Sources and review notes
These links support definition-level research and do not establish the regulatory status, safety, or suitability of any product.
CMS's Measures Management System defines risk adjustment for quality measures as a mathematical method for accounting for population characteristics to support fairer comparison of outcomes across measured entities, while risk stratification compares performance within defined groups and can expose differences that adjustment might obscure. CMS advises measure developers to justify the variables and method, use valid and comprehensive data, separate development and validation samples, assess model adequacy, and document limitations; the appropriate approach is measure- and purpose-specific. Medicare Advantage uses separate CMS-HCC, RxHCC and ESRD payment models, with CMS publishing year-specific software and ICD-10 mappings, including initial and midyear or final 2026 artifacts. Those payment models should not be generalized to a quality measure, commercial payer, clinical risk score, individual treatment decision, or different payment year. CMS's Medicare Advantage RADV program confirms whether diagnoses submitted for risk-adjusted payment are supported in enrollee medical records and may recover overpayments for unsupported diagnoses. A diagnosis suggestion, code-to-HCC mapping, calculated score, model output or prior-year acceptance therefore does not establish current documentation support or submission eligibility. A scoped implementation should name the program, contract, payment or performance year, model family and version, model artifact and checksum, ICD version and mapping release, data-collection and submission periods, eligible provider and encounter types, diagnosis source, demographic inputs, coefficients, hierarchy and interaction rules, normalization, and responsible coding, actuarial, quality and compliance reviewers. Teams should preserve the original encounter and signed documentation, author and provider type, service and signature dates, diagnosis and code version, source-system identifiers, extraction and normalization steps, mapping result, suppressed or hierarchical categories, demographic factors, model run, candidate rationale, reviewer decision, correction, submission record, response and audit package. AI may surface candidate conditions, possible documentation gaps or inconsistent data, but it must not create a diagnosis, infer unsupported specificity, copy a historical condition forward, select a code solely for payment impact, or submit a result without the required qualified review and current evidence. Review queues should show the exact text and encounter supporting a candidate, conflicting and absent evidence, provenance, model and mapping version, confidence and reason, and the financial or quality context without hiding rejected candidates. Validation should use CMS or program reference artifacts and known-answer cases to test demographic segments, new and deleted codes, hierarchy suppression, disease and demographic interactions, multiple encounters, duplicates, corrected and voided data, late records, ineligible sources, missing signatures, conflicting diagnoses, model transitions, retroactive changes and reruns. Results should be reconciled independently at member, diagnosis, category, coefficient and aggregate levels, with deterministic replay and change reports between software and mapping releases. Monitoring should distinguish data completeness, coding and mapping concordance, supported and unsupported candidates, false positives and negatives from reviewed samples, reviewer overrides, score changes by reason, submissions accepted or rejected, retrospective corrections, audit findings, workload and results by relevant population and provider strata. Quality comparisons should present unadjusted and stratified results when needed to avoid concealing disparities, and model performance should be assessed for calibration and discrimination in the intended population rather than inferred from a single aggregate score. Contracts should cover program and model updates, source-data rights, transparency of mappings and transformations, human review, audit support, correction and deletion, security, subcontractors, retention, export, liability allocation and vendor exit. A higher risk score, more surfaced diagnoses, a successful model replication or an accepted submission does not by itself prove better documentation, coding accuracy, care quality, compliance, reimbursement, equity or patient need.
