Combating Healthcare Fraud with MST

This is an overview of how MST fights healthcare collusion and fraud. We’ve discovered healthcare fraud is prevalent, but is difficult to detect with typical rule-based systems.

MST’s Vectorized Knowledge Graph

  • Data Ingestion and Entity Extraction: On an ongoing basis, we ingest massive amounts of diverse healthcare data, including claims data, patient records, provider information, medical guidelines, regulatory policies, research papers, and even unstructured clinical notes. Our LLM, trained on medical terminology and concepts, is crucial here to extract entities (e.g., patients, providers, procedures, diagnoses, medications, facilities, insurance plans) and relationships (e.g., “provider treated patient for diagnosis,” “medication prescribed for condition,” “patient seen at facility”) from both structured and unstructured data.
  • Vectorization: The extracted entities and relationships are transformed into high-dimensional numerical representations (vectors). This vectorization allows for semantic similarity searches and captures complex relationships that traditional rule-based systems often miss. Similar entities or relationships will have similar vector representations.
  • Graph Construction: These vectorized entities and relationships form the nodes and edges of the knowledge graph. Our graph isn’t just a static database; it’s a dynamic, interconnected network where every piece of information is linked to relevant contexts.

2. Fraud Detection and Collusion Identification:

  • Pattern Recognition and Anomaly Detection:
    • Graph Algorithms: We apply various graph algorithms to traverse the knowledge graph. This includes algorithms for community detection (finding groups of colluding providers or patients), pathfinding (tracing suspicious referral patterns), and centrality analysis (identifying key players in fraudulent networks).
    • Vector Similarity Search: By comparing the vectors of new claims or activities to known fraudulent patterns or even to normal, legitimate patterns, we can quickly identify anomalies. For example, if a provider’s billing patterns (vectorized) suddenly deviate significantly from their historical, legitimate patterns, it flags a potential issue.
    • Multi-hop Reasoning: Our knowledge graph allows for “multi-hop” reasoning, meaning it can analyze indirect connections. For instance, it can identify collusion between two providers who never directly interact but frequently refer patients to a shared, questionable laboratory.
    • Rule-based Integration: While leveraging advanced AI, we also integrate traditional rule-based fraud detection. These rules (e.g., “billing for deceased patients,” “excessive services for a diagnosis”) are encoded into the knowledge graph’s logic, and the LLM helps interpret and apply these rules to new data.
  • LLM for Contextual Understanding and Explanation:
    • Hypothesis Generation: When the knowledge graph flags a suspicious pattern, the LLM analyzes related unstructured text (e.g., doctor’s notes, patient complaints, insurance appeals) to generate hypotheses about the nature of the potential fraud. It understands the nuances of medical language that might indicate fabricated conditions or unnecessary procedures.
    • Explainable AI: A key benefit of combining our LLM with a knowledge graph is explainability. The LLM can articulate why a particular claim or set of activities is flagged as suspicious by referencing specific entities, relationships, and rules within the knowledge graph, providing auditability and supporting human investigators.
    • Querying and Augmentation: When an investigator has a specific query (e.g., “Show me all providers who billed for procedure X more than 10 times in a week, and also had common patients with provider Y”), the LLM can interpret this natural language query and use the knowledge graph to retrieve relevant, grounded information
    • Summarization and Reporting: Once potential fraud is identified, our system generates concise summaries and reports for human review, highlighting the key evidence and connections.

Our Results:

  • Increased Fraud Detection Rate: MST provides a significant uplift in the identification of fraudulent claims and collusive networks compared to traditional methods. We’ve seen a 30-50% improvement in detection accuracy or volume of identified fraudulent activities.
  • Reduced False Positives: By providing richer context and enabling more nuanced reasoning, the system drastically reduces the number of false positives that human investigators need to review. Roughly a 20-40% reduction in false positive rates, saving significant time and resources.
  • Faster Detection and Prevention: The ability to identify fraud before claims are paid (pre-payment review) or very soon after submission. This typically leads to a reduction in payment of fraudulent claims by 15-30%.
  • Discovery of New Fraud Schemes: Our system has the ability to uncover complex, previously unknown patterns of collusion and emerging fraud schemes that are too intricate for human analysts or simpler rule-based systems to detect.
  • Enhanced Investigator Efficiency: Human investigators receive higher-quality leads, with supporting evidence and explanations, allowing them to focus on complex cases and make informed decisions faster. This translates to a 2x-3x increase in the efficiency of fraud investigation teams.
  • Improved Recoveries: By identifying fraud earlier and more accurately, there’s a better chance of recovering improperly paid funds.
  • Deterrent Effect: Once the word gets out, our sophisticated detection system acts as a deterrent, discouraging potential fraudsters.

It’s important to note that implementing our system is complex and requires continuous refinement and expert oversight (human-in-the-loop) to ensure accuracy, prevent bias, and adapt to evolving fraud tactics.