Health risk can be defined as the expected health care costs or utilization of an individual or groups of individuals. Risk assessment is the measurement of that risk, linking the characteristics of an individual to their current and future resource use or clinical outcome. Risk adjustment is the mechanism that adjusts payment rates or measure results to reflect the differences in risks as measured by the risk assessment process. Risk assessment and adjustment tools have a number of applications for an ACO. They include setting payment rates more accurately, adjusting financial performance to reflect differences in population health status, measuring provider performance fairly across patient
populations, and identifying high-cost patients for care management.
Basics of Risk Assessment Models
Models of health risk assessment vary along a number of dimensions, including the type of data required, the applications for which the models are being used, the modeling techniques, the outcomes to be measured, and the model outputs. Many risk models used in health care provide an assessment of cost or utilization outcomes for an individual – particularly when applied for financial analysis or rate setting. A number of health care models also have been developed that focus on specific patient events, such as the likelihood of an adverse event or mortality related to a clinical intervention. This section describes modeling approaches that are primarily used in assessing the risk of financial outcomes.
The two basic components of most health risk models are risk markers and risk weights. Risk markers describe demographic clinical and other characteristics that distinguish one individual from another. Risk weights translate those markers into a measure of risk. Equation 1 below illustrates the basic structure of such a model. Riski is the measure of risk for individual i, Markeri,m describes the presence of risk marker m for individual i, and Wtm is the weight assigned by the model to Marker m. The risk for an individual is the sum of the weights for all of their risk markers observed – often expressed as a relative score centered around 1.00, where 1.00 represents the average risk for a reference population. Using this approach, a risk score of 0.50 represents a level of risk one half of that average, a risk score of 2.0 twice that average, and so on. The risk for a group of individuals can be expressed as the average risk for all individuals in that group.
(1) Riski = S Wtm*Markeri,m
Modeling and Data – Clinically-Based and Demographic Risk Models
Models of health risk assessment can be simple or quite complex. A demographic or age-sex model is an example of a simple model, where the markers describe the age and gender category for the individual and the weights represent the relative expected costs or resource utilization of individuals in that category. Demographic models are straightforward to administer, but offer little clinical information and perform poorly in terms of predictive accuracy for most applications (i.e., a demographic risk model does not predict costs well for an individual). Clinically-based models employ markers that leverage patient diagnoses and, in some cases, the use of medical services. In these models, the markers describe the presence of a clinical diagnosis or utilization event for an individual, and the weights represent
the incremental contribution to the risk of having that marker. Given the richness of the clinical information employed and the strong link between patient health status and expected resource use, the clinically-based models provide greater predictive accuracy, versus models using only demographic information.2
Many health risk assessment models leverage information readily available from administrative medical and pharmacy claims and enrollment data. Some models use clinical lab results, in particular where high-risk prediction is the objective. Models that employ pharmacy data or results from member surveys can also provide value when medical claims data are not available or incomplete.
Most risk models use data for a 12-month period to identify markers for an individual. The risk weight assigned to a marker is typically predefined by the model developer and can vary depending on the outcome being measured and available data. Software that encapsulates the health risk methodology and weightings is often employed to produce the risk assessment results. Relevant data are prepared and processed using the software to produce the risk scores by individuals and to develop information that is useful in understanding the measured risk.
Applications for Health Risk
In selecting a health risk assessment model, it is important to recognize the intended business use of the model. For example, risk assessment can be applied either retrospectively or prospectively. Both types of models have importance for ACOs. Retrospective or concurrent models use risk markers for an individual in a base year to measure risk for that same period of time. A prospective application uses markers in a base year to measure risk for a future time period. Retrospective models are most often used for comparing provider and health plan performance. Prospective models are often applied when setting payment rates and to
stratify populations for care intervention and disease management.
The intended business use is an important consideration when selecting a model – using the right tool for the right purpose. As described above, prospective models are often applied when setting payments. The risk assessment model used by the Centers for Medicare & Medicaid Services (CMS) to reimburse health plans for serving Medicare beneficiaries is one example.3 Many state Medicaid programs use similar models.4 As a third example, an ACO’s target benchmark may be risk-adjusted prospectively based on future risk expectations. In addition to being prospective, risk models used in payment most often include risk markers based on patient diagnoses and exclude markers describing utilization events. Where discretion is present, the risk assessment formula will not reward or penalize treatment decisions, such as the decision to admit a patient to the hospital, to perform a surgery, or to prescribe a medication. In this way, the payment systems provide appropriate incentives for medical practice.
A second business use for health risk assessment is high risk prediction. “Predictive models” are designed to identify patients of the highest risk in a population and to provide information useful in supporting care and health management. These models leverage all available, useful information – including diagnoses, history of medical service use, utilization events, and lab results – to identify patients who are expected to consume significant resources in the future and are good candidates for care management. As a result, these models provide enhanced predictive ability relative to models based exclusively on diagnoses.
Example – Assignment of Health Risks Using Three Different Models
A few examples can help illustrate how risk assessment models work and how they can be applied by an ACO. Table 1 summarizes the calculation of risk for three individuals using three
different approaches to health risk assessment. The first two models are clinically-based and describe retrospective and prospective applications. The third model is a demographic-only, age-sex model. The risk markers observed for each individual are shown along with the weights assigned to each marker for each model. As shown, the clinical models use markers based on diagnostic information and also give some weight to age and gender in calculating prospective risk. The age- sex model uses only the individual’s age and sex to assess risk.
The first individual, a 58-year-old male, is observed to have diabetes, congestive heart failure (CHF), ulcers, and a dermatology condition. Each of these markers receives a numeric weight describing the contribution of that marker to risk.5 The sum of the weights across the markers observed is the risk score for the individual. The total risk scores of 6.632 and 6.741 for the retrospective and prospective applications suggest a level of risk for this 58-year-old male is more than six times that of the average individual in the reference population.
The risk score based on the age-sex model provides a different assessment and is markedly lower than that for the clinical models. In an age- sex model, all individuals in the category of male age 55-64 are assigned the same risk factor of 2.25, indicating that the expected cost of health care for individuals in this category is more than twice that of the average individual in the reference population. A comparison of these models suggests that using an age-sex model alone would likely underestimate their health risk. Finally, note that the retrospective and prospective risk scores for this example are similar, both driven by the presence of two chronic, ongoing conditions (diabetes and CHF) that comprise the majority of the patient’s risk. The calculation of risk for the second and third individuals can be interpreted in a similar manner. The observed difference between retrospective and prospective risk for the 35-year-old female illustrates the impact of an acute event (pregnancy) on the two clinical models. The pregnancy is a significant driver of expected costs for the current year, but has a negligible impact on risk for the future year.