9.2 Risk Adjustment & Healthcare Severity Modeling
Key Takeaways
- Risk adjustment supports fairer comparison by accounting for measured differences in patient characteristics, but it cannot remove unmeasured confounding or repair poor source data.
- For Medicare Advantage payment year 2026, CMS uses the 2024 CMS-HCC risk-adjustment model (often called V28) at 100%; mappings, hierarchies, coefficients, normalization, and model segments are version-specific.
- A CMS-HCC score combines demographic terms, retained hierarchical condition categories, applicable interactions, and current CMS adjustments; analysts must use the official files for the payment year rather than memorize legacy HCC numbers or coefficients.
- Hospital profiling may use hierarchical models in which a risk-standardized rate is calculated as predicted divided by expected outcomes, multiplied by the national observed rate; this is different from a simple observed-to-expected ratio.
- Discrimination, calibration, validity, and fitness for the intended population must all be evaluated; no single metric establishes model quality.
Risk Adjustment & Healthcare Severity Modeling
Quick Answer: Risk adjustment estimates expected resource use or outcomes after accounting for measured patient differences. It makes comparisons more defensible, but it does not make unlike populations identical. For Medicare Advantage payment year 2026, CMS uses the 2024 CMS-HCC model (V28) at 100%. Exact diagnosis mappings, HCC numbers, coefficients, interactions, normalization factors, and model segments are year-specific and must be taken from the applicable CMS files.
1. Why Risk Adjustment Is Needed
Raw outcomes mix at least three influences: patient risk, care processes, and random variation. A tertiary center treating transferred patients with advanced disease may have a higher crude mortality rate than a low-acuity community facility even if its care is excellent. Conversely, a low event rate can reflect favorable selection rather than better care.
A risk model attempts to estimate what would be expected given measured characteristics such as age, sex, diagnoses, prior utilization, functional status, or other permitted factors. The intended use determines which predictors are appropriate. Payment models, public outcome measures, and internal clinical prediction models have different estimands and exclusions.
Risk adjustment does not justify controlling away a disparity that the analysis is meant to expose. For an equity question, decide whether race, language, disability, or social risk is a confounder, a stratification variable, a mediator, or part of the outcome definition before modeling. Always report important unadjusted and stratified results alongside adjusted findings when they answer different questions.
2. The Current CMS-HCC Conceptual Workflow
CMS-HCC is a prospective risk-adjustment approach used in Medicare Advantage. In a simplified workflow:
- Confirm the payment year, model segment, and official CMS model files.
- Validate eligible diagnoses from permitted data sources and dates of service.
- Map ICD-10-CM diagnoses to condition categories using the applicable mapping.
- Apply disease hierarchies so a more severe related condition can supersede a less severe category where the model specifies.
- Add demographic factors and any applicable disease or demographic interactions.
- Apply the current model’s coefficients, normalization, and other CMS adjustments.
- Validate reconciliation totals and investigate material changes from the prior model or year.
For payment year 2026, CMS completed the phase-in of the 2024 CMS-HCC model and uses it at 100%. Older V24 examples with familiar HCC numbers are not reliable proxies: categories were restructured and renumbered, mappings changed, and coefficients depend on model segment and year.
The generic structure is:
Risk score = demographic contribution + retained condition-category contributions + applicable interaction contributions
That expression is conceptual, not a production calculator. A hypothetical patient with a demographic contribution of 0.40, retained disease contributions of 0.31 and 0.18, and a permitted interaction of 0.07 would have an illustrative pre-adjustment sum of 0.96. Those numbers are invented for arithmetic practice; they are not CMS coefficients.
Hierarchies and interactions
A hierarchy prevents double-counting related levels of disease severity. If the current mapping places two conditions in one hierarchy, only the category specified by the hierarchy survives. Unrelated categories may remain additive. Interactions are separate model terms that capture combinations whose expected cost differs from the simple sum of their main effects.
Never infer the hierarchy from clinical intuition alone. Use the official software and mapping files. Also distinguish a condition documented in the chart from a diagnosis eligible for the particular risk-adjustment submission. Documentation, coding, data-source, and timing rules all matter.
3. Other Severity and Risk Systems
| System | Typical use | Unit of analysis | Important distinction |
|---|---|---|---|
| CMS-HCC | Medicare Advantage payment risk | Person/year or payment period | Prospective, payment-specific, and versioned |
| MS-DRG | Medicare inpatient prospective payment | Inpatient stay | Groups a stay by diagnoses, procedures, and severity logic |
| APR-DRG | All-payer severity and mortality comparisons | Inpatient stay | Adds severity-of-illness and risk-of-mortality subclasses |
| Charlson-based methods | Comorbidity adjustment in research | Patient/encounter | Implementations and weights vary; document the version |
| Elixhauser-based methods | Comorbidity adjustment in outcomes research | Patient/encounter | Condition definitions and scoring approaches vary |
| CDPS | Medicaid risk adjustment | Person | Developed around disability and Medicaid populations |
| ACG or similar grouping systems | Population management and utilization prediction | Person/population | Proprietary specifications may apply |
Do not substitute one system merely because both are called “risk adjustment.” A payment model may be unsuitable for estimating causal treatment effects. A stay-level severity grouper does not automatically provide a prospective annual cost score.
4. Observed-to-Expected Ratios
A simple observed-to-expected ratio is:
O/E = observed events / expected events
If 36 readmissions are observed and the validated model predicts 30, O/E = 36 / 30 = 1.20. The organization experienced 20% more events than expected under that model. If 24 are observed, O/E = 0.80, or 20% fewer than expected.
Interpretation requires uncertainty. An O/E above 1 is not automatically statistically distinguishable from 1, and “expected” is conditional on the model, data, and reference population. Report confidence intervals or another suitable uncertainty estimate, examine calibration, and check whether low volume makes the ratio unstable.
CMS hierarchical risk-standardized outcomes
Do not conflate simple O/E with every CMS measure. For several hospital outcome measures, CMS uses a hierarchical generalized linear model and reports a risk-standardized rate in this form:
Risk-standardized rate = (predicted outcomes / expected outcomes) × national observed rate
“Predicted” incorporates the hospital-specific effect for its patients; “expected” estimates outcomes for those same patients at an average hospital. Empirical Bayes shrinkage pulls unstable estimates from low-volume hospitals toward the overall mean. The current measure methodology must be consulted for exact cohorts, risk variables, time windows, and exclusions.
5. Validation and Governance
A model can discriminate well and still be poorly calibrated. Evaluate several dimensions:
- Discrimination: Can the model rank higher-risk patients above lower-risk patients? For binary outcomes, ROC AUC is common.
- Calibration: Do predicted probabilities agree with observed frequencies overall and across clinically meaningful risk ranges? Use calibration plots, calibration intercept/slope, and appropriate summaries.
- Internal validity: Use resampling or cross-validation to estimate optimism.
- External validity: Test performance in the target sites, populations, and time period.
- Data validity: Confirm coding completeness, lookback windows, enrollment, duplicate handling, and mapping versions.
- Fairness and transportability: Examine performance and error rates across relevant subgroups without assuming equal aggregate AUC means equal impact.
- Change control: Version model files, mappings, code, input snapshots, approval records, and reconciliation outputs.
A non-significant goodness-of-fit test does not prove calibration, and an AUC threshold does not prove clinical utility. Model approval should tie performance to the intended decision and the harm of false positives and false negatives.
6. Practical Audit Checklist
Before releasing risk-adjusted results, ask:
- What decision and estimand does the adjustment support?
- Is the model current for the payment or measurement year?
- Are the cohort, source, and lookback rules implemented exactly?
- Were mappings, hierarchies, interactions, and model segments version-controlled?
- Are observed, expected, and predicted quantities labeled correctly?
- Were discrimination, calibration, uncertainty, and subgroup performance assessed?
- Can another analyst reproduce the result from retained code and approved inputs?
The exam-relevant principle is durable: choose an approach that fits the question, explain it, validate it, and interpret the adjusted result without overstating what adjustment can accomplish.
For Medicare Advantage payment year 2026, which approach should an analyst use to calculate CMS-HCC risk scores?
A hospital quality analytics department is evaluating its inpatient surgical outcomes. The multivariable risk-adjustment model calculates an Expected Mortality count of E = 50.0 deaths for the hospital's surgical case mix. The hospital records an actual Observed Mortality count of O = 35.0 deaths. What is the hospital's Observed-to-Expected (O/E) Mortality Ratio, and what does it indicate?
Which of the following correctly describes the structural difference between the 3M APR-DRG inpatient classification system and the CMS-HCC risk adjustment model?