Prospective Risk Adjustment Research Articles

Incorporating machine learning and social determinants of health indicators into prospective risk adjustment for health plan payments

BackgroundRisk adjustment models are employed to prevent adverse selection, anticipate budgetary reserve needs, and offer care management services to high-risk individuals. We aimed to address two unknowns about risk adjustment: whether machine learning (ML) and inclusion of social determinants of health (SDH) indicators improve prospective risk adjustment for health plan payments.MethodsWe employed a 2-by-2 factorial design comparing: (i) linear regression versus ML (gradient boosting) and (ii) demographics and diagnostic codes alone, versus additional ZIP code-level SDH indicators. Healthcare claims from privately-insured US adults (2016–2017), and Census data were used for analysis. Data from 1.02 million adults were used for derivation, and data from 0.26 million to assess performance. Model performance was measured using coefficient of determination (R2), discrimination (C-statistic), and mean absolute error (MAE) for the overall population, and predictive ratio and net compensation for vulnerable subgroups. We provide 95% confidence intervals (CI) around each performance measure.ResultsLinear regression without SDH indicators achieved moderate determination (R2 0.327, 95% CI: 0.300, 0.353), error ($6992; 95% CI: $6889, $7094), and discrimination (C-statistic 0.703; 95% CI: 0.701, 0.705). ML without SDH indicators improved all metrics (R2 0.388; 95% CI: 0.357, 0.420; error $6637; 95% CI: $6539, $6735; C-statistic 0.717; 95% CI: 0.715, 0.718), reducing misestimation of cost by $3.5 M per 10,000 members. Among people living in areas with high poverty, high wealth inequality, or high prevalence of uninsured, SDH indicators reduced underestimation of cost, improving the predictive ratio by 3% (~$200/person/year).ConclusionsML improved risk adjustment models and the incorporation of SDH indicators reduced underpayment in several vulnerable populations.

Demand elasticities and service selection incentives among competing private health plans

We examine selection incentives by health plans while refining the selection index of McGuire et al. (2014) to reflect not only service predictability and predictiveness but also variation in cost sharing, risk-adjusted profits, profit margins, and newly-refined demand elasticities across 26 disaggregated types of service. We contrast selection incentives, measured by service selection elasticities, across six plan types using privately-insured claims data from 73 large employers from 2008 to 2014. Compared to flat capitation, concurrent risk adjustment reduces the elasticity by 47%, prospective risk adjustment by 43%, simple reinsurance system by 32%, and combined concurrent risk adjustment with reinsurance by 60%. Reinsurance significantly reduces the variability of individual-level profits, but increases the correlation of expected spending with profits, which strengthens selection incentives.

Multivariate Risk Adjustment of Primary Care Patient Panels in a Public Health Setting: A Comparison of Statistical Models.

We compared prospective risk adjustment models for adjusting patient panels at the San Francisco Department of Public Health. We used 4 statistical models (linear regression, two-part model, zero-inflated Poisson, and zero-inflated negative binomial) and 4 subsets of predictor variables (age/gender categories, chronic diagnoses, homelessness, and a loss to follow-up indicator) to predict primary care visit frequency. Predicted visit frequency was then used to calculate patient weights and adjusted panel sizes. The two-part model using all predictor variables performed best (R = 0.20). This model, designed specifically for safety net patients, may prove useful for panel adjustment in other public health settings.

Population Based Resource Allocation: The Use of Hybrid Risk Adjustment

The emphasis on integrated care implies new incentives that promote coordination between levels of care. Considering a population as a whole, the resource allocation system has to adapt to this environment. This research is aimed to design a model that allows for morbidity related prospective and concurrent capitation payment. The model can be applied in publicly funded health systems and managed competition settings. Methods: We analyze the application of hybrid risk adjustment versus either prospective or concurrent risk adjustment formulae in the context of funding total health expenditures for the population of an integrated healthcare delivery organization in Catalonia during years 2004 and 2005. Results: The hybrid model reimburses integrated care organizations avoiding excessive risk transfer and maximizing incentives for efficiency in the provision. At the same time, it eliminates incentives for risk selection for a specific set of high risk individuals through the use of concurrent reimbursement in order to assure a proper classification of patients. Conclusion: Prospective Risk Adjustment is used to transfer the financial risk to the health provider and therefore provide incentives for efficiency. Within the context of a National Health System, such transfer of financial risk is illusory, and the government has to cover the deficits. Hybrid risk adjustment is useful to provide the right combination of incentive for efficiency and appropriate level of risk transfer for integrated care organizations.

Hybrid Risk Adjustment for Total Health Expenditures

Objectives: Efficient use of health resources has become a major concern for health policy makers and researchers when health expenditures are continously increasing. Prospective risk adjustment, using ex ante information and relating health expenditures and individual or population characteristics and morbidity, is generally recommended as a tool for controlling the cost and promoting incentives for efficiency in the health provision. However, it also can promote incentives for risk selection. Concurrent risk adjustment, using ex post or actual information on health conditions and costs, improves the access and quality of the health services provision and avoids the incentives for risk selection although it also reduces the efficiency. This paper follows a relatively new literature on risk adjustment in which a mixed formula, using both prospective and concurrent risk adjustment, seeks to maximize the efficiency incentives involved in the prospective risk adjustment and minimize the incentives for selection through a concurrent risk adjustment for high risk individuals suffering a set of specific conditions. Methods: We utilize individual data on total health expenditures, including pharmaceutical, primary care, and hospital and specialised attention expenditures, and health conditions for the population belonging to an integrated healthcare delivery organization in Catalonia in years 2004 and 2005 in order to show the gain of a hybrid risk adjustment versus either an only prospective or only concurrent risk adjustment formula. The individual health conditions are classified by the Clinical Risk Group classification system. Results: Concurrent risk adjustment obtains better results in terms of R-squared than prospective risk adjustment although it decreases incentives for efficiency. However, the use of a hybrid formula with concurrent payments for individuals with specific conditions can solve the problem of selection while incentives for efficiency are maximum for a high proportion of the population. Conclusions: A hybrid risk adjustment system is expected to solve the trade-off between efficiency in provision and the risk selection problem and thus promoting efficient use of health resources. The interest of a hybrid risk adjustment proposal relies also on its feasibility and easy implementation.

The Performance of Administrative and Self‐Reported Measures for Risk Adjustment of Veterans Affairs Expenditures

To evaluate the performance of different prospective risk adjustment models of outpatient, inpatient, and total expenditures of veterans who regularly use Veterans Affairs (VA) primary care. We utilized administrative, survey and expenditure data on 14,449 VA patients enrolled in a randomized trial that gave providers regular patient health assessments. This cohort study compared five administrative data-based, two self-report risk adjusters, and base year expenditures in prospective models. VA outpatient care and nonacute inpatient care expenditures were based on unit expenditures and utilization, while VA expenditures for acute inpatient care were calculated from a Medicare-based inpatient cost function. Risk adjusters for this sample were constructed from diagnosis, medication and self-report data collected during a clinical trial. Model performance was compared using adjusted R2 and predictive ratios. In all expenditure models, administrative-based measures performed better than self-reported measures, which performed better than age and gender. The Diagnosis Cost Groups (DCG) model explained total expenditure variation (R2=7.2 percent) better than other models. Prior outpatient expenditures predicted outpatient expenditures best by far (R2=42 percent). Models with multiple measures improved overall prediction, reduced over-prediction of low expenditure quintiles, and reduced under-prediction in the highest quintile of expenditures. Prediction of VA total expenditures was poor because expenditure variation reflected utilization variation, but not patient severity. Base year expenditures were the best predictor of outpatient expenditures and nearly the best for total expenditures. Models that combined two or more risk adjusters predicted expenditures better than single-measure models, but are more difficult and expensive to apply.

Identifying Future High-Healthcare Users

Diagnosis-based risk-adjustment measures are increasingly being promoted as disease management tools. We compared the ability of several types of predictive models to identify future high-risk older people likely to benefit from disease management. Veterans Health Administration (VHA) data were used to identify veterans ≥65 years of age who used healthcare services during fiscal years (FY) 1997 and 1998 and who remained alive through FY 1997. This yielded a development sample of 412 679 individuals and a validation sample of 207 294. Prospective risk-adjustment models were fitted and tested using Adjusted Clinical Groups (ACGs), Diagnostic Cost Groups (DCGs), a prior-utilization model (prior), and combined models (prior + ACGs and prior + DCGs). Prespecified high use in FY 1998 was defined as ≥92 days of care (top 2.2%) for an individual (i.e. the number of days during the year in which an individual received inpatient or outpatient healthcare services). We developed a second outcome, defined as ≥164 days of care (top 1.0%), to explore whether changing the criterion for high risk would affect the number of misclassifications. The diagnosis-based models performed better than the prior model in identifying a subgroup of future high-cost individuals with high disease burden and chronic diseases appropriate for disease management. The combined models performed best at correctly classifying those without high use in the prospective year. The utility for efficiently identifying high-risk cases appeared limited because of the high number of individuals misclassified as future high-risk cases by all the models. Changing the criterion for high risk generally decreased the number of patients misclassified. There was little agreement between the models regarding who was identified as high risk. Health plans should be aware that different risk-adjustment measures may select dissimilar groups of individuals for disease management. Although diagnosis-based measures show potential as predictive modeling tools, combining a diagnosis-based measure with prior-utilization model may yield the best results.

Risk adjustment effect on stroke clinical trials.

The ischemic stroke population is heterogeneous. Even in balanced randomized trials, patient heterogeneity biases estimates of the treatment effect toward no effect when dichotomous end points are used. Risk adjustment statistically addresses some of the heterogeneity and can reduce bias in the treatment effect estimate. The purpose of this study was to estimate the treatment effect of tissue plasminogen activator (tPA) in the National Institute of Neurological Disorders and Stroke (NINDS) tPA data set with and without adjustment for baseline differences. Using a prespecified predictive model, we calculated unadjusted and risk-adjusted odds ratios (ORs) for favorable outcome for the Barthel Index, National Institutes of Health Stroke Scale, and Glasgow Outcome Scale for the patients in the NINDS tPA stroke trial. To assess the importance of the difference, a new sample size was calculated through the use of the risk-adjusted analysis. We analyzed 615 subjects. The ORs for the Barthel Index were 1.76 (unadjusted) and 2.04 (adjusted). The National Institutes of Health Stroke Scale and Glasgow Outcome Scale analyses also demonstrated increased ORs after adjustment. The estimated sample size required for the adjusted comparison was 13% smaller than the unadjusted sample. Risk adjustment in this data set suggests that the true treatment effect was larger than estimated by the unadjusted analysis. Stroke clinical trials should include prospective risk adjustment methodologies.

The costs of decedents in the Medicare program: implications for payments to Medicare + Choice plans.

To discuss and quantify the incentives that Medicare managed care plans have to avoid (through selective enrollment or disenrollment) people who are at risk for very high costs, focusing on Medicare beneficiaries in the last year of life-a group that accounts for more than one-quarter of Medicare's annual expenditures. Medicare administrative claims for 1994 and 1995. We calculated the payment a plan would have received under three risk-adjustment systems for each beneficiary in our 1995 sample based on his or her age, gender, county of residence, original reason for Medicare entitlement, and principal inpatient diagnoses received during any hospital stays in 1994. We compared these amounts to the actual costs incurred by those beneficiaries. We then looked for clinical categories that were predictive of costs, including costs in a beneficiary's last year of life, not accounted for by the risk adjusters. The analyses were conducted using claims for a 5 percent random sample of Medicare beneficiaries who died in 1995 and a matched group of survivors. Medicare is currently implementing the Principal Inpatient Diagnostic Cost Groups (PIP-DCG) risk adjustment payment system to address the problem of risk selection in the Medicare+Choice program. We quantify the strong financial disincentives to enroll terminally ill beneficiaries that plans still have under this risk adjustment system. We also show that up to one-third of the selection observed between Medicare HMOs and the traditional fee-for-service system could be due to differential enrollment of decedents. A risk adjustment system that incorporated more of the available diagnostic information would attenuate this disincentive; however, plans could still use clinical information (not included in the risk adjustment scheme) to identify beneficiaries whose expected costs exceed expected payments. More disaggregated prospective risk adjustment methods and alternative payment systems that compensate plans for delivering care to certain classes of patients should be considered to ensure access to high-quality managed care for all beneficiaries.

Risk adjustment and risk sharing: the Israeli experience

Israel, like several other countries, introduced a national risk adjusted capitation system during the 1990s. However, the Israeli move was drastic, implementing from the beginning a fully prospective risk adjustment scheme based on age, supplemented by a 100% five condition-specific risk sharing. That scheme, together with open enrollment (periodic switching options), was intended to transform an unregulated competitive health insurance market, characterized by adverse selection and preferred risk selection, into managed competition assuring quality of care, efficiency and fairness. This paper presents the Israeli experience during the first 6 years of the reformed system, focusing on issues related to the risk adjustment and risk sharing arrangements.

The prospective risk adjustment system.

The Episode Classification System is intended to perform two tasks. First, it will be a prospective capitation risk adjuster and predict future health care costs. It will do this by assigning each individual a single capitation risk adjustment category based on an analysis of the medical history and of health care services rendered during a specific period of time. Second, the Episode Classification System will create retrospective severity adjusted Episodes of illness or Episodes of Care. These latter Episodes will provide a framework for relating patient characteristics to the amount, type, and duration of services provided during the treatment of a specific disease. These Episodes will give users the ability to understand past costs and the risk of mortality. As such they will form the basis for provider profiling by allowing users to analyze a complete clinical episode.

Expenditure models for prospective risk adjustment: choosing the measure appropriate for the problem.

Risk adjustment models for health care expenditures typically have been disappointing in their predictive ability for individual enrollees. However, practical applications of risk adjustment often involve groups, not individuals. Therefore, model performance on groups of enrollees may be more meaningful than performance on individuals. The authors propose a new measure of model fit, the sorted-group R2, which measures the correlation between predicted and actual cost averages for groups of enrollees sorted by risk prior to grouping. The authors use this tool to evaluate four models with different explanatory variables and functional complexities. Results indicate that even a model which performs poorly for individual subscriber units and for unselected groups can indeed distinguish between groups of subscriber units where there has been moderate risk selection.