Mukundan Srinivasan, Peter L. Levin
Amida Technology Solutions, Inc.
January 2026
Executive Summary
Medicare handles over $1 trillion in annual claims, but its conventional fraud detection methods sometimes fail to identify sophisticated schemes that hide within the data. Amida addresses this challenge with a graph-based framework that transforms semantic data into a visual map of interconnected entities and behaviors.
Our approach reveals discrepancies that appear legitimate in isolation but may be anomalous in context. This methodology enables payers to adopt a proactive defense and prioritize high-impact cases for human investigators. Amida provides a scalable, explainable defense against the evolution of organized fraud; we were a finalist in the Centers for Medicare & Medicaid Services (CMS) Crushing Fraud Chili Cook-Off (top ten of 259 submissions).
The Challenge: Identify Fraud Invisible to Traditional Detection
Medicare manages over $1 trillion in annual outlays for nearly 70 million beneficiaries. The sheer volume of transactions makes it difficult to identify and block improper payments that legitimate transactions can conceal.
Traditional fraud detection relies on static rules to identify inconsistencies. It fails to detect sophisticated schemes because criminals operate through intricate networks of interconnected behaviors, entities, and dynamic historical patterns. In this short white paper, we focus on fraud in healthcare, however, our approach applies to fraud, waste, and abuse (FWA) across claims-based payment systems in the public and private sectors.
The Graph-Based Approach: A Geometric Framework
Amida uses an innovative technique that maps raw data into a graph, and then transforms it into a “vector space.” This transformation renders semantic relationships as geometric structures, and our framework systematically and reproducibly groups claims into “likely valid” and “potentially invalid” clusters.
In this model, nodes represent entities such as participants, transactions, and service locations. Edges represent the relationships between them, including shared addresses, common financial accounts, and recurring interaction patterns. The graph captures the entire relational context of an event: a transaction that appears legitimate in a table might expose fraud risk when its geometric proximity to known unusual or suspicious entities become visible.
The model applies an embedding technique that translates structural relationships into numerical coordinates. Embedding serves as the foundation for detection and explainability. The process interprets explicit connections – the frequency of interaction, the strength of a bond, and transactional constraints – as spatial relationships. The distance between embedded points reveals the strength and relevance of their tabular connections. This transformation enables inference and anomaly detection in a new, effective, and scalable way.
Applicability for Real-World Use Cases
Standard pattern detection identifies certain types of outliers, but graph-based transforms reveal the underlying drivers (the hidden connections) behind suspicious activity. This AI-based strategy provides context by converting raw data into evidence-based indicators that investigators can trust. This novel approach supports human-in-the-loop oversight and enables program integrity teams to provision precious investigative resources in a systematic, methodical, and fair way.
In August 2025, CMS initiated its Crushing Fraud Chili Cook-Off. This market-based challenge tasked participants to develop explainable artificial intelligence (AI) and machine learning (ML) models that detect anomalies in complex Medicare claims data. Ten finalists from 259 competitors were announced in November, and Amida made the cut.
Through our evaluation, we uncovered clusters with abnormally high payment levels. Our report flagged these clusters as potential indicators of overpayment.
Value Proposition: Turn Patterns into Action
We classify our findings into easy-to-understand clusters. Our methodology identifies the following:
- High-priority claims that warrant investigation
- Peripheral claims that indicate key differences within a cluster
- Cluster-level insights that reveal suspicious networks or behavior types
These outputs enable human investigators to validate cases, discover new fraud typologies, and refine internal controls. Outliers identified through this process may also expose providers that overcharge for the same service. Over time, insights derived from clusters can train automated detection systems and investigative staff.
Conclusion
Amida’s graph-based approach redefines fraud detection through the conversion of complex semantic relationships into actionable quantitative insights. This framework provides a scalable, explainable, and forward-looking defense against the changing landscape of organized FWA. The successful application of this model to Medicare Fee-for-Service (Phase 2 of the CMS Crushing Fraud Chili Cook-Off) data demonstrates its readiness for large-scale operational use. As fraud tactics increase in sophistication, this methodology offers program integrity teams a roadmap to stay ahead of the curve, and protect vital (and scarce) resources through superior data intelligence and human-in-the-loop oversight.
For more information on how to use graph-based AI/ML to enhance your organization’s fraud detection capabilities, please contact Amida at [email protected] or (202) 735-1798.
