MST’s Risk Adjustment System

 

Building the Risk Adjustment Knowledge Graph

  • Comprehensive Data Ingestion: MST pulls in an even wider array of healthcare data than for fraud detection, focusing specifically on data relevant to patient acuity and risk. This includes:
    • Claims Data: ICD-10-CM diagnosis codes, CPT codes, HCPCS codes.
    • Electronic Health Records (EHRs): Structured data (diagnoses, medications, lab results, vitals) and crucial unstructured data (clinical notes, discharge summaries, physician progress notes).
    • Pharmacy Data: Medication lists and dispensing records.
    • Lab Results: Test results and their interpretations.
    • Social Determinants of Health (SDOH): Data on socioeconomic factors, living conditions, and other non-clinical factors that impact health.
    • Patient Demographics: Age, gender, location, etc.
    • Medical Guidelines and Ontologies: ICD-10-CM guidelines, Hierarchical Condition Categories (HCC) mapping rules, SNOMED CT, LOINC, RxNorm, etc. These are critical for accurate coding and risk score calculation.
  • Entity and Relationship Extraction (LLM-Powered): Our LLM plays a vital role here, especially for unstructured data:
    • Named Entity Recognition (NER): Identifies all relevant medical entities (diseases, symptoms, procedures, medications, anatomical sites, labs) from clinical notes.
    • Relation Extraction: The system understands the relationships between these entities (e.g., “patient suffers from diabetes,” “medication prescribed for hypertension,” “lab result indicates high cholesterol,” “history of heart attack”).
    • Contextual Understanding: Crucially, the LLM interprets the nuance of clinical language. It can differentiate between a “rule-out” diagnosis, a past history, a current active condition, or a family history, which is vital for accurate HCC assignment. It can also identify conditions documented across multiple visits or by multiple providers that, when combined, point to a chronic condition.
  • Vectorization and Graph Construction: Each extracted entity and relationship is vectorized, creating embeddings that capture their semantic meaning and relationships within the healthcare domain. These vectors form a complex, interconnected knowledge graph where:
    • Nodes represent entities (patients, providers, diagnoses, medications, encounters, lab tests, specific clinical findings).
    • Edges represent relationships (e.g., “patient P has diagnosis D,” “provider X documented diagnosis D for patient P on date Z,” “medication M is prescribed for diagnosis D,” “lab test L is indicative of condition C”).
    • Crucially, explicit links are then established between clinical documentation and the relevant HCCs, ensuring traceability and compliance.

Leveraging the Knowledge Graph and LLMs for Risk Adjustment:

  • Comprehensive Condition Capture (Pre-payment/Prospective):
    • Automated Chart Review: MST’s system automatically reviews vast quantities of patient records (structured and unstructured) to identify all documented conditions that qualify for HCCs. The LLM’s ability to read and understand clinical notes is transformative here, as much of the critical information for risk adjustment resides in unstructured text.
    • Identifying Undocumented Conditions/Gaps in Care: By analyzing patterns in claims, pharmacy data, and lab results, the system flags suspected conditions that may not have been explicitly coded. For example, a patient on multiple diabetes medications with high HbA1c results, but no documented diabetes HCC, would be flagged. The LLM then searches the full clinical record for supporting evidence.
    • Hierarchical Condition Category (HCC) Mapping: The knowledge graph, pre-populated with HCC hierarchies and mapping rules, can automatically assign the correct HCCs based on the identified diagnoses and their supporting documentation. It can also identify potential “upcoding” or “downcoding” based on supporting clinical evidence.
  • Documentation Improvement and Query Generation:
    • Clinical Abstraction Support: When a potential HCC is identified but the documentation is ambiguous or incomplete, our system can generate specific, context-aware queries for providers or medical coders, pointing them to the exact sections of the patient record that need clarification. For example, “Is the patient’s CHF still active and being treated?” or “Please confirm the severity of the patient’s COPD.”
    • Point-of-Care Insights: By integrating with EHRs, the system can provide real-time suggestions to providers during patient encounters about potential HCCs that need to be documented or re-affirmed, ensuring accurate capture at the source.
  • Audit Readiness and Compliance (RADV Support):
    • Evidence Tracing: The knowledge graph inherently provides a transparent trail of how each HCC was derived, linking it directly to specific documentation within the patient record. This makes Risk Adjustment Data Validation (RADV) audits significantly easier and more efficient, as the system can quickly pull all supporting evidence.
    • Automated Validation: The system can perform automated checks against CMS guidelines and internal coding rules, identifying potential unsupported diagnoses before submission, minimizing the risk of audit penalties.
  • Population Health Management:
    • Longitudinal Patient View: The knowledge graph provides a comprehensive, longitudinal view of each patient’s health status, allowing for better identification of high-risk individuals and proactive care management.
    • Cohort Analysis: It can identify patient cohorts with similar risk profiles or care gaps, enabling targeted interventions.

Typical Results

The impact on risk adjustment can be substantial, leading to both financial and clinical benefits:

  • Increased HCC Capture Rate: A significant improvement in identifying and capturing all legitimate HCCs for a population. We often achieve a 10-25% increase in accurately captured HCCs, leading to more accurate risk scores and appropriate reimbursement.
  • Improved Risk Adjustment Factor (RAF) Accuracy: Due to more complete and accurate HCC capture, the overall RAF scores for a population will be more precise, reflecting the true burden of illness. This could translate to a 5-15% increase in overall RAF scores, ensuring health plans are correctly compensated.
  • Reduced Manual Effort for Chart Review: This means automation of large parts of the chart review process. We’ve seen a 40-60% reduction in the time and resources spent on manual chart abstraction and coding.
  • Faster and More Efficient RADV Audits: The ability to quickly and accurately pull supporting documentation for audits, leading to a 50-70% reduction in audit preparation time and a higher success rate in defending submitted risk scores.
  • Decreased Audit Penalties: By proactively identifying and correcting unsupported diagnoses, the system can significantly reduce potential CMS audit penalties by 20-50%.
  • Enhanced Compliance: Consistent application of coding guidelines and robust documentation, leading to higher confidence in regulatory compliance.
  • Better Patient Outcomes (Indirectly): By more accurately identifying chronic conditions and risk factors, health plans and providers can better target care management programs, leading to earlier interventions and improved patient health.
  • Improved Provider Engagement: By making it easier for providers to accurately document patient conditions at the point of care, it can reduce “provider abrasion” associated with risk adjustment queries and retrospective chart reviews.

The combination of a vectorized knowledge graph providing structured, interconnected knowledge and an LLM capable of understanding the nuances of clinical language is a powerful synergy for tackling the complexities of risk adjustment.