DI Analysis Report

Synthetic FHIR
Multimodal Ophthalmic
Dataset

Dimensional Intelligence decomposition of a linked synthetic care graph: patient, hospital, and practitioner FHIR bundles combined with fundus and OCT imaging across three format layers — FHIR Media, JPEG, and DICOM.

5 evidence layers confirmed FHIR R4 · LOINC · SNOMED · US Core Synthetic only — not clinical truth

Dataset Manifest

Patient bundle entries2,143

Observation resources905

Procedure resources238

Claim + EOB pairs155 × 2

DiagnosticReport153

Encounter105

ImagingStudy16

Media (embedded img)12

DICOM filesMultiple

ClassificationConfidential

Dataset Summary

Bundle entries

2,143

Patient longitudinal chart · life-course from birth to 2024

Imaging assets

16+ studies

12 Media resources · standalone DICOM · JPEG renders

Billing artifacts

310

155 Claim + 155 EOB · end-to-end reimbursement testable

DI evidence layers

Identity · chronology · imaging · provider · billing

Multimodal Evidence Dimensions

D1 · FUNDUS

2D Surface

R+2

Vessel structure, optic disc, macula. 1024×1024 JPEG. Minor acquisition artifact (central shadow). Sufficient for screening — not surgical fidelity.

D2 · OCT

Cross-Sectional

R+3

Retinal layer integrity, foveal contour, depth information. 512×512 JPEG. Higher diagnostic signal than fundus — critical differentiator.

D3 · FHIR

Structured Data

High

Diagnoses, coded observations, DiagnosticReport, billing. LOINC + SNOMED + US Core compliant. Standards-oriented synthetic data.

D4 · PROVIDER

Attribution

Weak*

Org linkage, referral chains, NPI-style IDs. *NUCC specialty codes unreliable — many coded as General Practice despite ophthalmic org names. Synthetic artifact.

DI non-interchangeability rule: These dimensions are never conflated. Fundus ≠ OCT ≠ structured report ≠ coded observation ≠ billing artifact. Cross-modal output must declare one of three states: Agreement · Disagreement · Unresolved. Collapse is the failure mode this system is built to prevent.

Care Graph — Concept Map

Temporal Decomposition

Chronic disease progression timeline — life-course longitudinal synthetic data

1987

Prediabetes

Initial metabolic anomaly — begins disease trajectory

2020

Type 2 Diabetes Mellitus

Confirmed · annual ophthalmic surveillance begins

2021

Neuropathy due to Type 2 DM

Systemic complication — confirms progression pattern

2023

Nonproliferative Diabetic Retinopathy (NPDR)

Due to Type 2 DM · ophthalmic complications established

2024-05-17

Complete Multimodal Event Cluster

Encounter → Procedure (exam + OCT) → ImagingStudy → DiagnosticReport → Observation (left/right OCT + retinopathy severity) → Claim + EOB. Full five-layer record in a single date.

Cross-Modal Consistency — 2024-05-17

Evidence signal by modality

OCT observations (structured LOINC coded) No abnormal findings

Fundus imaging (visual assessment) No strong pathology detected

Structured FHIR diagnosis (SNOMED coded) Mild NPDR confirmed

DI output — correct

"Mild retinopathy present per structured clinical data; imaging does not show strong advanced features — consistent with early-stage NPDR below JPEG resolution threshold."

✕ "No disease" — image-only collapse error
✕ "Disease confirmed visually" — overclaim from image alone

Four-dimensional confidence scoring

Structural confidence

DICOM UID → ImagingStudy; Practitioner → Encounter performer; Media → DiagnosticReport

Clinical signal confidence

Fundus image quality score; OCT availability; report/observation agreement

Temporal confidence

Procedure before billing; imaging before interpretation; diabetes onset before retinopathy

Attribution confidence

Capturing / ordering / interpreting provider per imaging study; facility of service

Composite score produced only after all four sub-scores are resolved

Deterministic Ingestion Pipeline

Competitive Differentiation

Approach	What they output	DI / MosaicDM
Image AI only	"No disease" — JPEG classifier result, FHIR diagnosis ignored	Cross-modal reconciliation: image + coded report + structured finding
FHIR ETL only	"Disease present" — structured data alone, no imaging context	Full provenance graph with modality-separated evidence at every node
Clinical dashboard	Shows both layers with no reconciliation logic — user resolves disagreement	Deterministic consistency engine: Agreement / Disagreement / Unresolved — explicitly labeled
Note summarization AI	Summarizes text — loses imaging and billing context entirely	Bounded conclusions scoped to confirmed signal strength across all four dimensions

Core differentiator: Deterministic graph assembly + modality-separated evidence handling + traceable provider attribution + bounded conclusions. Every competitor collapses dimensions. DI is the only framework built to prevent that collapse. This is not an ophthalmology AI — it is a domain-agnostic deterministic multimodal clinical reasoning engine whose first domain is ophthalmology.

DI Guardrails — System Level

Information Conservation

Do not infer disease severity beyond available signal. A single JPEG cannot support strong disease claims. Signal scope determines conclusion scope — always. The image is one evidence dimension, not the arbiter.

Dimensional Coherence

Maintain hard separation between modalities: 2D fundus image, cross-sectional OCT, structured coded findings, clinical note text, billing artifacts. These are different dimensions of evidence — not interchangeable representations of the same truth.

Evidence Boundary

Cap conclusions at the strongest confirmed layer. When image suggests mild findings, OCT report says no abnormal findings, and structured diagnosis says mild NPDR — the output must preserve the disagreement rather than collapsing to a single label. Disagreement is data.

Client-Facing Output Schema

"patient_id": "...",
"episode_id": "...",

"modalities_present": ["fundus_jpeg", "dicom", "fhir_media", "observations", "diagnostic_report", "claim"],

"provider_context": {
  "performing_org": "HARBORVIEW MEDICAL CENTER",
  "linked_practitioners": [...],
  "attribution_confidence": 0.84
},

"imaging_summary": {
  "fundus_quality": "usable",
  "dicom_present": true,
  "oct_present": true,
  "laterality_resolved": ["left", "right"]
},

"clinical_summary": {
  "structured_findings": [...],
  "report_findings": [...],
  "image_only_findings": [...]
},

"consistency_check": {
  "agreement": [...],
  "disagreement": ["mild NPDR in structured data; not confirmed visually"],
  "needs_review": [...]
},

"trace": {
  "source_files": [...],
  "resource_ids": [...],
  "linkage_paths": [...]
}

DI Utility Bounds

High-value applications

Validating multimodal FHIR ingestion pipelines

Building patient-centric graphs across encounters, studies, images, and billing

Testing image extraction and indexing from FHIR Media

OCT/fundus linkage to structured FHIR findings

Simulating diabetic retinopathy longitudinal surveillance workflows

Validating joins between patient bundles and provider/location registries

Building demo dashboards and synthetic sandboxes — no PHI risk

Proving the DI cross-modal consistency engine approach

Do not use for

Real-world epidemiologic truth or disease prevalence inference

Clinical performance claims without external validation

Specialty routing based on NUCC codes — synthetic artifact, unreliable

Ground-truth imaging assessment — synthetic generation inconsistencies present

Causal medical conclusions beyond what DI analysis confirms

Production diagnostic performance benchmarking

Confirmed synthetic data — not ontologically equivalent to real care data

Synthetic FHIR Multimodal Ophthalmic Dataset