RESEARCH DOSSIER
Morbidity Classification and Predictive Health Risk Scoring
Abstract. Morbidity classification and predictive health risk scoring transform diagnoses, comorbidity histories, utilization and financial observations into risk estimates for defined insured and public populations. This monograph develops a source-verified international synthesis of the underlying classification and inference systems, their historical progression and institutional uses, and the consequences of coding incentives, population selection, target mismatch and model drift. It distinguishes WHO ICD-11 and national ICD-10-CM coding, AHRQ CCSR and Elixhauser software, classical comorbidity indices, the U.S. CMS-HCC and HHS-HCC payment frameworks, and prospective clinical and claims-cost predictions. Evidence includes original epidemiological methods, official classification and regulatory records, and empirical research on upcoding, algorithmic equity and calibration. Equations and numerical examples are explicitly synthetic. The resulting methodological framework links classification provenance to statistical validation, insurance exposure, clinical meaning and contract-aware financial risk, identifying clear limits on transfer across health systems and insurance products.
Keywords: morbidity; clinical classification; health risk adjustment; ICD-11; ICD-10-CM; CCSR; Elixhauser; Charlson; CMS-HCC; HHS-HCC; predictive modeling; calibration; upcoding; case mix; medical costs; insurance exposure; health equity
Introduction
Morbidity classification and predictive health risk scoring organize observed conditions, healthcare use, demographic information and historical risk exposure into structured predictors of future disease, medical service use, disability or covered loss. Health insurers, public risk-equalization schemes, and population-health programs use related variables for different institutional decisions. A scientific assessment must establish the predicted variable, the population at risk, the time of information availability, and the decision for which a score is used. The same diagnosis history may support clinical stratification, payment adjustment or statistical forecasting, but the resulting outcomes, incentives and validity conditions are distinct.
A diagnostic classification locates a condition within an international or national taxonomy. A comorbidity grouping summarizes conditions coexisting with a primary illness or with each other. A risk score is a model output, typically an estimated relative expected outcome or a probability for a stated horizon. A payment adjustment factor is a statutory or contractual calculation that may use a risk score while imposing additional normalization, eligibility and transfer rules. These objects occupy different levels of an insurance decision system.
The WHO ICD-11 implementation guidance establishes the international role of disease classification and explains the 2022 worldwide effective date of the eleventh revision while allowing national transition schedules. The 2026 WHO ICD-11 fact sheet describes interoperability and digital health applications. Country-specific claims systems may continue to use national ICD-10 modifications and local revisions, and the existence of a global classification does not mean that an insurance claims code in one country can be mapped without loss to another.
This monograph develops an original comparative synthesis of classification architectures, risk adjustment mechanisms, predictive cost and morbidity estimands, coding and selection incentives, statistical validation, and international transfer. It separates empirically documented findings, mathematical accounting identities and illustrative hypothetical cases. Its framework uses existing medical classification and predictive science and institutes no new proprietary medical or actuarial doctrine.
Foundations and Source-Critical History
The development of prognostic comorbidity indices
Formal comorbidity adjustment arose from the need to compare patients whose baseline conditions influence prognosis and resource use. Charlson and colleagues' 1987 longitudinal prognostic index classifies clinically specified conditions and weights them to predict mortality. Its original outcome was survival risk among observed patients, with later validation in a separate cohort. Those weights express the association between specified conditions and the original mortality target in the derivation setting; a weighted mortality index cannot be interpreted automatically as an expected dollar expenditure or as a disease-prevalence census.
The original Elixhauser and colleagues (1998) methods created a broader administrative-data collection of comorbidity indicators to describe outcome differences in hospitalized populations. Using multiple indicators preserves information about particular conditions and allows interactions with age, service line, treatment and health system characteristics. Combining those indicators into one weighted score changes the target and the statistical assumptions. Van Walraven and colleagues (2009) demonstrated a mortality-specific point system condensed from the earlier Elixhauser group set; the validation discrimination they report is a hospital-death metric. It cannot validate a separate insurer's annual claim expenditure forecast.
Contemporary diagnosis grouping standards
The U.S. AHRQ Clinical Classifications Software Refined maps more than 70,000 ICD-10-CM diagnosis codes to over 530 categories and distinguishes use cases such as inpatient and outpatient administrative observations. AHRQ lists a v2027.1 release dated November 10, 2026. This illustrates how quickly an ICD-based analytic system can change: versions, code validity windows, input field conventions and first-listed diagnosis rules must be stored alongside an extracted dataset. Applying a new mapping file retroactively to historical claims changes analytic definitions even when clinical events remain the same.
AHRQ's Elixhauser Comorbidity Software Refined identifies 38 pre-existing conditions for current ICD-10-CM uses, employing secondary diagnosis information and present-on-admission indicators for selected conditions. The earlier original study's 30 groups and the modern software's 38 categories refer to different published specifications. Researchers must not splice the two outputs under an unchanged column name. A hospital discharge field, a community outpatient diagnosis and a payer's longitudinal diagnosis history have distinct selection mechanisms and often different coding incentives.
A hierarchy of evidence and applicability
For scientific classification, primary records include the international WHO classification, national adaptation documentation, versioned software mappings and user guides, linked clinical records, payer claims, and original published validation studies. Official risk-adjustment regulations define a different evidence object: which diagnoses and demographic attributes enter a particular payment formula for a specified year. Empirical research on prediction, fairness and endogenous recording identifies mechanisms that a technical mapping manual cannot settle on its own.
Evidence should therefore be recorded on three levels: the rules that define a coding or payment system, the observed data that populate it, and the outcome study that tests how well the system works. A national regulation can establish a legally defined HCC risk score and still be silent about predictive performance for a specific private insurer. A high retrospective prediction score can coexist with administrative coding incentives that distort population comparison. Proper scientific interpretation preserves these levels as separate auditable records.
Diagnosis and Utilization Predictors
Units, exposure and information timing
Let i denote a covered person, t an observation year, X_i,t a vector of diagnosis, service and demographic predictors measured by a declared cutoff, and Y_i,t+1 the outcome in the subsequent period. Prospective health insurance risk scoring seeks an estimate E[Y_i,t+1 | X_i,t] or a probability Pr(Y_i,t+1 exceeds a stated threshold | X_i,t), with an eligibility and coverage definition attached. The prediction target may be all medical expenditures, plan-paid allowed charges, insurer indemnity, hospital admission, disability onset, morbidity progression, mortality or another explicitly stated event. Their denominators and predictor validity are different.
An insurance person-year ordinarily measures a member observed with eligible coverage for one year; a member enrolled for six months can contribute roughly half a member-year under a standard earned-exposure convention. If the unit is a member-month, twelve months correspond to one person-year. An office visit count, hospital episode count, coded diagnosis count and incident disease onset are distinct numerators. Claims paid in a calendar year may include services incurred in earlier years, while claims incurred in a year may remain unpaid at extraction. Any forecast must state whether it uses service date, occurrence date, claim reporting or payment date.
A diagnosis code may indicate disease or merely the recognition and recording of disease during a healthcare encounter. Its probability of appearance depends on access to care, screening, clinical specialty, coding specificity, patient preferences, encounter volume and policy obligations. A score built from diagnoses alone is therefore partly a measurement of healthcare contact and institutional recording. The Obermeyer and colleagues (2019) study provides direct evidence that predicting medical cost as a proxy for illness need can yield divergent classification of similarly sick populations when their spending differs. This is a substantive warning about the chosen target rather than a purely statistical defect.
Hierarchical condition categories and payment factors
Condition categories frequently group diagnosis codes by clinical similarity and expected resource burden; hierarchies can suppress less severe diagnoses when a more severe diagnosis within the same group is present. CMS publishes a particular regulatory implementation through the 2026 Medicare Advantage rate announcement. For organizations outside specified exceptions, calendar-year 2026 calculations complete the phase-in and use the 2024 CMS-HCC model for 100% of the risk score. This is a U.S. program-year fact, not a statement about a universally optimal disease scoring system.
The U.S. individual and small-group market operates under a distinct HHS-operated model. The 2026 benefit-year HHS final rule describes recalibration on enrollee-level 2020, 2021 and 2022 EDGE data and program-specific adjustments. Medicare Advantage CMS-HCC, ACA HHS-HCC, a hospital Elixhauser index and a proprietary insurer severity score differ in purpose, coefficients, enrolled population, statutory basis, output normalization and date of applicability. Their numerical values are not exchangeable without deriving a validated transformation.
A generic linear risk-score representation is R_i = α + Σ_j β_j HCC_ij + γᵀZ_i + Σ_k δ_k interaction_ik, where HCC indicators, demographics and allowed interactions reflect the particular program specification. It is a mathematical template, not a claim to reproduce the proprietary or statutory coefficient file. A risk-transfer payment may then transform R_i using a program's published weights, normalization and budget-neutral balancing conditions. Analysts must read the applicable official policy before interpreting a model coefficient as an actual insurer dollar transfer.
Predictive utilization and expenditures
Clinical severity and historical utilization interact in medical-cost forecasts. Past emergency visits, inpatient stays, pharmacy claims, durable medical equipment use, home-care services and diagnostic clusters can convey different persistence profiles. Some predictors represent chronic disease burden; others capture transient procedures, administrative billing variations or a near-term treatment episode. The model must distinguish persistent risk from temporary shocks if its prediction horizon covers an entire future policy year.
A conventional aggregate outcome can be decomposed as E[C_i] = E[N_i] × E[S_i | N_i event] under a homogeneous event-cost definition with suitable weighting and conditional expectations. In a more realistic heterogeneous setting, E[C_i | X_i] = E[Σ_{j=1}^{N_i} S_ij | X_i], where severity may depend on frequency and encounter type. A mixture model can separately represent zero claims, routine utilization and high-severity events. A logarithmic model for expenditures requires retransformation and checks for the large share of zeros often present in short observation windows. A multi-outcome model can predict hospital admission and total cost simultaneously, but one target's validity never guarantees the other's calibration.
Claims history contains contractual information. An insurer-paid amount depends on plan design, negotiated prices, network, benefit limits, copayments and claim adjudication. Allowed charges and total economic resource use have different meanings. In long-term care and disability products, illness may generate severity-conditioned waiting periods, benefit elimination periods or contract-specific payment obligations. The same clinical condition has different actuarial outcomes under different products, so the intended coverage definition belongs inside the predictive model's data contract.
Classification Accuracy, Bias, and Institutional Incentives
Three mechanisms of systematic error
Classification error arises when a true clinical condition is not recorded, an administrative diagnosis is coded inaccurately, or the mapping from code to clinical grouping is inconsistent. Outcome-target error arises when the variable being predicted does not represent the welfare, medical need or insured cost under study. Population transport error arises when a model's validation data differ from the target insurance pool by enrollment, utilization, provider, coding system or calendar period. Each mechanism can be evaluated separately; improving coding precision does not automatically improve outcome-target alignment or out-of-sample transportability.
The 2019 population-health algorithm investigation by Obermeyer and colleagues showed how an otherwise apparently useful cost predictor could assign misleadingly equivalent risk to groups with differing baseline illness burden and care access. The choice between spending and need is an explicit target decision. Where the insurer's decision concerns reserve or expected claims cost, monetary spending may be a relevant contractual outcome. Where the decision allocates care-management resources to sick people, cost alone may mismeasure the desired allocation target. A rigorous scientific report states the target before evaluating subgroup disparities, rather than treating one demographic metric as a universal model evaluation.
Endogenous coding and upcoding
Codes can respond to financing incentives. In Geruso and Layton's Medicare study, economically meaningful differences in diagnosis-based risk scores were linked to organizational settings and coding intensity, with score differentials ranging from 6% to 16% in reported comparisons. The analysis separates enrollment sorting from coding changes and shows why a higher recorded risk score may reflect a changed reporting process as well as underlying patient health. The result is contextualized to the U.S. Medicare market and its historical institutional changes; it is not a cross-country or cross-product numeric coding correction.
CMS's Risk Adjustment Data Validation program audits whether submitted Medicare Advantage diagnoses have medical-record support and may recover unsupported excess payments. Such an audit checks one material compliance dimension. It does not by itself estimate the counterfactual set of diagnoses that would have been coded under an alternative payer incentive. A diagnosis can be clinically supported and still be more likely to be documented when financial incentives change. Research should therefore separate unsupported coding, defensible but selectively intensified coding, and actual disease incidence.
The 2025 Health Affairs study on survey-based adjustment examines combining patient-reported health measures with diagnosis-based HCC information to reduce coding incentives and selection distortions. Its findings support investigation of hybrid risk measures while also acknowledging survey nonresponse and sampling variability. The general problem is a tradeoff between outcome-relevant information and manipulability of the input. Even a variable recorded by a third party can be selected if individuals with worse health are less likely to respond.
Hypothetical sensitivity and audit arithmetic
Consider a synthetic population of 1,000 covered persons with 200 people having a specified true condition. If administrative records identify 160 of those people and correctly avoid a positive classification for 720 of the 800 people without the condition, sensitivity equals 160/200 = 0.80 and specificity equals 720/800 = 0.90. The positive predictive value is 160/(160+80) = 2/3 and the negative predictive value is 720/(720+40) = 0.9474. These values are sample-conditioned functions of prevalence and testing/reporting behavior. They are not empirical estimates for any particular diagnosis or insurer.
If the same classifier were applied to a different population with a 5% underlying prevalence while preserving the same hypothetical sensitivity and specificity, the positive predictive value would be (0.8×0.05)/[(0.8×0.05)+(0.1×0.95)] = 0.2963. An apparently stable clinical coding process can thus generate substantially different predictive value across populations because the base rate changes. This illustrates why both the diagnosis prevalence and the observation frame must be specified when transporting model validation results from one insurer to another.
A separate illustrative example considers recorded risk-score inflation. Suppose observed average scores increase from 1.00 to 1.10 while a consistently measured independent clinical outcome remains stable. This ten-percent score rise is compatible with a change in diagnosis capture, policyholder composition, true disease risk in dimensions not represented by the chosen clinical outcome, or some combination. Attribution requires comparable independent health measures and a research design that distinguishes enrollment selection from coding intensity. The arithmetic alone establishes no wrongdoing, and it supplies no empirical estimate of actual Medicare payment overstatement.
Equity, privacy and accountability
Subgroup validation requires measurement of calibration and clinically or contractually relevant errors within defined populations. A score's discrimination can remain high even if predicted expenditures are systematically lower than actual expenditures for certain subgroups. Whether an observed disparity indicates inappropriate risk treatment depends on the targeted outcome, lawful permitted variables, access to treatment, quality of diagnosis recording and market rules. A statistical audit should provide subgroup counts, uncertainty, plausible mechanisms and any mitigation's effect on overall performance and coverage eligibility.
Insurance databases contain sensitive clinical information, sometimes combined with diagnosis histories, pharmacy records and financial claims. Privacy controls, lawful purposes, retention and access restrictions must be implemented under the jurisdiction's applicable rules. Aggregated evidence may be published without exposing individual policyholders. Research data linkage must document consent or other lawful bases and test the possibility that deidentification procedures unintentionally alter the observed morbidity distribution. A model that is scientifically valid in a restricted dataset may not be legally or operationally deployable as a production underwriting or pricing tool.
Prospective Cost and Severity Forecasts
Prediction horizon, target and distribution
Prospective forecasting requires a model to use only information available before the outcome interval. For example, a policy-year 2027 predictor using diagnoses and benefit costs recorded through December 31, 2026 must not incorporate claims reported in 2027, retrospectively amended 2026 claims discovered later, or treatment decisions made during the outcome year unless explicitly modeling a revised-information retrospective analysis. A production model needs an information-cutoff timestamp for each feature and a data-extraction time separate from service and reporting dates.
A risk score used for payments and a predictive cost model may use different loss functions. If the target is a future binary hospitalization, logistic or survival models can estimate event probability. If the target is annual plan-paid medical expense, an estimator may optimize squared error, deviance or another proper scoring criterion for cost. If the decision is to identify members exceeding a high expense threshold, the relevant conditional tail probability and policy capacity can differ from the mean. For the same patient and diagnoses, predictions of 12-month cost, 24-month admission and long-term disability are scientifically distinct estimates.
For a covered population, the expected future total cost under model predictions is Σ_i w_i μ_i, where w_i is earned coverage exposure or another appropriate eligibility weighting, and μ_i is the expected cost for that exposure unit. If predicted annual costs assume full-year coverage but a member has only six months of expected enrollment, a weighting or an explicitly modeled partial-year trajectory is required. Incurred claims must be developed to a declared valuation date if late reporting affects outcome completeness. Clinical morbidity categories should remain consistent during the past and future periods or be bridged with documented recoding.
Mixture structures and high-cost tails
Health insurance costs are often concentrated among a small portion of insured lives. A modeling distribution may contain a positive probability of zero eligible claims, a routine-services component and a heavy-tailed expensive-episode component. A two-part representation is E[C | X] = Pr(C>0 | X)×E[C | C>0,X]. It clarifies which component of expected expense arises from the probability of incurring any claim and which from positive cost intensity. The factorization holds algebraically for nonnegative cost C but makes no assertion that these factors share predictors or that a particular parametric family is superior.
Suppose in an entirely synthetic group 20% of members have zero annual costs, 70% have $1,000 of costs and 10% have $15,000. The mean is 0.2×0+0.7×1,000+0.1×15,000 = $2,200. The conditional positive mean is $2,750 because 80% have positive expense; multiplying 0.8×$2,750 reproduces $2,200. The median is $1,000. For a hypothetical risk contract whose decision requires covering adverse high cost, reporting only $2,200 conceals the 10% tail event and its operational meaning.
Uncertainty needs a coherent distributional description. Conditional mean predictions can be compared with observed-to-expected ratios, calibration intercept and slope, mean absolute error and risk-stratified residuals. For a risk pool, predicted total cost should be validated at member, provider and insurer aggregation levels. If enrollment changes or prices trend, comparison must state whether predictions include medical inflation, contracted unit costs, clinical treatment changes or utilization intensity. The stochastic uncertainty of future claims is distinct from parameter uncertainty and from structural uncertainty about how a healthcare payment system will change.
Forecast backtesting and sample leakage
Chronological out-of-sample testing should reserve later enrollment and outcome windows for validation. A randomly partitioned encounter-level dataset can leak the same person's disease history between training and testing. A randomly partitioned claim line sample can place records from the same episode on both sides. A legitimate prospective backtest groups observations by individual or policyholder as appropriate and freezes predictor extraction dates. Time-series data from changing provider networks may require geographic and insurer-holdout validation in addition to a future temporal holdout.
The TRIPOD reporting standard provides a methodological anchor for transparent description of multivariable predictive model development, validation, outcome definition and predictors. Its application does not establish that a model has been externally validated; it is a reporting framework. A scientifically complete insurance risk-score study should publish the data dictionary, exclusion and inclusion rules, missingness, cutoffs, model tuning choices, versioned code maps and target prevalence alongside performance metrics.
A model can overfit systematic patterns in historical coding even when apparent discrimination is strong. Releasing a new coding software revision or a payment rule may change the distribution of predictors. If physicians alter encounter frequency, the same true disease burden can produce more diagnoses. This observation motivates out-of-period and out-of-institution evaluation, with a deployment monitoring plan for shifts in code frequency, present-on-admission flags, severity group counts, cost inflation and enrollment selection.
Data, Measurement, and Validation
Three linked data layers
A reproducible data structure separates diagnostic terminology, clinical events and contract accounting. Diagnosis terminology contains national ICD version, code, effective dates, official mapping version and grouping rules. Clinical event data contain patient or pseudonymous member, encounter date, source institution, medical-record support, primary and secondary diagnosis and relevant observation flags. Insurance data contain policy effective dates, eligibility, covered benefits, claims, allowed amounts, insurer payment, time of reporting, processing and reserves. Joining these systems by person and date requires explicit linkage logic, privacy controls and documented temporal ordering.
The effective date of a diagnosis must be distinguished from the date on which it was first recorded. A condition detected in February 2026 but entered into the insurer system in March is known at different times to the treating clinician and the risk-score administrator. Using a later claims extraction to reconstruct a December 2025 risk score can introduce look-ahead information. Data pipelines should store original event timestamps, ingestion timestamps, correction timestamps and the software release used to transform diagnosis codes. Reproducibility requires that historical scores can be recalculated using the exact information set originally available.
Accuracy, calibration and model comparison
Calibration evaluates whether scores quantitatively match the intended outcome; discrimination evaluates relative ranking. The Van Calster and colleagues analysis of calibration emphasizes the practical danger of treating discrimination as sufficient model quality. A perfectly ranked cohort can still have expected expenditures systematically overstated by 50%, and such error matters for budgets, transfer rules and reinsurance attachment. Conversely, an average pool calibration ratio of one can conceal major errors among high-risk members or protected groups.
For a cost model, an aggregate observed-to-expected ratio can be written O/E = (Σ_i C_i)/(Σ_i μ_i) for identical exposure and valuation bases, with O/E=1 representing equality of aggregate observed and predicted amounts in the tested sample. Consider two hypothetical groups each containing 100 earned person-years: group A has $100,000 observed cost against $50,000 predicted; group B has $50,000 observed against $100,000 predicted. The pooled O/E is 150,000/150,000=1, but group-specific O/E values are 2.0 and 0.5. A budget model could appear calibrated at portfolio level while allocating resources inaccurately between groups.
Calibration drift should be evaluated by risk decile, clinical severity, age, population class, product and geography, with intervals that account for claim variability and cluster dependence. Calibration slopes, residual quantile plots and group mean comparisons answer different aspects of the same problem. Discrimination measures for binary events are not directly interpretable as model performance in dollars. Proper scoring rules should match the target outcome and the decision's consequences; confidence intervals should account for repeated observations and high-cost concentration.
Source version comparability and national variation
AHRQ's CCSR v2027.1 mapping and Elixhauser v2027.1 release illustrate date-specific classification products. A dataset covering multiple fiscal years should either retain each year's contemporaneous code groupings or apply a harmonized reference mapping with transparent treatment of codes not valid in earlier years. Neither approach is universally superior: contemporaneous mappings reproduce the information structure at the time, whereas fixed retrospective mapping can aid comparability at the cost of availability and recoding assumptions.
A U.S. program-year rule must not be generalized to markets that use age bands, community rating, solidarity-based risk equalization or diagnosis systems based on different coding standards. Some national systems draw from person-level registries, others from health insurers' cost claims, hospital discharge files or population surveys. Each approach measures a mixture of clinical disease, healthcare access, reporting and payment incentives. International comparison should report population coverage, insurance penetration, disease taxonomy, legal model, outcome horizon and whether risk adjustment redistributes payments among private plans or funds public reimbursement.
Practical audit and reproducibility
A scientific reproducibility archive for an insurer risk-scoring project should retain (i) the versioned diagnostic mapping files and software license conditions, (ii) predictor and outcome data dictionaries, (iii) a reproducible temporal feature construction, (iv) exact eligibility criteria and coverage units, (v) statistical model specification and hyperparameter search decisions, (vi) source and transformation hashes, (vii) aggregate subgroup and temporal validation reports, and (viii) a clear model review record. This list describes a research protocol, not a claim that any actual private insurer has disclosed these items.
If external researchers cannot access sensitive person-level data, credible verification may still use privacy-safe synthetic datasets, open source mapping tables, simulated error mechanisms and independently recomputed summary statistics. However, synthetic replication does not establish empirical predictive accuracy in an actual insurance market. Regulatory or operational approval must depend on appropriately authorized data, legal review and institution-specific validation. The distinction should remain explicit in any scholarly dissemination.
International Synthesis and Scientific Conclusions
The integrated scientific interpretation
Morbidity classification is a data standard and inferential input. Predictive risk scoring is a stochastic or statistical estimate for a specified outcome and time horizon. Financial risk adjustment is an institutionally defined redistribution or payment calculation. These instruments interact because payment models draw on diagnostic categories and because the financial environment can change which diagnoses are recorded. Their distinct definitions enable a more accurate interpretation of both medical science and insurance economics.
The evidence supports a layered evaluation. The WHO classification establishes comparable medical language but permits national transitions. AHRQ grouping tools make a large diagnosis vocabulary tractable but change across code and software versions. Charlson and Elixhauser indices demonstrate how comorbidity can summarize prognosis and hospital outcomes, while their original dependent variables and derivation cohorts restrict transport to annual insurance costs. CMS risk-adjustment systems demonstrate how public programs turn disease recording into monetary transfers. The empirical literature of Geruso, Layton, Obermeyer and McWilliams identifies incentives, target mismatch and potential alternatives that have direct implications for policyholder classification.
A scientifically defensible prospective insurance model states its target, identifies its population and observation frame, records code and contract versions, predicts with information available at the correct cutoff, validates monetary and clinical performance separately, and documents fairness and coding incentives. A model cannot establish clinical need from cost prediction alone, and an audit of coding validity cannot estimate all incentives or counterfactual patient health. These are distinct tasks requiring different evidence.
Comparative research questions for further work
International research needs independent replication of diagnosis-based forecast performance across differently financed systems, including comparisons of ICD-11 and national ICD-10 transitions. Such studies should distinguish true morbidity progression from changes in health access or documentation. A second priority is methods for incorporating survey and clinical measures that are less sensitive to reimbursement incentives while handling nonresponse, measurement errors and legal access limits. A third is joint assessment of cost calibration, quality of care, disease burden and distributional equity after risk-adjustment rules change.
In reinsurance and prudential applications, future studies can examine how annual health-claim tails, disability transitions and medical inflation interact with risk score stratification and portfolio capital. The model should preserve uncertainty about policy duration and eligible payment obligations. A coherent international approach would compare inferential objects using shared statistical principles while respecting the contracts, regulations and information systems that make each insurance market unique.
The resulting general conclusion is affirmative: classification systems are indispensable scientific infrastructure for observing morbidity, and predictive health scores are useful when anchored in validated outcomes, transparent observation mechanisms and properly defined insurance exposures. Their effectiveness is determined through their measurable performance and institutional use. Accurate source classification, calibrated forecasting and carefully designed risk equalization can together strengthen the scientific understanding of health insurance portfolios and the decisions that govern them.
Outcome-specific comorbidity indices and their limits
The historical development of comorbidity scoring makes a substantive point about target choice. The AHRQ Elixhauser outcome-index study by Moore and colleagues (2017) estimated separate weighted indices for in-hospital mortality and 30-day readmissions using all-payer hospital administrative observations from 18 U.S. states in 2011–2012. The original publication reported a mortality-index c-statistic of 0.777 and a readmission-index c-statistic of 0.634 when used without other covariates. These outcomes differed even though the common source vocabulary was a similar set of clinical comorbidity indicators. Their discriminative accuracy is a property of the dependent variable, predictor specification and validation sample together.
A payer's annual expenditure prediction cannot inherit those reported c-statistics because dollar costs are neither in-hospital death nor readmission within 30 days. Mortality, short-term hospital use and annual insurer cost each have different censoring, benefit and exposure definitions. In particular, once a patient leaves the source hospital, the administrative file may contain incomplete observation of readmissions to other institutions, whereas a payer whose contracts cover the person may retain a wider claims history. The reverse can also occur if eligibility ends or insurance changes. This makes outcome-specific source provenance as important as the chosen comorbidity score.
An illustrative family of predictive models may include a separate logistic equation for next-year hospital admission, a cause-specific survival representation for death or disability transitions, and a heavy-tailed cost forecast conditional on alive-and-covered status. A common covariate such as diabetes severity may appear in all three. Its effect sizes, interactions and calibration targets are fitted separately. Combining predictions can improve a payer's analysis of uncertainty only if the joint event structure, competing risks and financial eligibility are treated coherently. A simple addition of independently predicted probabilities of overlapping events can count the same person twice.
Risk equalization, selection, and international comparison
Risk adjustment in competitive insurance markets seeks to reduce incentives for insurers to attract relatively profitable lives and avoid more expensive populations. The classic van de Ven and Ellis international synthesis relates expected expenditures, risk-adjusted payments, enrollment sorting and residual risk-sharing in health plan competition. The chapter's central contribution is institutional: a statistical prediction of expected cost becomes part of a payment or solidarity arrangement whose behavioral incentives depend on market rules.
Suppose a regulator transfers resources among insurers in proportion to their enrolled population's measured risk. One insurer may face a strong incentive to improve diagnosis completeness, another to improve care efficiency, and a third to alter enrollment composition. These incentives interact because the payment formula conditions on information that healthcare providers and payers can influence. Better prediction of historical spending reduces residual unexplained cost variation only within the model's population and observed treatment regime. It does not guarantee that an insurer cannot select or discourage chronically ill applicants through factors excluded from the score.
The appropriate analytic objective therefore combines predictive risk calibration with a separate incentive analysis. One can compare groups whose predicted costs are identical but whose true morbidity, access to care or claims patterns differ. If payment corrections depend on coded disease, assessing care access and coding procedures becomes necessary. Researchers can study the consequences through quasi-experimental payment changes, independent clinical surveys, insurance enrollment records and provider-level documentation. They should report whether differences concern real medical needs, diagnosis discovery, coding intensity, treatment price or formal redistribution.
The international evidence base also includes settings with universal statutory health coverage or risk equalization independent of individual underwritten premiums. Applying a U.S. Medicare Advantage payment coefficient to a solidarity-financed European sickness fund would incorrectly combine institutional rules. A comparative analysis should preserve each country's eligibility system, disease classification version, available risk adjusters, regulatory restrictions, ex-post compensation rules and outcomes used for recalibration. The statistical idea of adjusting expected risks is widely shared; the exact legal payments and observable risk profile remain local.
Time-at-risk and competing outcomes
Health insurance exposure requires a clinically and contractually valid risk interval. Suppose 1,000 members are eligible for one year, but 100 leave at midyear and the rest remain. Under uniform eligibility until departure, total earned member-years are 900 + 100 × 0.5 = 950. A claims count of 190 over those member-years corresponds to 0.20 claims per earned member-year. Dividing the same claims by 1,000 enrolled individuals would give 0.19 claims per member initially enrolled and answers a different question. Neither quantity alone explains morbidity risk if departures are correlated with disease or disability.
Competing-event analysis must also distinguish outcomes conditional on survival and policy persistence. In a health claims forecast, death may terminate a contract's eligible claim period, yet preceding terminal-care utilization may be unusually expensive. A disease progression model can treat death as an absorbing state, while a health insurer's financial model must incorporate the covered costs before termination. If the model censors death as if it were unrelated to condition severity, morbidity or expenditure probabilities can be biased. A prospective study should define its outcome as an unconditional covered cost, a cost conditional on continued enrollment, or an alternative clinically specified process.
A fully synthetic two-group example illustrates selection from coverage changes. Group A has 50 people with 12 eligible member-months and 50 with six; their exposure is 50×1+50×0.5=75 member-years. Group B has 100 full-year members and therefore 100 member-years. If both groups experience 30 eligible hospitalizations, annualized rates are 0.40 and 0.30 hospitalizations per member-year, respectively, despite identical raw counts. If the six-month enrollees departed after expensive hospitalization, even these annualized rates need interpretation conditioned on informative loss of coverage.
Quantitative calibration and score normalization
A technical risk score may be normalized to a population mean or annual reference benchmark. Suppose three synthetic members have unnormalized expected costs $500, $2,000 and $3,500 for equal one-year exposure. Their mean is $2,000. Dividing by the reference mean yields relative scores 0.25, 1.00 and 1.75, whose arithmetic average is one. Changing the reference population or coverage duration changes that normalization while leaving the person's estimated dollar expense unchanged. A value of 1.75 is therefore not an intrinsic medical severity unit; it denotes expected cost relative to a defined benchmark.
If the same three patients have annual insured amounts predicted under a different benefit design, their predicted scores may change even with identical medical need. The denominator of the index can be a plan's base cost, a statutory reference population or another declared scale. Correct interpretation requires a model identifier, coefficient file, prediction year, benchmark population and monetary price year. Without these the numerical label does not support comparison across programs, countries or time.
Clinical decision thresholds require a different type of normalization. A hospital admission probability of 0.25 denotes a predicted chance of an event over a stated horizon and is not an insurance relative-risk score of 0.25. The predicted number of admissions in a homogeneous cohort of 200 such members would be 200×0.25=50 under an unbiased model, but uncertainty depends on correlation and heterogeneity. The score's utility for triage depends on program capacity and the benefits and burdens of intervention; predictive accuracy alone does not specify a socially or clinically optimal selection threshold.
Transparent machine-learning research and monitoring
The TRIPOD+AI reporting update extends reporting expectations to regression and machine-learning prediction studies. Its scientific relevance is broader than an algorithm label. Prediction studies should disclose development and validation populations, clinical outcomes, predictor availability, preprocessing, missingness, calibration, subgroup evaluation, and intended deployment. A boosted tree or deep representation does not eliminate the requirement to verify a score's clinical meaning, insurer exposure and data provenance.
An operational morbidity model should maintain a data-version register with the mapping of old to new ICD codes, AHRQ code software releases, HCC model year, clinical record validation status and policy benefit definition. Drift monitoring can track diagnosis counts per encounter, proportion of coded severe conditions, share of incomplete present-on-admission flags, clinician or network mix, and the effect of a new documentation system. A change in any one of these variables can create spurious temporal risk differences. External validation should therefore compare performance under both source definitions and a harmonized retrospective map, stating which version was legitimately available at each forecast origin.
Calibration and fairness must also be evaluated at the decisions actually made. A score may predict insurer-paid costs accurately and still provide a misleading basis for allocating disease management. A score may indicate high clinical need and still be irrelevant to a product that excludes the event or has a waiting period. Modeling is more effective when the predicted event, use-case and observable input are selected together, then periodically retested with prospective data rather than inherited from an unrelated historical setting.
Reproducibility, adversarial tests and source uncertainty
A useful source audit asks whether every diagnosis score can be traced to contemporaneous input records, software mapping version, hierarchy assignment, normalization parameters, eligible coverage and the predicted outcome. The audit should challenge apparently stable predictors with counterfactual tests: remove post-outcome claims, substitute a new valid ICD map, change the service-date cutoff, allow for different provider coding intensity, alter the coverage window and separately re-estimate cost and clinical-need outcomes. These are sensitivity tests of an existing science framework, not a claim to know which perturbation will matter most in a given insurer.
Synthetic test cases can demonstrate formula correctness and software integrity while preserving privacy. For example, a simulated member with two diagnosis categories can be passed through the hierarchical mapping to confirm that a less severe category is suppressed when the official version requires it. A deliberately invalid code date can test whether the pipeline rejects the input rather than creating a false condition. These checks cannot prove clinical accuracy: authentic evaluation requires appropriate validated records and correctly observed outcomes.
The literature yields a stable conclusion about scientific accountability. A patient can be unwell without having a recorded diagnosis; a recorded diagnosis may be correct but prompted by different financial incentives; a patient with high morbidity can have low spending because of poor access; and a predictive score can rank patients well yet misstate absolute cost. Each phenomenon has a distinct observable implication. Combining WHO classification integrity, administrative coding supervision, genuine cohort validation and explicit incentive analysis is therefore essential for accurate interpretation of health insurance risk statistics.
Conclusion
Morbidity classification and prospective health risk scoring require consistent definitions of health conditions, eligibility, contractual outcomes and observation time. The reviewed evidence establishes the historical development of comorbidity indices, the version-specific character of contemporary code groupings, and the institutional differences between health prediction and regulated payment. It also demonstrates how coding incentives, access-dependent cost outcomes and population selection can change risk-score interpretation while clinical disease burden remains incompletely observed. An insurer or public regulator can make scientifically useful decisions from such scores by specifying valid targets, preserving classification and contract provenance, checking subgroup and temporal calibration, and independently validating the eventual use. These conclusions derive from established international classification, epidemiological research, documented U.S. regulatory examples and original reproducible hypothetical calculations; they retain explicit limits for cross-market transfer.
References
World Health Organization. (2025). ICD-11 implementation: Frequently asked questions. https://www.who.int/standards/classifications/frequently-asked-questions/icd-11-implementation
World Health Organization. (2026, April 30). ICD-11 fact sheet. https://www.who.int/publications/m/item/icd-11-fact-sheet
Agency for Healthcare Research and Quality. (2026). Clinical Classifications Software Refined for ICD-10-CM Diagnoses, v2027.1. Healthcare Cost and Utilization Project. https://hcup-us.ahrq.gov/toolssoftware/ccsr/dxccsr.jsp
Agency for Healthcare Research and Quality. (2026). Elixhauser Comorbidity Software Refined for ICD-10-CM, v2027.1. Healthcare Cost and Utilization Project. https://hcup-us.ahrq.gov/toolssoftware/comorbidityicd10/comorbidity_icd10.jsp
Centers for Medicare & Medicaid Services. (2025). 2026 Medicare Advantage and Part D Rate Announcement: Fact Sheet. https://www.cms.gov/newsroom/fact-sheets/2026-medicare-advantage-part-d-rate-announcement
Centers for Medicare & Medicaid Services. (2025). HHS Notice of Benefit and Payment Parameters for 2026 Final Rule. https://www.cms.gov/newsroom/fact-sheets/hhs-notice-benefit-and-payment-parameters-2026-final-rule
Centers for Medicare & Medicaid Services. (n.d.). Medicare Advantage Risk Adjustment Data Validation Program. https://www.cms.gov/data-research/monitoring-programs/medicare-risk-adjustment-data-validation-program
Charlson, M. E., Pompei, P., Ales, K. L., & MacKenzie, C. R. (1987). A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation. Journal of Chronic Diseases, 40(5), 373–383. https://doi.org/10.1016/0021-9681(87)90171-8
Elixhauser, A., Steiner, C., Harris, D. R., & Coffey, R. M. (1998). Comorbidity measures for use with administrative data. Medical Care, 36(1), 8–27. https://doi.org/10.1097/00005650-199801000-00004
van Walraven, C., Austin, P. C., Jennings, A., Quan, H., & Forster, A. J. (2009). A modification of the Elixhauser comorbidity measures into a point system for hospital death using administrative data. Medical Care, 47(6), 626–633. https://doi.org/10.1097/MLR.0b013e31819432e5
Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453. https://doi.org/10.1126/science.aax2342
Geruso, M., & Layton, T. (2020). Upcoding: Evidence from Medicare on squishy risk adjustment. Journal of Political Economy, 128(3), 984–1026. https://doi.org/10.1086/704756
McWilliams, J. M., Weinreb, G., Landrum, M. B., & Chernew, M. E. (2025). Use of patient health survey data for risk adjustment to limit distortionary coding incentives in Medicare. Health Affairs, 44(1), 48–57. https://doi.org/10.1377/hlthaff.2023.01351
Van Calster, B., McLernon, D. J., van Smeden, M., Wynants, L., & Steyerberg, E. W. (2019). Calibration: The Achilles heel of predictive analytics. BMC Medicine, 17, 230. https://doi.org/10.1186/s12916-019-1466-7
Collins, G. S., Reitsma, J. B., Altman, D. G., & Moons, K. G. M. (2015). Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): The TRIPOD statement. Annals of Internal Medicine, 162(1), 55–63. https://doi.org/10.7326/M14-0697
Moore, B. J., White, S., Washington, R., Coenen, N., & Elixhauser, A. (2017). Identifying increased risk of readmission and in-hospital mortality using hospital administrative data: The AHRQ Elixhauser Comorbidity Index. Medical Care, 55(7), 698–705. https://doi.org/10.1097/MLR.0000000000000735
Collins, G. S., et al. (2024). TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ, 385, e078378. https://www.bmj.com/content/385/bmj-2023-078378
van de Ven, W. P. M. M., & Ellis, R. P. (2000). Risk adjustment in competitive health plan markets. In A. J. Culyer & J. P. Newhouse (Eds.), Handbook of Health Economics, Vol. 1 (pp. 755–845). Elsevier. https://www.hopkinsacg.org/document/risk-adjustment-in-competitive-health-plan-markets/