Research

---

Claims Sciences develops novel machine learning techniques to aid in the detection of Medicare Fraud, Waste, and Abuse with a specific focus on provider utilization. Our current research focus centers around identifying abnormal provider behavior. 

We hypothesize that providers who later face enforcement or exclusion exhibit detectable patterns in their billing behavior well before those actions occur. Our tools flag these providers long before traditional channels would detect such behaviors, stopping FWA nearly as soon as it begins.

Claims and utilization data

• Medicare claims feeds used by Accountable Care Organizations (e.g., CCLS, BCDA)
• Provider-level utilization and payment data aggregated by NPI (MPUPD)

Reference enforcement data

• Office of Inspector General (OIG) List of Excluded Individuals and Entities (LEIE)
• Used to validate whether anomalous provider behavior correlates with known enforcement outcomes

This combination allows us to study provider behavior longitudinally and test whether statistical anomalies precede formal enforcement actions.

Step 1: Data normalization

• Standardize specialty, credential, and taxonomy fields
• Clean and normalize categorical provider attributes

Step 2: Deterministic matching

• Join claims, utilization, and enforcement datasets on exact NPI matches
• Maintain role-based distinctions where available (rendering, billing, ordering)

We convert raw claims data into structured provider behavior profiles.

Each provider is represented by a set of statistical and behavioral features, including:

Utilization patterns

• Services per beneficiary
• Unique procedures billed
• High-intensity service concentration

Cost behavior

• Charge and payment distributions
• Extreme outlier spending patterns
• Procedure-level cost concentration

Coding behavior

• Abrupt changes in procedure mix
• Specialty-incongruent billing
• Unusual code combinations

Temporal dynamics

• Sudden growth in volume or charges
• Short-term spikes in utilization

Categorical variables such as specialty or taxonomy are encoded after normalization to preserve provider context.

We employ an agentic approach to FWA detection, as we believe there isn't a one-size-fits-all solution to the difficult problem of FWA detection. An agent is tasked with provider evaluation and has a myriad of tools available to assist in realistic, logic based anomaly detection.  Think of the agent as an investigator with intimate knowledge of all billing codes, Medicare rules and regulations, provider history, and knowledge of thousands of fraud schemes. 

Each provider rigorously examined through the lens of mathematical anomaly detection

• Top contributing anomaly features
• Peer comparison within specialty and geography
• Procedure-level drill-downs

This first step allows the agent to quickly assess whether a case warrants further investigation and returns a risk score. 

After a provider is flagged for further investigation, the agent uses a self-contained foundation model to determine a plain language explanation for the behavior and creates a case for human decision makers which includes

• Confidence score 
• Suspected pattern (Fraud, Waste, Abuse, or some combination)
• Explanation of pattern and suggested next steps

We use two complementary validation methods.

Retrospective validation

• Compare anomaly scores against providers later appearing on OIG exclusion lists
• Measure whether flagged providers are disproportionately represented among known enforcement cases

Operational validation

• Review flagged providers with ACO analysts and compliance teams
• Measure:
  • Actionable case rates
  • Time-to-detection improvements
  • Dollars at risk identified per analyst hour

Our methodology is designed to support compliance teams, not replace them.

Known limitations

• Enforcement datasets do not capture all fraud or improper behavior
• Some attaching and detaching patterns in exclusion lists are not fully observable over long horizons
• Aggregation at the provider level may obscure individual claim-line issues

Operational controls

• Anomaly flags are treated as investigation leads, not determinations of wrongdoing
• All flagged cases require chart review or compliance workflows
• Models are monitored and retrained to account for evolving billing patterns

Accountable Care Organizations receive regular claims feeds and bear direct financial risk for abnormal spending. Our methodology is designed to integrate into existing ACO workflows by:

• Ranking providers by utilization behavior
• Prioritizing investigation queues
• Identifying high-risk patterns months earlier than retrospective SQL-based reviews

This enables ACOs to intervene earlier, protect shared savings, and reduce compliance risk.