Diagnosis-Based Risk Score

Diagnosis-Based Risk Score in Health Plan Payment and Resource Allocation

A diagnosis-based risk score turns the conditions coded on a person's insurance claims or hospital records into an expected level of healthcare spending, usually expressed relative to the average person in a reference population. Payers use such scores to set capitated payments and to transfer money between competing insurers, so that a plan enrolling sicker people receives more. Every scoring system rests on choices about which codes count and how they are grouped, and those choices decide both how accurate the score is and how strongly it rewards coding. The score is one kind of risk score, and the wider practice of paying or comparing with it is risk adjustment.

Why payers added recorded diagnoses to demographics

Ellis cites the view of many researchers that diagnoses come as close as any widely available measure to consumer health status, so claims diagnoses are the most used information beyond demographics. Prescribing may vary more across doctors and over time than diagnostic coding, and surveys of self-reported health cost more to collect.

Ellis observes that every country with more elaborate risk adjustment began with age and gender. The Netherlands added Diagnostic Cost Groups built on 9 categories of inpatient diagnoses in 2004. US Medicare first used diagnoses in 2000 through the PIP-DCG model, which Pope and colleagues describe as using the worst principal inpatient diagnosis. Because only illnesses producing admissions counted, plans that kept people out of hospital through good ambulatory care could look healthier and be paid less, and the all-encounter model that followed is described on the hierarchical condition category page.

How diagnosis codes become one number

A score is built in three steps. Codes are grouped into clinically coherent categories, rules decide which categories count, and the categories are weighted by their association with cost. Ash and colleagues describe the DCG/HCC version, which maps ICD-9-CM codes to 543 diagnostic groups and then clusters these into 118 condition categories of related problems with similar expected costs. A hierarchy means a person placed in a severe category is not also placed in a lower category of the same disease, while unrelated conditions, such as cancer and diabetes, both count.

The developers also chose what not to count. Some categories were excluded, the number of hospitalisations was left out so that organisations treating sick people with fewer admissions were not penalised, and repeat appearances of a diagnosis were ignored. Ellis contrasts rate cell systems, which put each person in one cell, the most detailed using over 700 cells, with additive regression models, which cope better with multimorbidity because 20 common conditions alone give 1,048,576 combinations.

$$\hat{C}_i = \alpha_{d(i)} + \sum_{k \in K_i} \beta_k , \qquad R_i = \frac{\hat{C}_i}{\bar{C}}$$

where $\hat{C}_i$ is the predicted annual cost for person $i$, $\alpha_{d(i)}$ is the coefficient for the demographic cell $d(i)$ that the person falls in, $K_i$ is the set of condition categories retained for the person after hierarchies and exclusions, $\beta_k$ is the cost increment for category $k$, $\bar{C}$ is the mean annual cost in the reference population, and $R_i$ is the relative risk score, equal to 1.0 for a person of average expected cost.

Design choices that move the score

The period from which diagnoses are drawn is a major choice. Ellis explains that prospective models use a base year to predict the next year's spending, while concurrent models use same-year diagnoses and are much more predictive, a contrast developed on the prospective risk adjustment and concurrent risk adjustment pages. The outcome is most often total annual spending.

Ellis reports that ordinary least squares on untransformed spending is the most common approach and the one Medicare uses, with part-year enrolees' costs annualised and weighted by the fraction of the year enrolled, a form of weighted least squares. Calibration data matter too. In a Society of Actuaries comparison by Hileman and Steele, a model whose supplied weights came from Medicaid improved substantially when recalibrated to commercial data, while most others barely changed.

How well diagnosis-based scores predict

In the Ash and colleagues validation, prospective DCG/HCC models had validated R-squared values of 9.1% in a privately insured population, 23.1% in Medicaid and 8.5% in Medicare. In a US Medicare sample reported by Ellis, age and gender alone explained 1.1% of the variation in next-year charges, prior-year charges 8.9%, diagnoses grouped by DCG/HCC 10.4% and all available information together 16.9%.

Hileman and Steele compared over 40 risk scoring models on a sample of one million people from 2012 and 2013 MarketScan commercial claims. Uncensored R-squared for concurrent diagnosis-only models lay between 24.2% and 52.7%, against 12.9% to 30.1% for pharmacy-only models, and adding pharmacy data to diagnoses gave only a modest gain. For prospective diagnosis-only models the values lay between 9.1% and 20.7%. The authors stress that R-squared alone is not sufficient, because a few extreme values move it sharply, and censoring costs at USD 250,000 per person raised every value.

Group measures are often more relevant to payment. A predictive ratio compares predicted with actual spending for a subgroup.

$$PR_g = \frac{\sum_{i \in g} \hat{C}_i}{\sum_{i \in g} C_i}$$

where $PR_g$ is the predictive ratio for group $g$, $\hat{C}_i$ is predicted cost and $C_i$ is actual cost for person $i$ in the group. A ratio of 1.0 means the group is paid its cost on average, and Hileman and Steele cite a paper treating a ratio within 10% of 1.0 as reasonably accurate. Regression towards the mean is strong. Among prospective diagnosis-only models, people in the lowest two cost deciles were overpredicted by a factor of between 8.9 and 10.0, and the top 2% of spenders were predicted at only 23.7% to 31.0% of their actual cost.

Coding, persistence and the incentive to record

Recording is imperfect and responds to payment. Ellis cites studies finding that, even for serious chronic conditions such as diabetes and multiple sclerosis, fewer than 75% of people with a diagnosis coded in one year have it coded again the next. He also notes that once payment depends on codes, coded severity tends to rise, as it did after hospital diagnosis-related groups were introduced, a pattern described on the DRG creep page.

Designers respond with structure. Ash and colleagues anticipated DCG creep and used exclusions and hierarchies to reduce sensitivity to variation in coding practice, to deliberate coding for higher payment and to inconsistent coding of vague conditions. When NHS England tested a count of repeated diagnoses for its allocation model, described next, it capped the count at three to avoid building access effects into the model.

Recorded diagnoses in NHS England allocations

England uses diagnoses to share a national budget between areas. NHS England's technical guide for 2026/27 allocations explains that need per head for general and acute services has been estimated since 2014/15 with a person-based approach first developed by the Nuffield Trust, refreshed for 2022/23 and unchanged for 2026/27. Two years of inpatient diagnoses from the Secondary Uses Service dataset are turned into morbidity flags at ICD sub-chapter level, with extra flags for pairs of diagnoses whose joint effect differs from the sum. The guide states that morbidity (previous diagnoses) and age were the most important variables.

Variables for the availability of services are estimated but then set to the national average, so that low use caused by low capacity is not read as low need. The diagnoses still come from hospital records, which reflect access as well as illness, so the weights estimate relative cost-weighted need for a capitation style allocation to integrated care boards, not illness itself.

Worked example with three enrolees and one coding change (illustrative)

All figures here are illustrative. An additive prospective model is fitted on a reference population whose mean annual cost is GBP 1,000. Demographic coefficients are GBP 900 for a woman aged 75 to 84 and GBP 450 for a man aged 40 to 64. Condition increments are GBP 600 for diabetes without complications, GBP 1,100 for diabetes with complications, GBP 1,500 for heart failure and GBP 800 for chronic obstructive pulmonary disease, and complicated diabetes excludes uncomplicated diabetes in the hierarchy.

1. Enrolee A. A woman of 80 has base-year codes for heart failure and for diabetes both with and without complications. The hierarchy keeps only complicated diabetes, so her predicted cost is 900 + 1,500 + 1,100 = GBP 3,500 and her score is 3,500 / 1,000 = 3.50. Without the hierarchy the extra 600 would give 4,100, a score of 4.10.

2. Enrolee B. A man of 50 with no qualifying diagnoses is predicted at GBP 450, a score of 450 / 1,000 = 0.45.

3. Enrolee C. A man of 55 with chronic obstructive pulmonary disease is predicted at 450 + 800 = GBP 1,250, a score of 1.25.

4. A plan of 10,000 enrolees. Suppose the plan's average score is 1.12, so its expected cost and payment are 1.12 × 1,000 = GBP 1,120 per head. Its actual cost is GBP 1,180 per head, so the predictive ratio is 1,120 / 1,180 = 0.949, an underpayment of GBP 60 per head, or GBP 600,000.

5. A coding change. In the next base year the plan records uncomplicated diabetes for 500 more enrolees whose health has not changed. Predicted cost per head rises by 500 × 600 / 10,000 = GBP 30, to GBP 1,150, a score of 1.15, and total payment rises by 500 × 600 = GBP 300,000. The predictive ratio becomes 1,150 / 1,180 = 0.975.

6. Interpretation. The hierarchy stopped Enrolee A being paid twice for one disease, while her unrelated heart failure still counted. The plan's fit appears better after the coding change, yet nobody is sicker, and in a budget-neutral scheme the extra GBP 300,000 is drawn from other plans.

Use in economic evaluation and health services research

Outside payment, diagnosis-based scores are common covariates for case-mix adjustment in claims data studies, where matching on them balances predicted cost but leaves confounding by unrecorded illness in place. Predicting cost is also different from predicting need. Obermeyer and colleagues found that a commercial algorithm predicting cost left Black patients considerably sicker than White patients at the same score, because less is spent on patients with poorer access to care, a point that matters for health equity.

Boundaries with neighbouring scores

A diagnosis-based risk score is the general idea, and the hierarchical condition category is one grouping of diagnoses within it, used by CMS. Pharmacy-based scores draw on prescriptions, and demographic adjustment uses basic characteristics such as age and sex. A comorbidity index such as the one Charlson and colleagues developed also weights recorded conditions, but it was built to classify comorbid conditions that might alter the risk of death, not to predict cost. A diagnosis-related group classifies one hospital stay to price that episode, whereas a risk score summarises a person's conditions over a period. Selection by plans on features the score misses is treated under adverse selection and cream skimming.

Misreadings and limits

Most misreadings come from treating a payment prediction as a clinical measure. Four of them recur often.

1. Reading a higher score as a sicker person. The score rises with documentation and contact with services as well as with illness.
2. Expecting accurate individual prediction. Prospective R-squared values near 10% to 20% support fair group averages but leave most of one person's spending unexplained.
3. Comparing scores across systems or versions. Each score is relative to its own reference population and calibration, so 1.2 in one model is not 1.2 in another.
4. Treating the score as need. A model calibrated on spending reproduces existing patterns of access, and it misses conditions that were never diagnosed.

Sources

• Ash AS, Ellis RP, Pope GC, Ayanian JZ, Bates DW, Burstin H, Iezzoni LI, MacKay E, Yu W. Using diagnoses to describe populations and predict costs. Health Care Financing Review. 2000;21(3):7-28. https://pmc.ncbi.nlm.nih.gov/articles/PMC4194673/
• Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. Journal of Chronic Diseases. 1987;40(5):373-383. https://doi.org/10.1016/0021-9681(87)90171-8
• 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. p. 177-222. https://doi.org/10.1002/9783527611294.ch8 (author manuscript, February 2007: https://people.bu.edu/ellisrp/EllisPapers/2007_Ellis_Riskadjustment25.pdf)
• Hileman G, Steele S. Accuracy of Claims-Based Risk Scoring Models. Society of Actuaries; October 2016. https://www.soa.org/globalassets/assets/files/research/research-2016-accuracy-claims-based-risk-scoring-models.pdf
• NHS England. Technical guide to allocation formulae and convergence for 2026/27 allocations. Publication reference PRN02116, December 2025. https://www.england.nhs.uk/wp-content/uploads/2025/12/allocations-2627-technical-guide-to-formulae-v2.pdf
• 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, Ash AS, Ayanian JZ, Iezzoni LI, Ingber MJ, Levy JM, Robst J. 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/