Risk Score

Risk Score in Health Plan Payment and Clinical Decisions

In health systems, a risk score is the figure a payer or clinician reads off a prediction model to decide how much money or which treatment should follow a particular person. In health insurance the figure is usually a relative cost index, so a score of 1.20 means the model expects spending 20% above the average of its reference population, and a capitated payment is the score multiplied by a base rate. In clinical care and health technology assessment the figure is more often an absolute probability, such as a 10-year chance of cardiovascular disease, compared with a treatment threshold. The scale of the number, the reference population and the outcome predicted decide what a score can be used for. The method behind payment scores is on the risk adjustment page and the coding route on the diagnosis-based risk score page.

Three scales a risk score can use

The same word covers numbers with different units, and most misuse starts with confusing them. Three scales are common in health economics.

A relative cost index expresses predicted spending for one person as a multiple of average predicted spending. In its Rate Announcement for calendar year 2027, CMS states that the Medicare Advantage risk score is the sum of relative factors assigned to a beneficiary and determines the expected relative cost. A score of 2.0 on this scale is not a 200% chance of anything.

An absolute probability gives the chance of a defined event within a stated horizon. Steyerberg and colleagues define calibration as the agreement between observed outcomes and predictions, so a predicted 20% risk should correspond to about 20 events in 100 similar people. A probability can be checked against observed event rates in this way, and a relative index is checked against observed spending through predictive ratios.

A points total rounds a fitted model into integer points for bedside use, so a total of 3 means nothing until a published table maps it to a probability.

How a payment risk score is assembled

A payment score starts as a regression of annual spending on demographic cells, condition categories and interactions. Describing its Part D model, CMS explains that the dollar coefficients are divided by the average per capita expenditure the model predicts for a given year, giving relative factors and an average score of 1.0 in that year. Pope and colleagues, describing the CMS hierarchical condition category model introduced in 2004, note that categories for unrelated diseases accumulate and that some combinations, such as diabetes with congestive heart failure, carry interaction terms.

$$S_i = \frac{1}{F}\left(\sum_{k=1}^{K} f_k X_{ik}\right)(1-c), \qquad P_i = B \times S_i$$

where $S_i$ is the payment risk score for enrolee $i$, $f_k$ is the relative factor for risk marker $k$, $X_{ik}$ equals 1 if enrolee $i$ has marker $k$ and 0 otherwise, $K$ is the number of markers, $F$ is the normalisation factor, $c$ is the coding adjustment as a proportion, $P_i$ is the payment for enrolee $i$ and $B$ is the base rate for an average enrolee. The form simplifies the Medicare Advantage structure.

Two programme-level corrections act on every score. CMS states that its normalisation factors offset the trend in risk scores and keep the average fee-for-service risk score at 1.0. For 2027 it set the factor for the 2024 CMS-HCC model at 1.079, after fee-for-service risk scores grew by 1.8% a year on average between 2022 and 2025. MedPAC's payment basics note, revised October 2026, states that because Medicare Advantage plans tend to document diagnoses more thoroughly than providers in traditional Medicare, the Secretary must reduce their risk scores by at least 5.9%, and CMS finalised exactly 5.9% for 2027. Neither correction changes how enrolees rank against each other, but both change how much money a given score buys.

Reading scores at plan and population level

A payer pays for populations, so a plan's average score matters more than any single enrolee's. Pope and colleagues measured accuracy for individuals with R-squared and for groups with predictive ratios, the ratio of mean predicted to mean actual spending for a subgroup. They read predictive ratios below 1.0 as underprediction.

Ellis states that an ideal model would have predictive ratios close to one for every group of interest, and that validated R-squared in a sample not used for estimation is often 5 to 50% lower than in the development sample. Where a subgroup's predictive ratio sits well below one, an insurer gains by avoiding that group, the incentive behind cream skimming, and the sorting of people across plans can feed adverse selection.

Clinical risk scores and treatment thresholds

Clinical scores usually come from a logistic or survival model that converts a weighted sum of predictors into a probability. For a binary outcome over a fixed horizon the logistic form is standard.

$$p_i = \frac{1}{1 + \exp\left(-\beta_0-\sum_{j=1}^{J} \beta_j x_{ij}\right)}$$

where $p_i$ is the predicted probability of the event for person $i$ within the horizon, $\beta_0$ is the intercept, $\beta_j$ is the coefficient on predictor $j$, $x_{ij}$ is the value of predictor $j$ for person $i$ at the time of prediction, and $J$ is the number of predictors. A survival model also needs a baseline survival function before its linear predictor becomes an absolute risk at a given time.

The NICE guideline on cardiovascular disease risk assessment and reduction (NG238, published 14 December 2023) shows a score becoming a resource allocation rule. It recommends the QRISK3 tool to estimate 10-year cardiovascular risk for people aged 25 to 84 without the disease, and offering atorvastatin 20 mg for primary prevention when the 10-year QRISK3 score is 10% or more. The committee's rationale records that statins were cost effective for people with 10-year risk scores below 10%, yet the 10% threshold was retained. National audit data showed only about half of people scoring 10% or more were prescribed lipid-lowering treatment, and the committee judged that raising uptake among those with most potential to benefit would have more impact than lowering the threshold. A treatment threshold on a risk score is therefore a policy choice that cost-effectiveness evidence informs but does not settle.

What validation has to show

A score that ranks people well can still give the wrong numbers, and the two failures have different costs. Steyerberg and colleagues describe the concordance statistic, identical to the area under the ROC curve for a binary outcome, as the usual measure of discrimination. They summarise calibration by an intercept, showing whether predictions are systematically too high or too low, and a slope, which should equal 1. At validation a slope below 1 reflects overfitting. They conclude that discrimination and calibration should always be reported, with decision-analytic measures such as the net benefit in decision curve analysis when the model guides clinical decisions.

For reporting, the TRIPOD+AI statement of Collins and colleagues is a 27 item checklist that supersedes the 2015 TRIPOD checklist and gives particular emphasis to fairness. For appraisal, PROBAST+AI from Moons and colleagues replaces the 2019 PROBAST tool.

Choice of outcome and equity

The outcome a score is trained to predict decides whose needs it sees. Obermeyer and colleagues studied a commercial algorithm that health systems used to pick patients with complex needs for extra help. At a given risk score, Black patients were considerably sicker than White patients, because the algorithm predicted healthcare costs and unequal access meant less was spent on Black patients. Correcting the disparity would have raised the share of Black patients receiving extra help to 46.5%, against 17.7% under the original algorithm. A score built to predict spending reproduces past patterns of access, so health equity requires checking whether the label stands for need.

Worked example across payment and clinical scales (illustrative)

All relative factors, rates, costs and effects in this example are illustrative, including the normalisation factor of 1.05. Only the 5.9% coding adjustment comes from the 2027 Medicare Advantage rules.

1. Raw payment score. An enrolee has a demographic factor of 0.45, factors of 0.30 and 0.33 for two chronic conditions and an interaction factor of 0.11 for having both. The raw score is 0.45 + 0.30 + 0.33 + 0.11 = 1.19.

2. Normalisation and coding adjustment. Dividing by the normalisation factor gives 1.19 / 1.05 = 1.133. The coding adjustment multiplies the result by 1 − 0.059 = 0.941, giving 1.133 × 0.941 = 1.066.

3. Payment. At a base rate of USD 1,000 per enrolee per month, the plan receives 1,000 × 1.066 = about USD 1,066 a month, 6.6% more than the base rate for an average enrolee.

4. Group accuracy. Among enrolees with the second condition, mean predicted annual spending is USD 13,500 and mean actual spending USD 15,000, a predictive ratio of 13,500 / 15,000 = 0.90. A plan enrolling 500 such people is underpaid in expectation by 500 × (15,000 − 13,500) = USD 750,000 a year.

5. Clinical score and absolute benefit. Two people have 10-year cardiovascular risk scores of 12% and 6%, and a preventive drug with an assumed constant relative risk reduction of 25% has absolute risk reductions of 0.12 × 0.25 = 0.030 and 0.06 × 0.25 = 0.015, so the number needed to treat is 1 / 0.030 = 33.3 and 1 / 0.015 = 66.7.

6. Cost per event avoided. At an undiscounted drug cost of GBP 300 per person over 10 years, the cost per event avoided is 300 / 0.030 = GBP 10,000 at 12% risk and 300 / 0.015 = GBP 20,000 at 6% risk.

7. Miscalibration. If people scored at 12% in fact have a 9% risk, the absolute risk reduction is 0.09 × 0.25 = 0.0225 and the cost per event avoided rises to 300 / 0.0225 = GBP 13,333. The ranking is unchanged, but the cost per event avoided in the higher group is a third above what the score implied.

Risk scores inside economic evaluation

Economic models use risk scores in three distinct places. A score can supply the baseline risk of events for a modelled cohort, so a trial-based treatment effect is applied to the risk of the population under consideration, and absolute risk reduction and cost effectiveness then vary across risk groups, as step 6 shows. A score can define strata for a prespecified subgroup analysis. Observational cost studies use payment scores from claims data as covariates, where a score built from diagnoses recorded after treatment starts can absorb part of the treatment's effect and weaken confounding control. A constant relative effect across risk levels has to be assumed or tested, and a score validated in one population may need recalibration before it sets baseline risk in another.

Boundaries with neighbouring concepts

Risk adjustment is the method of setting payments or comparisons from expected risk, and the risk score is its output for one person, which capitation payment multiplies by a base rate. Predicted cost is the money amount a model forecasts, while a relative score divides that amount by an average. A propensity score predicts the probability of receiving a treatment, not of an outcome.

Common misreadings of risk scores

Most errors come from reading the number on the wrong scale or for the wrong group. Each has a direct cost for a payer or a guideline committee.

1. Treating a relative cost index as a probability. A score of 1.5 means 50% above average predicted spending, not a 150% or 1.5% risk.
2. Comparing scores across models or years. Recalibration, normalisation and coding adjustments change the scale, so rising average scores do not prove a sicker population.
3. Assuming a threshold is a fact of the model. NG238 kept 10% for practical reasons although statins were cost effective below it.
4. Assuming the training outcome measures need. A score built on spending inherits differences in access.

Sources

• Centers for Medicare & Medicaid Services. Announcement of Calendar Year (CY) 2027 Medicare Advantage (MA) Capitation Rates and Part C and Part D Payment Policies. 6 April 2026. https://www.cms.gov/files/document/2027-announcement.pdf
• Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. https://doi.org/10.1136/bmj-2023-078378
• Ellis RP. Risk adjustment in health care markets: concepts and applications. In: Lu M, Jonsson E, eds. Financing Health Care: New Ideas for a Changing Society. Wiley-VCH; 2008:177-222. Author manuscript dated February 2007. https://people.bu.edu/ellisrp/EllisPapers/2007_Ellis_Riskadjustment25.pdf
• Medicare Payment Advisory Commission. Medicare Advantage program payment system. Payment Basics, revised October 2026. https://www.medpac.gov/wp-content/uploads/2026/10/MedPAC_Payment_Basics_26_MA_FINAL_SEC.pdf
• Moons KGM, Damen JAA, Kaul T, et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ. 2025;388:e082505. https://doi.org/10.1136/bmj-2024-082505
• National Institute for Health and Care Excellence. Cardiovascular disease: risk assessment and reduction, including lipid modification. NICE guideline NG238. Published 14 December 2023. https://www.nice.org.uk/guidance/ng238/chapter/Recommendations
• Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-453. https://doi.org/10.1126/science.aax2342
• Pope GC, Kautter J, Ellis RP, et al. Risk adjustment of Medicare capitation payments using the CMS-HCC model. Health Care Financing Review. 2004;25(4):119-141. https://pmc.ncbi.nlm.nih.gov/articles/PMC4194896/
• Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models: a framework for some traditional and novel measures. Epidemiology. 2010;21(1):128-138. https://doi.org/10.1097/EDE.0b013e3181c30fb2