Detecting Fraud Through Document Intelligence

How FHIR Data, Embeddings, and Graph Analysis Help Agencies Identify and Prevent Fraud

Fraud inside government programs is rarely a single event. It emerges from subtle inconsistencies across medical records, claims, provider histories, and long‑running case files. And because every agency — and every enterprise — experiences fraud differently, there is no universal model or one‑size‑fits‑all detection strategy.

Fraud is contextual. Fraud is behavioral. Fraud is local to the organization experiencing it.

This is why modern NLP systems must do more than read documents — they must help agencies understand the meaning behind their structured data and analyze it at scale. Corpus Crystal was built for this reality.

Structured Data as the Foundation: FHIR and Beyond

Unlike generic NLP platforms that rely on probabilistic entity extraction, Corpus Crystal works directly with structured data, especially FHIR.

The platform can:

• Ingest FHIR bundles
• Render them into human‑readable PDFs
• Auto‑annotate every element with full traceability back to the original FHIR structure
• Preserve relationships between patients, providers, encounters, diagnoses, and procedures

This creates a unified view of the case: the human‑readable document and the structured data behind it, perfectly aligned.

For fraud detection, this alignment is essential. It ensures that every claim, diagnosis, provider reference, and encounter is grounded in verifiable, structured truth.

Using Embeddings to Capture Meaning, Not Just Text

Fraud patterns rarely appear as simple keyword matches. They emerge from semantic similarity and behavioral patterns across cases.

Corpus Crystal generates embeddings for:

• FHIR elements
• Provider histories
• Encounter summaries
• Claim narratives
• Supporting documents
• Case‑level metadata

These embeddings capture meaning, not just words. They allow agencies to cluster similar cases, identify unusual patterns, and detect outliers that deviate from normal behavior.

Examples include:

• Providers whose encounter notes are semantically similar across unrelated patients
• Claims whose narratives differ significantly from typical cases with the same diagnosis
• Encounter patterns that don’t match expected clinical pathways
• Clusters of cases with unusually similar supporting documentation

Embeddings turn unstructured and semi‑structured data into comparable signals.

Graph Databases: Connecting the Dots Across Cases

Fraud is often a network problem. Providers, patients, addresses, diagnoses, and timelines form relationships that only become visible when analyzed as a graph.

Corpus Crystal exports structured data and embeddings into graph databases such as:

• Neo4j
• Amazon Neptune
• Azure Cosmos DB (Gremlin)
• TigerGraph

This enables agencies to build:

• Provider networks
• Patient‑provider relationships
• Temporal sequences of encounters
• Clusters of similar claims
• Cross‑case linkages

Graph analysis reveals:

• Providers connected to unusually high volumes of similar claims
• Patients appearing across multiple unrelated provider networks
• Encounter patterns that deviate from expected clinical pathways
• Reused documentation patterns across cases

This is where fraud rings, coordinated activity, and systemic anomalies become visible.

Offline Aggregate Analysis: Building Enterprise‑Specific Fraud Models

Because fraud varies dramatically between agencies, the most effective models are enterprise‑specific.

Corpus Crystal supports this by enabling agencies to:

1. Export structured data + embeddings + relationships
2. Run offline analysis across millions of cases
3. Build custom models that reflect their unique fraud landscape

These models can include:

• Clustering to identify unusual case groups
• Outlier detection on provider or claimant behavior
• Similarity scoring to detect reused narratives or templated documentation
• Temporal anomaly detection for suspicious encounter sequences
• Graph‑based risk scoring for providers or networks

The result is a fraud‑detection strategy that is tailored to the agency, not borrowed from another domain.

Hypothetical Use Cases: How Agencies Could Apply These Capabilities

The following examples illustrate how agencies could use Corpus Crystal’s structured‑data alignment, embeddings, and graph‑based analysis to support fraud detection and prevention.

Social Security Administration (SSA)

SSA processes enormous volumes of medical evidence, disability claims, and supporting documentation. Fraud in this domain often appears as:

• Repeated medical narratives across unrelated claimants
• Providers submitting templated encounter notes
• Inconsistent timelines between medical evidence and claimant statements
• Clusters of claims tied to the same provider or representative

Using Corpus Crystal, SSA could:

• Align FHIR medical records with human‑readable evidence
• Cluster similar cases to identify unusual patterns
• Detect outliers in provider behavior
• Build graph‑based models to identify coordinated activity
• Support analysts with retrieval‑driven chat to explore evidence quickly

Centers for Medicare & Medicaid Services (CMS)

CMS fraud often involves:

• Upcoding
• Phantom billing
• Unnecessary procedures
• Provider networks coordinating claims
• Reused documentation across patients

With Corpus Crystal, CMS could:

• Analyze FHIR claims and encounter data at scale
• Identify semantic similarities in provider documentation
• Build provider‑centric graphs to detect unusual referral or billing patterns
• Cluster claims to find outliers in cost, diagnosis, or treatment patterns
• Support auditors with transparent, traceable evidence trails

A Partnership Approach: Interactive Consulting Services

Fraud detection is not solved by a single model, a single workflow, or a single platform. It requires a deep understanding of each agency’s data, processes, and operational realities. That’s why Interactive Consulting Services pairs Corpus Crystal with hands‑on collaboration.

Our team works with agencies to:

• Understand their unique fraud landscape
• Identify the structured data sources that matter most
• Build embedding strategies tailored to their domain
• Design graph‑based analysis pipelines that reflect real investigative workflows
• Develop offline models that evolve with the agency’s needs
• Integrate findings back into day‑to‑day casework

Fraud is intimate to each enterprise — and the solutions must be as well. ICS brings the technical expertise, the domain understanding, and the collaborative approach needed to help agencies not only detect fraud, but prevent it, explain it, and operationalize the insights at scale. Our staff is ready to partner with you, work alongside your teams, and help you turn your document ecosystems and structured data into actionable intelligence.