Cover Page
●
●
●
●
●
Team name
Submission title
Name and address (address city, state, zip) of the institution(s), company(ies), or
organization(s) represented (must be U.S.-based entities)
Team lead information including name, title, address, email, and phone number
Additional team members' names, titles, addresses, emails, phone numbers, and
whether each team member meets the competition eligibility requirements. Each team
must be a group of two to five individual participants (see eligibility requirements)
PRESENTS
Tunable Large Relational Model for CMS
Payment Integrity, Efficiency, and
Accountability (CMS-LRM-IEA)
A Practical, Explainable AI Approach to Crushing Fraud, Waste, and Abuse in Medicare Feefor-Service
Ted Nadeau, Chief AI Officer & CMS Team Lead ted@polyviewhealth.com
Tejaswini Ashok, Lead AI Engineer - tejaswini.ashok@polyviewhealth.com
Greg Deocampo, Chief Technology Officer - GD@polyviewhealth.com
Dimitri Arges, Co-Founder, FWA expert - darges@polyviewhealth.com
Jason Hwang. Co-Founder, Physician - jhwamh@polyviewhealth.com
1000 Marsh Road, Menlo Park, CA 94025
609-915-6928
Goals of this document: ( this section will not appear in final submitted
document)
Phase 1 deliverable: a clear 10-page PDF white paper by 9/19/25 ( 9/16/2025 to get
feedback )
1. >= 1-inch margins, no more than 10 pages ( excluding cover page & reference page ),
body font size >= 11 point arial, fonts in tables and graphs >= 8 point arial
2. A proposed explainable AI/ML approach for detecting anomalous data patterns that
may be indicative of fraud in Medicare Fee-for-Service (FFS) claims data.
3. A detailed explanation of how model outputs will be translated into actionable and
transparent indicators of fraud.
4. A clear description of how the proposed solution can be scaled to detect systemic
vulnerabilities and how analytic insights will inform policy and operational changes
necessary to support broad adoption.
Judging Criteria:
Phase 1 criteria (20% each):
1. Relevance & Impact
2. Innovation & Technical Merit
3. Explainability & Interpretability
4. Feasibility & Implementation Potential
5. Clarity & Quality.
Phase 2 criteria (25% each):
1. Technical Merit
2. Explainability
3. Feasibility,
Clarity
Address:
"2022-2024 Limited Data Sets (LDS) data containing Medicare FFS Hospice, Part B, and DME
claims for a random 5% sample of Medicare beneficiaries"
https://www.cms.gov/Regulations-and-Guidance/Guidance/Manuals/Downloads/bp102c09.pdf
1. Executive Summary
Fraud, Waste, and Abuse (FWA) in Medicare and Medicaid represent a persistent and costly
challenge. Each year, tens of billions of taxpayer dollars are lost to improper payments,
undermining the financial sustainability of the programs and eroding public trust.
Existing detection systems—rules-based engines and standard machine learning models—are
limited. They struggle to adapt to evolving fraud tactics, often overwhelm investigators with false
positives, and produce opaque results that investigators cannot easily interpret.
This paper introduces Polyview Health’s CMS Large Relational Model for Integrity,
Efficiency, and Accountability (CMS-LRM-IEA). It is a layered fraud detection platform that
combines robust, well-understood tools with advanced AI techniques. The design ensures that
CMS can achieve measurable improvements quickly while building toward more sophisticated
capabilities over time.
The solution offers:
●
Practicality: A layered architecture that starts with proven rules and gradient boosting,
and adds a powerful Large Relational Model (LRM) that fuses Variational Autoencoders
(VAEs) and Heterogeneous Graph Neural Networks (GNNs).
●
Explainability: Plain-language outputs investigators can trust, with optional advanced
visualizations.
●
Scalability: A phased rollout from pilots to national deployment, built on cloud-ready
infrastructure.
●
High Return on Investment: Even modest improvements—such as a 1% increase in
detection accuracy—can save CMS over $1 billion annually.
●
Enterprise AI interfaces powered by Polyview Health Focus
This approach provides CMS with both the reliability of a pickup truck and the power of a
Formula 1 engine, aligning advanced AI with CMS’s mission to promote Integrity, Efficiency,
and Accountability—the direct opposites of Fraud, Waste, and Abuse.
2. Background and Problem Statement
2.1 The Scale of the Problem
The Office of Management and Budget (OMB) estimates that improper Medicare and Medicaid
payments reach tens of billions of dollars annually. Fraudulent behaviors include upcoding,
phantom billing, provider collusion, and misuse of emerging billing categories such as
telehealth. These losses drain program resources, divert funds from patients, and weaken
public confidence.
In particular we can focus on two topics: Hospice Fraud and Durable Medical Equipment (DME)
Fraud.
2.1.1 - Hospice Fraud Schemes
Fraudulent practices—particularly around the 180-day hospice recertification checkpoint—erode
trust and inflate costs. Schemes include falsifying documentation, enrolling non-eligible patients,
creating phantom patients, or ignoring improvements in patient condition.
• Falsified eligibility: Billing Medicare for patients past the 180-day recertification threshold who
no longer meet the criteria for terminal illness.
• Enrollment of chronic/non-terminal patients: Targeting dementia or long-term care patients
whose prognosis exceeds six months.
• Phantom patients: Creating false hospice entities or billing for services never provided.
• Failure to discharge improved patients: Continuing claims for patients who have recovered
beyond a terminal prognosis.
2.1.2 - DME Fraud
DME fraud commonly includes phantom rentals, upcoding, and patient kickbacks.
2.2 Limitations of Current Approaches
1. Statistical and Rules-based detection
Rules can effectively capture known fraud patterns, but they are brittle. As fraud tactics
evolve, rules require constant updates, creating costly maintenance and leaving the
system vulnerable to novel schemes.
2. Tabular & deep machine learning models
Techniques such as logistic regression or gradient boosting offer incremental
improvement over rules, but they treat claims as isolated rows of data. They cannot
detect relational anomalies, such as collusive provider networks. Their scores are often
opaque, producing “black-box” outputs that investigators cannot easily interpret.
3. Operational burden
High false positive rates overwhelm investigators, consuming limited resources. False
negatives allow systemic vulnerabilities to persist.
The gap is clear: CMS requires adaptive, relational, and explainable fraud detection that
integrates seamlessly into investigator workflows and supports policy refinement.
3. A Layered Fraud-Detection Solution
Our solution combines a robust baseline layer with an advanced relational AI layer, ensuring
both immediate practicality and long-term innovation.
3.0 External Data
The provider information will be compared to existing public data stores such as:
● CMS Open Payments
● OIG Exclusion Database
● OFAC Exclusion Database
● SAM Exclusion Database
● State Medical Board Provider Sanctions List
3.1 Baseline Layer: Reliable and Familiar
The first layer implements baseline standard Anomaly Detection techniques per
https://www.fraud.com/post/anomaly-detection
And integrates a rules engine. This layer provides CMS with:
●
A familiar, low-risk approach that is easy to adopt and maintain.
●
Transparent, investigator-friendly outputs suitable for immediate operational use.
●
A foundation that alone delivers measurable improvements in fraud detection.
This will form a base layer for validation & augmentation
3.2 Advanced Layer: Large Relational Model (VAE-GNN)
The innovation lies in the second layer, the Large Relational Model (LRM), which fuses
Variational Autoencoders and Heterogeneous Graph Neural Networks.
●
Variational Autoencoders (VAEs):
VAEs model the distribution of “normal” claims without labels. They identify anomalies
through posterior variance thresholds, KL divergence monitoring, reconstruction error,
and combined measures such as negative ELBO scores.
●
Heterogeneous Graph Neural Networks (GNNs):
GNNs capture complex relationships across entities such as patients, providers,
facilities, diagnosis codes, procedure codes, geography, and time. They reveal collusion
patterns, unusual referral loops, and regional anomalies invisible to tabular methods.
●
Fusion Scoring:
VAE anomaly signals and GNN embeddings are combined into a Suspicion Score,
which can be tuned by jurisdictional policy and investigative capacity.
This layered approach ensures CMS is not forced to choose between reliability and innovation.
The baseline delivers immediate value, while the LRM extends detection power and
adaptability.
( Diagram showing entities & relationships: Patient-Beneficiary to Claim Header to ( Claim Line to
Procedure ) and to [ Provider, Physician, Diagnosis ] )
PatientBeneficiary
Claim Header
Claim Line
Role
Provider
Role
Physician
Diagnosis
Procedure
3.3 Machine Reasoning
We will implement Machine Reasoning based on CMS Guidelines ( for instance: “Medicare
Benefit Policy Manual - Chapter 9 Coverage of Hospice Services Under Hospital Insurance Rev 13133; Issued 03-20-25 ) - https://www.cms.gov/Regulations-andGuidance/Guidance/Manuals/Downloads/bp102c09.pdf
This will transform the document into core logic-bearing concepts compatible with Satisfiability
Modulo Theories ( https://en.wikipedia.org/wiki/Satisfiability_modulo_theories ). This logic will
then be applied to the source data to validate the claim.
3.4 Learning & Continuous Improvement ( editor: move this to later in the
document )
In the early phase, the majority of the training will be ‘unsupervised’ anomaly detection.
Over-time, we will be able to add labeled data which will then enable ‘supervised’ training &
improve the Machine Learning Model.
Additionally, other ‘rules and guidelines’ can be defined that will augment & change the Machine
Reasoning rule-set.
4. Explainability and Interpretability
Fraud investigators are experts in billing and policy, not neural networks. A successful solution
must therefore present results in a way that investigators can trust and act on quickly.
4.0 Claim Adjudication Notation
The system will produce Claim Adjudication Reason Notation compatible with RARC, CARC
( Diagram showing Flagged & Tagged Claim )
Claim
ClaimAdjudic
ation
StayClaim
Stay
ClaimAdjudic
ationReason
RARC Code
CARC Code
4.1 Plain-Language Explanations
The system prioritizes narrative explanations before technical visualizations. Examples include:
●
“Provider A billed for Procedure X at a rate 300% higher than peers in the same county.”
●
“Patient B visited 12 providers across three states in a single month—well above typical
patterns.”
4.2 Transparency Features
For investigators who wish to dive deeper, the system provides:
●
Heatmaps highlighting the features that drove the anomaly score.
●
Network diagrams showing unusual provider-patient clusters.
●
Feature attribution summaries in plain English.
●
Audit trails documenting thresholds, model versions, and decision paths.
This dual approach—plain-language summaries with optional technical detail—ensures
accessibility for all users.
4.3 Chat Interface
We will enable a ‘chat’ interface which will enable the users to Chat with the data and with the
system
The system will also produce candidate ‘Payment Guidelines’ and ‘Clinical Guidelines’
5. User Workflow and Dashboard
The effectiveness of any fraud detection platform depends not only on its accuracy but on its
ability to improve investigator efficiency.
5.0 Integration w/ existing systems
Our proposed system will augment the existing CMS Fraud Protection Systems ( FPS & FPS2 )
and continue to move away from "pay and chase" to implement additional pre-payment edits &
post-payment models that generate alerts and prioritize claim-items and claim-groups for further
review.
https://www.cms.gov/files/document/dasg-leaflet-fps2.pdf
5.1 Dashboard Features
The proposed dashboard provides:
●
Ranked alerts prioritized by potential ROI (dollar impact × suspicion score).
●
Claim summaries with ID, provider, suspicion score, rationale, and denial codes.
●
Drill-down views for deeper exploration of anomaly patterns.
●
Search and filtering by provider, region, or billing code.
5.2 Workflow Integration
The workflow mirrors investigator processes:
1. Receive ranked alerts.
2. Review plain-language rationale.
3. Drill down into supporting data if desired.
4. Take action: denial, referral, or provider education.
This design minimizes time spent parsing raw data and maximizes time spent on resolution.
6. Business Case and ROI
6.1 Quantifiable Impact
●
With Medicare and Medicaid improper payments estimated near $100 billion annually,
even modest gains have outsized impact.
●
A 1% improvement in detection translates to $1 billion in annual savings.
6.2 Pilot Example
In a $5 billion claims category such as telehealth:
●
A 5% improvement in fraud detection yields $250 million in recoveries.
●
Reducing false positives by 20% frees substantial investigator capacity.
6.3 Broader Value
●
Improved morale among investigators.
●
Data-driven policy design that prevents fraud before it occurs.
●
Reinforced public trust through visible accountability.
7. Feasibility and Scalability
7.1 Practical Implementation
●
Batch mode pilots in high-yield categories demonstrate ROI with minimal disruption.
●
Cloud-native architecture using Spark, Dask, and PyTorch Geometric ensures
scalability.
●
Hybrid storage models (relational + graph databases) support complex queries.
●
Pre-training with lightweight inference enables near-real-time scoring without
excessive compute cost.
7.2 Phased Rollout
1. Pilot programs: Select states or high-value claim types.
2. State-level expansion: Adjust thresholds and workflows to capacity.
3. National deployment: Integrate with existing claims pipelines, dashboards, and
training.
This phased strategy allows CMS to scale confidently while ensuring stability.
8. Detecting Systemic Vulnerabilities
Beyond identifying individual fraudulent claims, the model highlights systemic vulnerabilities:
●
Collusion clusters among providers.
●
Overutilization of codes following policy changes.
●
Geospatial hotspots of unusual activity.
●
Temporal anomalies in billing patterns.
These insights enable CMS not only to detect fraud but also to strengthen policy through
targeted edits, provider education, and regulatory adjustments.
9. Governance and Monitoring
To ensure integrity and compliance, the solution incorporates:
●
Regular retraining cycles (e.g., quarterly or as drift is detected).
●
Threshold calibration tailored to jurisdictions.
●
Investigator feedback loops to refine outputs.
●
Comprehensive documentation of model versions and decision criteria for auditability.
10. Conclusion
Fraud, Waste, and Abuse represent one of CMS’s most persistent challenges. Traditional
methods—rules and tabular models—are no longer sufficient.
The CMS-LRM-IEA solution combines the reliability of established methods with the innovation
of advanced relational AI. It provides:
●
Practicality through incremental adoption.
●
Innovation through VAEs and GNNs.
●
Explainability via plain-language outputs.
●
Scalability through phased rollout.
●
High ROI with billions in potential annual savings.
By aligning with CMS’s mission and embedding the positive opposites of FWA—Integrity,
Efficiency, and Accountability—this solution offers a clear, actionable pathway to
safeguarding trust and resources in Medicare and Medicaid.
References
? to customers?
Links
Anomaly Detection, Fraud, Variational AutoEncoders, Graph Neural Networks, Machine
Reasoning,
Introduction to Anomaly Detection with Python
https://www.geeksforgeeks.org/machine-learning/introduction-to-anomaly-detection-with-python/