MANAGERIAL. APrediction Model for Targeting Low-Cost, High-Risk Members of Managed Care Organizations

Transcription

1 APrediction Model for Targeting Low-Cost, High-Risk Members of Managed Care Organizations Henry G. Dove, PhD; Ian Duncan, BPhil, BA; and Arthur Robb, PhD Objective: To describe the development and validation of a predictive model designed to identify and target HMO members who are likely to incur high costs. Study Design: Split-sample multivariate regression analysis. Patients and Methods: We studied enrollees in a member HMO with 1 claim in 1998 and The prediction model uses a combination of clinical and behavioral vaiables and 1998 and 1999 claims data. The prediction model was applied and used to rank low-cost patients (1998 cost <$2000) according to their estimated probability of incurring costs $2000 in For prospective testing, we applied our models to data that are not available in advance. The same prediction model was applied to rank a different set of lowcost patients (1999 cost <$2000) according to estimated probability of incurring costs $2000 in Because the predictions were used for disease management purposes, the outcomes of a randomly selected control group not intervened on for the disease management program was analyzed. The predictive accuracy of the model was tested by comparing the percentages of targeted vs all low-cost patients who incurred high costs in the subsequent year. Results: Of the low-cost, top-ranked 1998 patients, 47.8% incurred high ( $2000) medical expenses in 1999 vs 14.2% of randomly selected patients who were low cost in Of the top-ranked 1999 patients, 39.7% incurred high costs in 2000 vs 12.2% of the randomly selected low-ranked patients. Conclusions: The prediction model successfully identifies low-cost, high-risk patients who are likely to incur high costs in the next 12 months. (Am J Manag Care 2003;9: ) Presently, most managed care organizations (MCOs) manage population risk by applying disease management or case management programs. The targeting of specific events or diseases is a proxy for risk because members who were hospitalized or have 1 specific ( index ) diseases (such as asthma, congestive heart failure, or diabetes mellitus) experience higher claims costs than the average for the plan. However, there are many ways of trying to identify potentially high-cost patients, including using diseases (such as AIDS or renal failure) likely to require expensive treatments, previous hospital or emergency department utilization, or the level or rate of increase in recent medical expenses. Different interventions may be applied to each group depending on the identification method. Interventions are used on these members with the objectives of reducing costs and improving outcomes. This approach has 2 obvious disadvantages. First, some patients without one of the index diseases are likely to incur high medical expenses. Second, diseased populations are not homogeneous; thus, members whose medical conditions are well controlled are statistically less likely to experience higher-thanaverage costs in the future (and, therefore, are less likely to benefit from any intervention). The traditional approach has a third shortcoming: because of the intensity of resources required to manage the potentially high-risk members, health plans limit their interventions either to case management of the current high-cost patients or to broad, often untargeted, disease management programs. Population risk management consists of 3 components: Identification of high-risk populations or subpopulations. Use of specific interventions in the high-risk group to reduce the resource utilization and cost of the group. Application of pricing and underwriting techniques to convey financial signals to plan members and sponsors. A group s health insurance premiums, for example, could be based on their estimated future utilization rather than on the traditional underwriting method of a projection of historical costs. From the Division of Health Policy and Administration, Department of Epidemiology and Public Health, Yale University, New Haven, Conn (HGD); Lotter Actuarial Partners, Inc, New York, NY (ID); and LandaCorp, Inc, Montclair, NJ (AR). This study was suported by Landacorp, Inc, Atlanta, Ga. Corresponding author: Henry G. Dove, PhD, Division of Health Policy and Administration, Department of Epidemiology and Public Health, Yale University, 60 College St, PO Box , New Haven, CT dove@worldnet.att.net. VOL. 9, NO. 5 THE AMERICAN JOURNAL OF MANAGED CARE 381

2 Table 1. Distribution of Enrollees by Expense Categories and Percentage of Total Expenditures in 1998 and Medical Expense Average Average Category Per Capita Total Total Per Capita Total in 1998 Expenses, $ Enrollment, % Expenditures, % Expenses, $ Expenditures, % Low (<$2000) Medium ($2000-$24 999) High ( $25 000) It is well known in healthcare that a small percentage of the members of a health plan consume a significant percentage of its resources; it is assumed, often incorrectly, that the behavior of this minority is replicated period after period. This assumption is refuted by the data in Table 1. These data, representing approximately members of a member health plan, show that the highest-cost members, those costing $ in incurred claims in 1998 (and, hence, candidates for traditional case management), represented 1% of the members but 21% of the cost in 1998; in the following year, this cohort consumed only 7% of total plan costs. Conversely, enrollees in the lowest-cost medical class, with costs <$2000 in 1998, accounted for 58% of all costs in 1999 (Table 1). The phenomenon being illustrated here, regression to the mean (in which the resource consumption of most high-cost patients generally decreases, even in the absence of any intervention), is well known in health plans but seems to be overlooked as plans attempt to find and manage their high-risk members. In population risk management, a variety of clinical and behavioral variables are used to rank each patient according to his or her estimated probability of incurring high medical expenses in a subsequent period. This article describes the design, development, and validation of a prediction model targeting selected members in a large regional HMO in the southwest United States. The goal is to identify lowcost patients (<$2000 in the base year ) who are likely to become high-cost patients (in the absence of any intervention) in the subsequent year. OTHER COMPONENTS OF POPULATION RISK MANAGEMENT Two other components of a population risk management strategy are interventions (programs that aim to change patient behavior, healthcare delivery processes, and patient outcomes through education and coaching) and pricing and underwriting (techniques that aim to change behavior through price signals). Selecting appropriate, cost-effective intervention strategies for targeted patients and reflecting prospective risk in pricing and underwriting decisions (to the extent allowed by regulatory and ethical constraints) are related challenges that are topics for future articles. STUDY POPULATION, DATA SOURCES, AND GOALS The prediction model was developed on patients who were enrolled in a large HMO in 1998 and 1999 and had at least 1 medical or pharmacy claim in both years. Patient medical claims and pharmacy claims incurred in the subsequent year were the source of outcomes for these patients. Population risk management aims to reduce the cost of the targeted population; although this may result in improved health markers, the objective of risk management is to improve financial outcomes for the health plan. No reviews of patient medical records, questionnaires, or special surveys are conducted. Reliance on administrative data is an efficient, low-cost approach that is ideal for population risk management. DATA PREPARATION Before creating a prediction model, the demographic data and medical and pharmacy claims of MCO patients were checked for completeness, integrity, and consistency. Data preparation included several activities: Adopting an adequate run-out period (in this case 4 months), determined using standard actuarial methods for assessing the completeness of incurred claims data. 382 THE AMERICAN JOURNAL OF MANAGED CARE MAY 2003

3 Targeted Care for Low-Cost Members Separating medical expenses into categories such as professional services, hospital inpatient services, hospital outpatient services, laboratory and diagnostic tests, and pharmaceutical items. Data checks were used to measure the data's internal consistency against benchmarks. Data that were rejected based on the diagnostic reports were resubmitted and rechecked. Distinguishing the employee (policyholder) and his or her dependents by assigning each enrollee a unique member number. The claims for each patient were collected, tabulated, and grouped into a patient-centered database. Identifying covered charges that reflected only those medical services that the MCO was obligated to pay. Using the amount the MCO paid for each claim rather than the billed or charged amount or the amount patients paid. Members of the health plan are subject to different plan designs, with variable copays, limits, exclusions, and so on, as set by their employers. It can be argued that a member s behavior is influenced by the specific design of the benefit; however, this is one of many variables that we do not recognize in our modeling. The potential for incorrectly assigning a member (as high risk or not high risk) based solely on plan design is considered to be minor. OUTCOME MEASURES The concepts of patient risk and outcomes require clarification because these terms are often unclear or imprecise. Outcomes is a broad term that can have very different meanings for clinicians, epidemiologists, utilization management specialists, risk managers, actuaries, and quality assurance professionals. Outcomes may pertain to functional status, patient satisfaction, mortality, hospital utilization, or cost. Consistent with the objective of population risk management, our sole focus was on financial outcomes, as defined by total incurred medical claims in a 1-year period. USE OF A THRESHOLD COST FOR HIGH-RISK PREDICTION FOCUS ON MEMBERS WITH BASE YEAR COST <$2000 Our analyses focus on a $2000/<$2000 criterion, which raises 2 issues: Why a threshold? Why not predict dollar cost? If we are going to recognize a threshold, why $2000 and not another dollar amount? A cutoff value of $2000 for annual patientincurred medical (including pharmacy) expenses was used to establish a binary outcome variable, that is, enrollees were either low-cost patients (<$2000) or high-cost patients ( $2000). A component of our definition of patient risk, the probability that a patient will incur paid medical expenses $2000 in the succeeding 1-year period, is based on a patient s characteristics in the base year. The base year in this article is 1998, and the subsequent year is The $2000 threshold was chosen for practical and statistical reasons and to differentiate our approach from the traditional approach used in disease management and case management. Disease management and case management based methods for identifying high-risk members rely on diagnoses (using International Classification of Diseases, Ninth Revision, Clinical Modification, codes) and events (hospitalizations, emergency department visits, etc). Frequently, members traditionally labeled as high risk have base year costs $2000 and often considerably >$2000. Conversely, a population with costs <$2000 in the base year is not traditionally thought of as being at high risk. Yet, our data and models show that this population contains a considerable number of high-risk members. The models are general and may be applied to any dollar threshold. Because the subpopulation of low-cost consumers is large, the targeting process will deliver a large number of potential high-risk intervention candidates who generate a significant percentage of a health plan s total expense. Conversely, if the focus is on only the repeaters from the high-cost subgroups, relatively little total cost is identified for intervention and management. From a business perspective, health plans are interested in identifying subgroups whose management will have a noticeable effect on the plan s overall financial results. Our definition of risk extends beyond cost. In our work, we used a compound dependent variable designed to capture 2 important risk factors: the absolute amount of claims and the incidence or concentration of those claims. Thus, a patient whose claims are more highly aggregated represents a higher risk than a patient who has the same absolute amount of claims spread over 12 months. In general, healthcare costs are transformed if they are to be used as a dependent variable. As pointed out in the recent Society of Actuaries study of risk adjusters, 1 large claims should be truncated. Owing to the distribution, costs must be logarithmically transformed. In a sense, the binary variable ( /<$2000) is the simplest transform. This form of transform in turn allows VOL. 9, NO. 5 THE AMERICAN JOURNAL OF MANAGED CARE 383

4 us to include additional information (on the concentration of claim amount) in the dependent variable. There is a simple reason why categorical variables are advantageous: raw dollar amounts are not uniformly calibrated from patient to patient. In other words, 2 patients may have been subject to the same medical services but may incur significantly different costs due to differences in plan design, choice of providers, or provider billing practices. Because significant cost differences can be due to exogenous factors, it is preferable to use cost as a relative approximation of risk rather than as an absolute measure. In this sense, the $2000 threshold is merely a simple categorical variable. The specific threshold of $2000 was chosen for 2 reasons. One is a practical consequence of the fact that this model was used to select patients for an intervention. In general, to show positive return on investment in an intervention program, as well as to simply have statistically measurable outcomes, interventions require significant numbers of potential enrollees. Use of the $2000 threshold casts a wide net for intervention targets. Second, because higher costs are driven by events, it was deemed advantageous to choose a threshold that correlated strongly to the existence of an event as both a positive and negative proxy indicator. The $2000 threshold fulfills this need. PREVIOUS RESEARCH Few articles have appeared in peer-reviewed journals that attempt to identify patients likely to incur high medical expenses in a subsequent year among patients who incurred modest medical expenses in the preceding or base year. Meenan and his colleagues 2 at the Kaiser Permanente Center for Health Services Research developed and tested 3 models to identify high-cost risk status in a large sample of approximately HMO members from 3 health plans. LoBianco et al 3 studied high-cost Medicaid users. Forman and his colleagues 4,5 and Lynch et al 6 studied repeaters. We believe that our analytical technique addresses an important understudied group: individuals with no obvious previous costs who are less likely to be identified in the other studies referred to previously. The more common focus of research using administrative data sets has been for the purpose of risk adjustment The statistical methods, data sources, and variables used for identifying high-cost patients and creating risk adjusters are similar. However, the goal of risk adjustment is not to identify individual patients with high-cost conditions or to intervene in their care. The main objective of risk adjustment is to accurately predict the average annual expenditures for an individual patient to redistribute premiums to health plans. Thus, the coefficients or groupings models that researchers have created for risk adjustment are not designed for identifying high-cost patients, although these models also use disease categories, comorbidities, and demographic variables. Targeting the Right Patients in Population Risk Management The study objective was to develop a prediction model using variables in medical and pharmacy claims data sets to identify patients with medical expenses <$2000 in 1998 who were likely to incur high medical expenses ( $2000) in This study was the first phase of a population risk management project in which these targeted members were randomly assigned to an intervention consisting of a nurse-based, outbound-telephone survey with 3 purposes: To identify gaps in care. To further stratify the population to identify false positives whose diseases are well controlled. To help members become more compliant with the prescribed treatment regimen. Targeted members were assigned randomly to intervention and control groups to evaluate the interventions. The effectiveness of the interventions is a subject for future articles. Table 1, which was constructed from data before the introduction of a targeted intervention program, shows that a significant number of high-cost patients in 1998 became low-cost patients in Identification of high-risk members for interventions is just one aspect of population risk management. Equally important is the identification of high-cost 1998 members whose medical costs will decline because they represent a group for whom the application of resources should be limited. Modeling for population risk management should take into account a patient s characteristics in addition to his or her base year expenses. On a larger scale, claims-based prediction models should also take into account plan composition. Significant differences among MCO populations occur because of plan-specific factors such as plan type (independent practice association vs staff model, etc), capitation 384 THE AMERICAN JOURNAL OF MANAGED CARE MAY 2003

5 Targeted Care for Low-Cost Members agreements, copayment/deductible agreements, Medicare/commercial mix, and level of pharmaceutical caps. Regional differences in medical expenses may also reflect cultural differences, variation in clinical practices, and availability of medical resources. If an MCO has a large number of enrollees in a single region, we found that it is preferable to develop a prediction model for a single MCO rather than to pool and then try to make statistical adjustments to claims data from multiple heterogeneous health plans. The brunt of the effort is in the process of preparing data, not in making statistical calculations. DEVELOPMENT OF THE PREDICTION MODEL *International Classification of Diseases, Ninth Revision, Clinical Modification, code values of the comorbid conditions or diagnostic categories are available from the authors. The clinical logic that determines the presence or absence of a specific condition [based on pharmacy utilization patterns, demographic characteristics, laboratory test results, and physician visit patterns] is proprietary. Figure 1. Relationship Between the Number of Comorbid Conditions and the Probability of Annual Costs Being $2000 Medical Expenses $2000, % Comorbidities, No. The model was developed for a large HMO with an average enrollment of approximately persons. The patients studied had at least 1 claim in 1998 and 1999 and costs of <$2000 in The base year was 1998, and 1999 was the subsequent period in which financial outcomes were measured from medical and pharmacy claims. The standard split-sample technique (model developed on half of the 1998 patients and tested and validated on the other half) was used to prevent overfitting. The prediction model was developed using multiple regression model analysis. Several regression models were calculated using dependent variables such as the $2000/<$2000 binary variable, various transformations for the 1999 cost, and a proprietary cost grouper that measures the degree of cost concentration in a given period. The final independent patient variables included patient age, sex, number of specific comorbid conditions, number of distinct drug classes, number of physician visits, and nonphysician/nonhospital medical utilization. Binary (0=absent and 1=present) variables were created that flagged diabetes, cardiovascular, respiratory, and psychiatric diseases.* Other independent variables that reflect behavioral factors (eg, the primary treatment regimen for each disease state, the patient s prescription compliance pattern, and the patient s propensity to keep regular appointments with physicians) were created through a transformation of each patient s medical and pharmacy claims in the base year and were available for inclusion in the model. In the Appendix, we show an example of the specific coefficients and variable for one model used to develop predictions. The relationships between input variables and outcomes (eg, costs <$2000 annually/costs $2000 annually) may not be linear. However, input variables can often be transformed so that the resulting relationship between the aggregates of transformed values of the independent variable and the dependent variable is linear or nearly linear. For example the relationship between the number of comorbid conditions and the probability of annual costs being $2000 is nearly linear (Figure 1). RANKING ALL ENROLLEES, BASED ON CLAIMS AND THE PREDICTION MODEL By applying the regression coefficients to low-cost patients individual characteristics, a numerical score was calculated for every patient. The score directly corresponds to the probability that the patient will incur medical expenses $2000 in the subsequent year. For the purpose of designing an intervention program, the patients who are targeted for earliest deployment to health management and nurse intervention are those with the highest scores (ie, the patients with the highest probability of incurring costs $2000). The probability of identified members experiencing costs $2000 declines as the number of identified members increases. This inverse relationship, or yield curve, suggests that a population risk management program should focus on high-risk patients (Figure 2). VOL. 9, NO. 5 THE AMERICAN JOURNAL OF MANAGED CARE 385

6 VALIDATING THE TARGETING MODEL Retrospective Validation: 1999 Claims Expenses The prediction model was applied to approximately members who were enrolled in 1998 and had claims of <$2000. The members were ranked from high to low according to the probability of incurring medical expenses $2000 in Results are given in Table 2 and illustrate more clearly the relationship exhibited in Figure 2. The inverse relationship between these 2 variables indicates that the higher the percentage of lowcost 1998 enrollees targeted, the lower the proportion of those targeted patients incurring medical expenses $2000 in For example, the highest ranked 1054 low-cost 1998 patients, who were 0.5% of the 1998 low-cost enrolled population, had a 51.0% probability of incurring high costs, approximately 3.6 times the average of the entire low-cost population. Table 2 also gives (by rank) the average claims costs incurred by targeted members who had claims $2000. Prospective Validation: 2000 Claims Expenses For the prospective test, the prediction model was applied to low-cost 1999 patients. Members in 1999 and 2000 were ranked from high to low according to the estimated probability of incurring costs $2000 in Because this prediction was done as part of an intervention program, the highest-ranking highrisk patients were selected to receive nurse interventions. This group of 5535 members had a risk ranking similar to that of the group we reported previously herein (risk rank 34). Eighty percent of these targeted patients (n=4428) were randomly selected to receive an intervention and, thus, were inappropriate for prospective validation because Figure 2. Yield Curve Showing that the Probability of Identified Members Experiencing Costs $2000 Declines as the Number of Identified Members Increases Targeted Patients with Cost $2000 in 1999, % Patients Targeted, % their behavior was subject to change by the intervention. The remaining 20% of the high-risk patients (n=1107) composed the control group and received no intervention. The subpopulation identified by the prediction model as high risk, on average, was older, was more likely to be male, and had more comorbid conditions than the nontargeted population (Table 3). The prevalences of asthma, diabetes, and congestive heart failure for the high-risk population were also significantly higher than those for the low-risk population. Table 4 displays results for the 1107 members of the control group and of the remaining, untargeted members who experienced claims <$2000 in 1999 but who were not targeted for intervention because their risk scores were low (risk rank 33). Pharmaceutical and medical claims incurred in 2000 were used to calculate the relative frequency of patients with paid claims $2000 and to compare the average cost of the control high-risk patients with that of the low-risk patients. Table 4 provides the claims expenses for 2000 for the targeted and nontargeted patients. Of the targeted, highest-ranked patients, 39.7% incurred high ( $2000) claims expenses in 2000 compared with 12.2% of the nontargeted patients. DISCUSSION Although the prediction model was constructed using data from one large HMO instead of pooled data from other MCOs, the model can be generalized and modified to fit other populations. The key variables used in the regression model (patient age, sex, number of chronic conditions, number of distinct drug classes, number of physician visits, nonphysician/nonhospital medical utilization, and pre-sence or absence of diabetes, cardiovascular, respiratory, and psychiatric diseases) have been tested on several other large HMO data sets and different periods. The coefficients differ somewhat because of differences bet-ween plan-specific factors such as plan type, physician in-centives, copayment/deductible, and pharmacy benefit levels. However, the variables used in the prediction model in the member study group are identical to those found in HMOs with different patient and financial characteristics. If 2 years data are available for a large MCO, our preference is to construct MCO-specific prediction models rather than to make adjustments for the differences in plan characteristics. 386 THE AMERICAN JOURNAL OF MANAGED CARE MAY 2003

7 Targeted Care for Low-Cost Members Table 2. Patients With Claims in Both 1998 and 1999 and 1998 Costs <$2000 Cumulative Average 1999 Cumulative Probability Claims Cumulative Cost per Rank Patients, No. $2000 in 1999 Patients, % Costs in 1999, $ Patient, $ Our analyses shed new light on regression to the mean. We found that very few patients are consistently high-cost members (data not shown). Of those members who incurred catastrophic costs in 1999 ( $25 000), 39% were in the 1998 low-cost category and 43% came from the previous year s mediumcost segment. Only 18% of the 1998 high-cost category members were repeat high-cost consumers in In 2000, a very small percentage of the high-cost patients accounted for ~20% of total expenditures; our data thus suggest that expenditure levels in the base year are not a good predictor of high costs in the subsequent year. The transient nature of patients in MCOs is well known in the industry, with turnover rates of 20% to 25% per year. Our prediction models required 2 years data for model construction and validation. Disenrollment of > patients occurred in Additional work is perhaps needed to study the characteristics and utilization patterns of patients who enroll and disenroll. Patients who are identified as high-risk patients in the first year but then disenroll before an intervention can be undertaken and measured will continue to complicate the evaluation of population risk management programs. In our analytical approach, it was not practical to make any adjustments to reflect possible increases in provider reimbursement rates, which were modest in We doubt that price adjustments, which would require considerable effort, would change the basic results of our research. One alternative worth consideration for future modeling is to study expenditure patterns in the base and subsequent years by quintile or quartile. The small price increases may have caused a few patients to shift from low cost (<$2000) to high cost Table 3. Characteristics of High-Risk Patients With 1999 Costs <$2000 Disease Prevalence, % Sex, % Average Average Comorbid Congestive Male Female Age, y Diseases, No. Diabetes Heart Failure High risk, targeted, no intervention (n = 1107) All members VOL. 9, NO. 5 THE AMERICAN JOURNAL OF MANAGED CARE 387

8 Table 4. Year 2000 Outcomes of Patients With 1999 Cost <$ All Members Members With Claims $2000 High risk, targeted with no intervention (n = 1107) Average cost, % Costs >$2000, % Low risk, not targeted (n = ) Average cost, $ Costs >$2000, % ( $2000). Using different cutoff values ($2500, $3000, etc) did not materially affect our results. CONCLUSIONS The prediction model is based on historical claims data and was used to score each member s risk for incurring high medical expenses in the second year. The prediction model successfully identified patients with low medical expenses in 1998 who were 3.6 times as likely to incur high medical expenses in 1999 as the overall low-cost population. The prediction model was tested prospectively on 1107 patients who received no intervention and were identified as likely to incur high medical expenses in The 1107 patients were more likely to incur costs $2000 (39.7% vs 12.2% for the nontargeted group). Their average costs were $6602 vs $1108. The prediction model is only the first step in developing cost-effective intervention programs. Much hard work remains: Creating, evaluating, and implementing new interventions; Adopting objective criteria to evaluate interventions, which may eventually involve measuring clinical outcomes and cost savings; Testing and adopting enhancements to the model, including increasing the horizon beyond a single year to identify members at risk for longer-term events; Forecasting the financial impact of interventions by including cost estimates of interventions and estimated impact on medical expenses; Collecting data prospectively to assess the costeffectiveness of new and existing interventions; Providing tools for making better resource allocation, staffing, and intervention decisions; Finding ways to better identify and engage members who are not compliant but whose behavior may be changed by an intervention; and Developing tools to incorporate predictions into pricing and underwriting strategies. Population risk management depends on the development of accurate prediction models in which patients are selected for intervention according to their predicted risk. The second component of population health management is to devise new interventions, that is, programs that change healthcare delivery and hopefully improve patient outcomes. A third and final step, in the absence of randomization, is to use the prediction model to adjust patients outcomes so that actual-to-expected outcomes can be compared. REFERENCES 1. Cumming RB, Knutson D, Cameron BA, Derrick BA. Comparative Analysis of Claims-based Methods of Health Risk Assessment for Commercial Populations. Schaumburg, Ill: Society of Actuaries; Meenan RT, O'Keeffe-Rosetti C, Hornbrook MC, et al. The sensitivity and specificity of forecasting high-cost users of medical care. Med Care. 1999;37: LoBianco MS, Mills ME, Moore HW. A model for case management of high cost Medicaid users. Nurs Econ. 1996;14: , Forman SA. Breakthroughs in High Risk Population Health Management. San Francisco, Calif: Jossey-Bass Publishers; Forman SA, Kelliher M, Wood G. Clinical improvement with bottom-line impact: custom care planning for patients with acute and chronic illnesses in a managed care setting. Am J Manag Care. 1997;3: Lynch JP, Forman SA, Graff S, Gunby MC. High-risk population health management: achieving improved patient outcomes and near-term financial results. Am J Manag Care. 2000;6: Ellis RP, Pope GC, Iezzoni L, et al. Diagnosis-based risk adjustment for Medicare capitation payments. Health Care Financ Rev. 1996;17: Kronick R, Dreyfus T, Lee L, Zhou Z. Diagnostic risk adjustment for Medicaid: the disability payment system. Health Care Financ Rev. 1996;17: Pope GC, Ellis RP, Ash AS, et al. Principal inpatient diagnostic cost group model for Medicare risk adjustment. Health Care Financ Rev. 2000;21: Weiner JP, Tucker AM, Collins AM, et al. The development of a risk-adjusted capitation payment system: the Maryland Medicaid model. J Ambulatory Care Manage. 1998;21: THE AMERICAN JOURNAL OF MANAGED CARE MAY 2003

9 Targeted Care for Low-Cost Members Appendix 1. Specifics of the Prediction Model Independent Variable Coefficient Diabetes (drug and diagnosis based) Cardiac diagnosis (drug and diagnosis based) Respiratory disease (drug and diagnosis based) Psychiatric diagnosis (drug and diagnosis based) Physician visit variable Nonhospital, non emergency department, nonphysician medical claim variable Composite prescription drug variable: measure of prescription drug classes Comorbidities (truncated at 4) All variables are transformed to be used in the model. Because this is a relative risk stratification, rather than an absolute determination of cost, there is no need for an intercept variable. 2. Evaluation of models was through receiver operating characteristic (ROC) curves. The ROC curve shown below for continuous enrollment has an area of Sensitivity, % Specificity, % 3. Sample Demographic Statistics Eligibility, 1998 Cost <$ Age Female Male All Gender Unknown < All Ages Disease Classes: 1998 Cost <$2000 Congestive Heart Cardiac Respiratory Claimants Failure Condition Asthma Diabetes Condition % 0.2% 9.8% 9.2% 3.4% 22.5% 4. Cost Distribution (1998) All 1998 Cumulative Cost Group Cumulative % % $ Claimants Claimants Claimants Claimants Sex unknown 1647 All VOL. 9, NO. 5 THE AMERICAN JOURNAL OF MANAGED CARE 389