Risk Adjustment (HCC) Capture Agent

The Problem

Medicare Advantage (MA) plans receive capitated payments adjusted for member risk; the adjustment is based on documented chronic conditions coded as Hierarchical Condition Categories (HCCs). Undercoding,failure to capture all documented conditions,directly reduces capitation and plan profitability. Studies indicate 10 to 20% of eligible HCC conditions go undocumented in claims/ICD-10 coding, representing millions in lost revenue for large MA plans. Conversely, overcoding creates compliance risk (OIG RADV audits). The challenge is capturing legitimate conditions from unstructured provider notes where documentation may be implicit (patient on insulin → diabetes) but not explicitly coded.

Data Requirements

Data Sources:

• Unstructured clinical notes from provider EHRs (via FHIR/HL7 feeds) , office visit summaries, care plans, discharge summaries, progress notes
• Claims data (837/835) with diagnosis codes , current coded conditions for comparison
• CMS HCC mapping table (annual update) , diagnosis code to HCC mapping rules
• Payer member master file , current risk score, enrolled MA plan
• Provider directory , specialty, NPI, EHR system (to infer note format)

Data Classification:

• Public: CMS HCC mapping tables (published annually)
• Internal Structured: Claims data, member demographics, risk score history, provider directory
• Internal Unstructured: Clinical notes (unstructured free text), care summaries, prior provider queries and responses
• External Third-Party: EHR feeds (FHIR/HL7 from providers)
• Sensitive/Regulated: All clinical notes are highly sensitive PHI; risk adjustment coding is subject to OIG RADV audits and requires full audit trail

Data Quality Requirements:

• Timeliness: Clinical notes must be available within 5 to 10 business days of documentation for timely capture; annual RADV submission deadline is typically September 30.
• Completeness: At least 80% of provider encounters should have available notes; gaps trigger provider request for historical records.
• Accuracy: NLP must not introduce false positives (claiming conditions not present); conservative approach (only high-confidence conditions) is preferred; HCC mapping must use current CMS version.

Integration Complexity: Medium-High , Requires reliable FHIR/HL7 feed from provider systems (Epic, Cerner, Meditech), NLP model training on healthcare data (typically proprietary or third-party model), and CMS data integration. Needs careful testing to ensure no false positives (compliance risk). Typical implementation uses third-party NLP service (e.g., Nuance DAX, Optum360) or custom ML model.

Regulatory Alignment

• CMS Risk Adjustment Data Validation (RADV): Annual audit of HCC coding accuracy; agent's audit trail supports compliance defense; conservatism (avoiding false positives) is critical.
• CMS Interoperability Rule: Requires access to clinical data; supports broader data sharing requirements.
• HIPAA: All notes are PHI; full access controls and audit logging required.
• OIG Compliance Programs: MA plans should have documented processes for HCC capture; agent demonstrates proactive approach.

Comparable Implementations

• Optum360 CAC (Computer Assisted Coding) , Industry standard for HCC capture from unstructured notes
• Nuance DAX (Document Analysis & Coding) , Ambient AI that identifies missing codes in real-time
• 3M CodeFinder , Traditional coding assistance, increasingly AI-enhanced