string;           // e.g., "G2"
  };
  biomarkers?: {
    name: string;             // e.g., "EGFR"
    result: string;           // e.g., "Positive"
    value?: string;           // e.g., "L858R mutation"
  }[];
  treatments: {
    type: string;             // "Chemotherapy", "Radiation", etc.
    regimen?: string;         // e.g., "FOLFOX"
    startDate: string;
    endDate?: string;
  }[];
}

Week 2: Enhance AI Extraction​

// Update AI prompt for oncology
const oncologyPrompt = `
Extract oncology-specific information:
1. Cancer type and primary site
2. TNM staging (clinical and pathologic)
3. Histologic type and grade
4. Biomarkers (EGFR, ALK, PD-L1, etc.)
5. Treatment history (chemo regimens, radiation, surgery)
6. Performance status (ECOG/Karnofsky)
7. Key dates (diagnosis, treatments, progression)
`;

Week 3: Create Oncology Display​

// New oncology patient view components
<OncologyDashboard patient={patient}>
  <CancerSummary />
  <TreatmentTimeline />
  <BiomarkerPanel />
  <StagingDisplay />
  <ImagingResults />
</OncologyDashboard>

Week 4: Demo Preparation​

  • Create synthetic oncology record
  • Test ingestion pipeline
  • Verify FHIR compliance
  • Polish UI display
  • Record demo video

Path to Production (3-4 months)​

Month 1: mCODE Implementation​

  • Implement full mCODE 3.0 specification
  • Add all oncology-specific FHIR profiles
  • Validate against mCODE validators
  • Test with real-world data samples

Month 2: AI Training & Validation​

  • Fine-tune AI on oncology reports
  • Add specialized extractors for:
    • Pathology reports
    • Radiology reports
    • Molecular/genetic reports
    • Treatment summaries
  • Clinical validation with oncologists

Month 3: Clinical Features​

  • Tumor board preparation tools
  • Treatment planning support
  • Clinical trial matching (basic)
  • Outcomes tracking
  • Quality measure reporting

Month 4: Integration & Testing​

  • EHR integration for oncology
  • Lab system connections
  • Imaging system integration
  • End-to-end testing
  • Security audit

Specific Gaps for Demo​

Technical Requirements​

// Current: Generic extraction
const data = await extractPatientData(document);

// Needed: Oncology-specific
const oncologyData = await extractOncologyData(document, {
  documentType: 'pathology-report',
  extractors: ['tnm-staging', 'biomarkers', 'histology'],
  terminology: 'snomed-ct'
});

FHIR Mapping Gaps​

// Current: Basic Condition
{
  resourceType: "Condition",
  code: { text: "Cancer" }
}

// Needed: mCODE-compliant
{
  resourceType: "Condition",
  meta: {
    profile: ["http://hl7.org/fhir/us/mcode/StructureDefinition/mcode-primary-cancer-condition"]
  },
  category: [{
    coding: [{
      system: "http://snomed.info/sct",
      code: "64572001"
    }]
  }],
  code: {
    coding: [{
      system: "http://snomed.info/sct",
      code: "254637007",  // NSCLC
      display: "Non-small cell lung cancer"
    }]
  },
  bodySite: [{
    coding: [{
      system: "http://snomed.info/sct",
      code: "39607008",
      display: "Lung"
    }]
  }],
  stage: [{
    summary: {
      coding: [{
        system: "http://cancerstaging.org",
        code: "3A",
        display: "Stage IIIA"
      }]
    },
    type: {
      coding: [{
        system: "http://loinc.org",
        code: "21908-9",
        display: "Stage group.clinical"
      }]
    }
  }]
}

Risk Assessment​

Demo Risks (Low-Medium)​

  • AI might miss complex staging information
  • Manual cleanup required after extraction
  • Limited to common cancer types
  • Basic display only

Production Risks (High)​

  • No clinical validation
  • Missing critical safety checks
  • No integration with oncology systems
  • Compliance gaps (HIPAA, clinical standards)

Recommendation​

For Demo (2-4 weeks)​

  1. Use simplified synthetic oncology record
  2. Focus on lung or breast cancer (well-documented)
  3. Manual verification of extracted data
  4. Basic FHIR structure only
  5. Simple UI display

For Real Usage (3-4 months)​

  1. Partner with oncology institution
  2. Implement full mCODE specification
  3. Clinical validation process
  4. Integration with oncology EMRs
  5. Compliance certification

Bottom Line​

For Sumeet's Question: We can demonstrate basic ingestion of a synthetic oncology record within 2-4 weeks, but this would be a proof-of-concept only. A production-ready system that could safely handle real oncology data requires 3-4 months of dedicated development with clinical input.

Key Decision: Is the goal to show technical feasibility (2-4 weeks) or build a clinically viable system (3-4 months)?