Version 6.0 Release Notes

Transcription

1 The Johns Hopkins ACG Case-Mix System Version 6.0 Release Notes PC (DOS/WIN/NT) and Unix Version 6.0 April, 2003 (Revised June 4, 2003)

2

3 This document was produced by the Health Services Research & Development Center at The Johns Hopkins University Bloomberg School of Public Health The Johns Hopkins University. All Rights Reserved Documentation Production Staff Editor in Chief: Jonathan P. Weiner, Dr. P.H. Senior Editors: Chad Abrams, M.A., David Bodycombe, ScD Major contributors: Christopher B. Forrest, MD, PhD, Thomas M. Richards, MS Barbara Starfield, MD,MPH, Hoon Byun, M.A. Editorial Assistance: Klaus Lemke PhD, and Tracy Lieberman. If users have questions regarding the software and its application, they are advised to contact the organization from which they obtained the ACG software. Comments, criticisms, or corrections related to this document should be directed to the Johns Hopkins ACG team (see below). Such communication is encouraged. ACG Project Coordinator 624 N. Broadway - Room 600 Baltimore, MD USA Telephone (410) Fax: (410) askacg@jhsph.edu Website: i

4 Important Warranty Limitation and Copyright Notices Considerable attempts have been made to ensure that the materials included in this document are accurate and appropriate to users needs. However, the responsibility for the appropriate application of the Johns Hopkins University ACG Case-Mix System, its supporting software and this documentation rests with the end-user and not the Johns Hopkins University or its agents. No warranty is given or implied that any of the information, methods or approaches discussed in this document are error-free. All terms and conditions associated with the software license, including the Johns Hopkins University and its agents not bearing any liability for actions taken by the user on the basis of software output, are in place and should be understood by the user. THE JOHNS HOPKINS UNIVERSITY HEREBY DISCLAIMS ALL WARRANTIES, INCLUDING THE WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. Licensed users of the ACG software may copy and distribute this documentation within their organization. To facilitate this process, electronic files (in.pdf format) are available for download at Individuals not within a licensee organization may download one copy for personal use only, and under no circumstances may they reproduce or distribute this documentation on paper or in any electronic format, beyond this single personal use copy. All copies so downloaded and /or distributed shall contain this notice and the identification of the source of the software shall not be deleted. The terms The Johns Hopkins ACG Case-Mix System, ACG System, ACG, ADG, Adjusted Clinical Groups, Ambulatory Care Groups TM, Ambulatory Diagnostic Groups TM, Johns Hopkins Expanded Diagnosis Clusters TM, EDCs TM, Dino-Clusters TM, ACG Predictive Model, and acgpm, are trademarks of The Johns Hopkins University. All materials in this document are copyrighted by the Johns Hopkins University. It is an infringement of copyright law to develop any derivative product based on the grouping algorithm or other information presented in this document. Copyright 2003, The Johns Hopkins University. All rights reserved. ii

5 iii

6 Preface Welcome to ACG Version 6.0 We hope that you will share our excitement about this newest release of the Johns Hopkins ACG System. No previous ACG software update has been more significant in terms of additional features over the prior version. No other version has had as many practical benefits for its users. Health care risk adjustment and predictive modeling in general are not simple processes, but the goal of these release notes is to help ease your way into these domains and to get you started as quickly as possible in making full use of the array of features in ACG Release Version 6.0. How to Use This Document This document, the ACG Version 6.0 Release Notes, augments the last comprehensive Johns Hopkins ACG handbook, the Release 5.0 Documentation and Application Manual. The chapters provide existing ACG users with all of the information needed to begin to use the new features and capabilities of this latest release. We recommend that first-time users of ACGs also review the Release 5.0 Manual to learn about the core features of the ACG System, including the system s conceptual basis. Both sets of documents are distributed with the Version 6.0 software, but if you need a copy of the Release 5.0 documentation, it is available on our website at The release notes are organized into six sections as follows: 1. An overview of new features in ACG Release 6.0. This will quickly introduce you to both the new areas of functionality and the modifications and updates of previous features. 2. A comprehensive description of the new state-of-the-art predictive model, the acgpm, which includes performance data and application suggestions. 3. Details of the additions to and refinements of the expanded diagnosis cluster (EDC) disease-marker methodology, including several new reports that combine EDCs and ACGs. 4. Introduction of the new internal relative value weights for ACGs that are now included in the software. 1

7 5. Suggested guidelines on how to make the best use of the tools and methodologies now found in the ACG System. This section will help you quickly get to the bottom line, and thus supports mission-critical clinical and financial management decisions. 6. A detailed technical software installation guide that describes all updates and includes operating instructions like those found in the previous ACG version. You will need only this section (and no previous documentation) to run the software. We hope and expect that you will find the latest ACG software and this accompanying documentation of value to you and your organization. As always, the Johns Hopkins ACG development team welcomes your feedback. Please forward comments, questions, or suggestions for improvement about this release document or our software to askacg@jhsph.edu. 2

8 Section 1 Overview of New Features in ACG Version 6.0 The Johns Hopkins University ACG development team is very pleased to distribute version 6.0 of our ACG risk-adjustment/case-mix software. This software includes several enhancements that significantly expand both the breadth and depth of the Johns Hopkins ACG suite of diagnosis-based health care measurement tools. This version reflects our ongoing commitment to continued improvement and refinement of the ACG System. Now more than ever, the ACG toolkit can serve as the measurement engine for a full array of health care applications, including clinical care management, resource management, health services finance and payment, and applied evaluation and scientific research. Some of the new features in version 6.0 will advance and expand the clinical grouper and measurement tools available to the experienced user. Other improvements will make the ACG System easier to use for those just starting out. Many changes and enhancements will be of value to both groups. As noted in the Preface, the key enhancements included in Version 6.0 include: predictive modeling (including new reports), additions and refinements to the expanded diagnosis cluster (EDC) methodology (including new reports), updating ICD-9 codes, relative risk factors based on nationally representative data, and an updated installation guide. a. AcgPM: Predictive Modeling for High-risk Case Identification and Future Costs The ACG Predictive Model (acgpm) permits the rapid identification of high-risk patients who may benefit from care management services. The acgpm is the focal point of several notable innovations: focus on individualized care management, incorporation of a new hospitalization propensity index, ability to employ pharmacy use data where available, integration of elements of our EDC technology, and incorporation of a unique regression modeling strategy. 3

9 The acgpm remains grounded in the disease burden perspective unique to the ACG System, one that focuses on commonly occurring patterns of morbidity and assessment of all types of medical need. The model also shares the modest data requirements of the ACG System. This state-of-the-art predictive modeling risk identification module builds on the many facets of the ACG system and several years of intensive research and development at Johns Hopkins. According to all empirical assessments to date, the accuracy of the risk prediction measures calculated by the acgpm equals or surpasses other available methods. One of the outputs of the acgpm model is based on a sophisticated logistic regression equation that maximizes the tool s ability to identify your members who will be among the very highest cost users in the next year (or some other future period). With the same data that you use to assign standard ACGs, the acgpm module assigns each individual within your organization a risk Probability Score that can be employed to array the members of your population from the lowest to the highest risk. As described in the following section, evaluations of the acgpm Probability Score indicate that it is more accurate than case identification derived from prior use/prior (cost) experience. Unlike other case identification methodologies based on extended hospital stays or repeat specialist visits, the acgpm method helps to identify many persons before they actually become expensive. Another set of outputs is based on linear OLS (ordinary least squares) regression and can be used to provide an estimate of cost for each individual in the subsequent time period. Termed the Predicted Resource Index (PRI), this index can be used to assist clinical administrators, actuaries, and others interested in valuing future expectations of resource use. This index can be used for a wide range of actuarial/financial applications such as setting risk-adjusted capitation payments or setting premiums. In addition to offering measures of risk related to overall service use, the acgpm includes a special component for predicting an individual s risk of using pharmaceuticals. Moreover, the acgpm can incorporate prior pharmacy use data, when obtainable, to improve the accuracy of its predictions. If prior pharmacy use data are available within an organization, this combination of pharmacy and the typical ACG risk factor information derived from ICD diagnosis information is a powerful means for identifying individuals at risk of high future resource use. 4

10 As discussed in more detail in the main body of this document, automatically generated acgpm reports combine the Probability Score and the PRI. This enables users to focus attention quickly on those individuals with case-manageable conditions at greatest risk for future high expenditures and to provide estimates of what these individuals might cost if no intervention is taken. b. Additions and Refinements to the Johns Hopkins Expanded Diagnosis Clusters The expanded diagnosis cluster (EDC) disease marker system originally emphasized commonly occurring conditions treated primarily in ambulatory settings. For this version, the EDC mapping underwent considerable refinement. To provide a more comprehensive categorization approach for profiling morbidity, EDCs now include less commonly occurring conditions, many of which may require hospitalization. All in all, 40 new EDC categories have been added and affect 16 of the 27 major EDCs (MEDCs), and several existing EDCs have been subdivided to provide additional clinical specificity. To facilitate quick implementation of the EDC typology, the software print file now incorporates a series of EDC-based standardized morbidity ratios (prevalence of EDCs within your population, after controlling for age and sex) as well as a series of combined ACG/EDC tables to help support case-management and disease-management programs. c. Updating ICD-9 Codes The ICD-9 mapping tables for both the EDC and ACG/ADG groupers have been updated to include all new codes introduced in the Center for Medicare and Medicaid Services official ICD-9-CM (Version 20). As in years past, old or retired ICD-9 codes have been retained as part of our algorithm because they were once valid and/or their interpretation is reasonably clear. d. Available Relative Value Weights for ACGs and RUBs A key strength of the ACG suite of case-mix/risk adjustment measures is that the grouping algorithms and analytic approaches in the software are very accessible to most organizations. The straightforward ACG actuarial cell approach readily allows for the application 5

11 of an organization s own information to calibrate weights (or coefficients) to fine-tune the analysis to best match the local situation and context. While many users of ACGs have applied actuarial cell-based models, there is a growing body of users who wish to implement risk adjustment and for whom no local cost data are available To facilitate quick implementation of ACG technologies, Version 6.0 for the first time incorporates concurrent (i.e., same-year resource use) ACG weights based on a large research database (more than 2 million persons) younger than 65 years of age who were enrolled in several U.S. commercial health insurance plans. Version6.0 also makes available individual assignment to one of six ACG categories (from low to high) termed ACG Resource Utilization Bands, or RUBs. Relative value weights for each RUB category can be calculated by using local cost data or by using the concurrent ACG weights built into the software. The supplied RUB categories cover six morbidity levels: non-users, healthy users, low morbidity, moderate morbidity, high morbidity, and very high morbidity. Figure 1 illustrates the percentage of the population and percentage of total dollars associated with each RUB (based on a large research population); the bottom 40% of the population consume less than 3% of all health care dollars while the top 0.7% in the highest RUB category consume over 15%. RUBs can be used in the same fashion as ACG assignments. For many applications, they dramatically simplify risk adjustment computational tasks while retaining considerable ability to explain variations in resource use that are attributable to casemix. 6

12 Figure 1: Distribution by ACG Comorbidity RUB Distribution of ACG Comorbidity RUB % of P opulation % of Dollars 50.0% 45.1% 40.0% 30.0% 25.8% 28.3% 27.6% 24.4% 20.0% 13.9% 12.5% 15.2% 10.0% 0.0% 2.4% 3.7% 0.4% 0.7% Nonusers Healthy Users Low Moderate High Very High ACG Comorbidity RUB Level The relative values and RUB assignments available within the current version of our software are also used to generate tables that combine ACG and EDC information to develop a series of summary reports. These reports are discussed in Section 5 of the Release Notes, Selecting the Right Tool from the Expanding ACG Tool Box. The software-provided weights may be considered viable external or reference weights for concurrent ACG or RUB analyses. These weights can be used as substitutes for locally calibrated weights by those organizations with no available resource-use measures, or whenever the population may be too small to produce reliable local weights. (In addition to the section on this topic in the Version 6.0 Release Notes, please see Chapters 6, 8, and 11 of the Version 5.0 Documentation and Application Manual for further discussion of the advantages and disadvantages of local versus software-provided reference weights). e. Technical Notes Although the functionality of Version 6.0 is greater, there are only modest changes from Version 5.0 in terms of how to load and execute the software. Those familiar with the operation of ACG Version 5.0 will find it easy to implement 6.0. All new features follow from the 7

13 incorporation of a few new reserved words in the ACG control card file (used to pass data to the software and to control what fields are written to the software s output file). Readers should review the Installation and Usage section to familiarize themselves with the technical details and new specification requirements for 6.0. The guide included in this document can stand alone, and there should be no need for programmers or analysts running the software to refer to the previous Version 5.0 Manual. 8

14 Section 2 The ACG Predictive Model: Helping to Manage Care for Persons at Risk for High Future Cost This section introduces the ACG Predictive Model (acgpm), an advanced tool for projecting future resource use based upon concurrent data captured largely from standard claims files. In addition to background on predictive modeling and on the construction of the acgpm, the text also provides model performance specifications, and concludes with a discussion of output reports and how they may be used to support case and disease management. a. Introduction This chapter describes the newest addition to the ACG System s toolkit, the ACG Predictive Model (acgpm). The acgpm uses sophisticated statistical techniques to project the impact of co-morbidity and other factors on an individual s use of health care resources in a subsequent time period. The acgpm is designed for prospective high-risk case identification and will be of real value for assessing both the quality and appropriateness of patient care. We begin with an overview of the acgpm and then launch into a discussion of our acgpm development effort, the elements of the new models, and model performance assessment. We conclude the chapter by addressing application issues. i. Model Offers Fast Identification of High-risk Patients The acgpm permits the rapid identification of high-risk patients who may benefit from care management services. The acgpm remains grounded in the disease burden perspective unique to the ACG System, which focuses on commonly occurring patterns of morbidity and assessment of all types of medical need. This holistic method has repeatedly proved to have many advantages over comparable case-mix adjustment approaches that include only a limited set of disease or episode categories. Also, our predictive model s straightforward approach to integrating clinically relevant risk factors offers advantages over black box strategies based on complex clinical algorithms for data-mining or on artificial intelligence. The Venn diagram provided in Figure 1 graphically depicts the predictive modeling challenge. Traditionally plans have made use of data on prior experiences (the Actual High-risk 9

15 Year 1 ellipse) to project individuals likely to remain costly in the following year. This projection represents an amalgam of patients that includes those who have continuing chronic conditions and also those who go through acute events or injuries that do not recur. The acgpm identifies another group of individuals (the Predicted High Risk ellipse) that overlaps the priorcost ellipse but also identifies individuals who were NOT high cost in Year 1 (the shaded area). Thus an important attribute of predictive model performance is the size of the shaded area. Figure 1. Identifying High-cost Persons Who Were Not Previously High Cost ii. Model Outputs Serve Dual Purposes The acgpm software produces two types of predictive risk indicators: (1) a probability score representing the likelihood that a member will be among those persons using extraordinary health care resources in the coming year and (2) a predicted resource index that expresses anticipated resource use as a relative value. Probability scores are used because the clinical decision-making process is often couched in terms of likelihood or odds. When employing the probability score, users can set their own definition of high resource use by setting some minimum threshold. Thus, for the purposes of case management, a health plan might choose to consider only those individuals with a probability of.6 or more of falling within the high-risk group during the following year. Our experience suggests this cut-off would identify about one half of 1% of the plan s members. There are performance tradeoffs to be made (e.g., increased positive predicted values with lower sensitivities) when setting minimum thresholds. Setting higher probability thresholds permits 10

16 prediction with greater accuracy but with a greater chance of missing potentially high-risk cases. These tradeoffs are discussed later in this chapter. The acgpm probability score has been tailored to case identification and thus will be especially useful to case managers in targeting patients for intervention. The second acgpm output, the predicted resource index, can be applied to calculate expected resource dollars. Case managers will be able to calculate expected differences between current and projected future costs to prioritize interventions that could have the highest impact. Health plans and others will also find this model output to be a useful tool for rate setting and financially related decisions. iii. Models Forecast Overall and Pharmacy-specific Expenses The acgpm output provides two separate sets of probability scores and predictive resource indexes: the first for total cost, and the second for pharmacy-specific cost. The total cost output represents an overall measure of inpatient and outpatient resource use that is the main focus of the prediction model. The model also provides pharmacy resource predictors because pharmacy-specific cost has become a major component of overall expenditures and is currently the focus of many payment and delivery organizations. Although there is a correlation between medical care service use and pharmacy use, the relationship is not one-to-one. Disease management evaluations are showing that total cost savings can be achieved, primarily because of reduced inpatient care, while pharmaceutical use increases. iv. Identification by Prior Costs Alone Omits Many Cases Prior cost and other utilization experience (e.g., extended hospitalization) are often used as the basis for identifying individuals for inclusion in intensive case management. Chronically ill individuals who have had significant health care use in the past often do continue these patterns of high care use into the future. Prior use measures also identify patients with a highcost acute event, which may have no bearing on healthcare use in the future. Still others who had previously been less intensive users of health care resources may enter a high-use phase. For especially high-risk cases, reliance on prior use alone will yield a very incomplete picture. In 11

17 Figure 2 we have pooled together those persons within a large health plan data set whom were successfully predicted to be high use (cost) by either the acgpm or by prior cost experience. 1 Figure 2. Percentage of Cases Correctly Predicted to Be at High-risk for Using Extraordinary Health Care Resources, by Prediction Strategy acgpm 46% Both 27% Prior Cost 27% acgpm Prior Cost Both Overall, the acgpm identified 73% of the high-risk cases that were successfully captured by either approach. The acgpm uniquely accounted for nearly half of the successfully identified high-risk patients. These cases were not previously high resource users and would be missed if prior cost alone were used in case identification. v. The acgpm Adds Another Tool to the ACG System Without Increased Data Collection The acgpm substantially improves high-use case identification and does so without increasing the minimal data collection effort currently required for the ACG software. The acgpm constructs its risk factor information from claims data streams containing only age, sex, and diagnostic information. If a plan has historical pharmacy-cost information for each person (the full drug claims history is not needed), this summary measure of pharmacy use can optionally be added to the input data stream for enhanced model performance. Moreover, the 1 The high risk group in this comparison included those with an acgpm probability score of 0.6 or more of being among those persons with the top 5% of total costs next year. An equivalent number of the highest prior cost cases in year 1 were then identified for comparison. These two groups each account for about ½ of 1% of the test population. 12

18 acgpm is bundled within the current ACG System, and thus the full capabilities of ACGs, including provider profiling and risk-adjusted capitation determination, are available as part of the standard software package. b. Developing the acgpm i. What is a Predictive Model? Predictive modeling in the health care management context is generally defined as a process that applies existing patient data to identify prospectively persons with high medical need who are at risk for above-average future medical service utilization. Such future resource use is often, although not always, linked to negative health outcomes. A predictive model can incorporate information from a wide array of sources and can rely on many statistical approaches from the very simple to the very complex. As noted, one simple approach is to use contemporary cost data as a predictor of future high cost. For example, a plan could assume that if a person is very high cost this year, he or she is likely also to have a high care-utilization experience in the future. Other models use multivariable statistical analysis and data available from a range of sources for their predictions. The basis for many of these efforts is ordinary least squares regression. Some predictive models are based upon neural networks. These are sometimes termed artificial intelligence approaches, given that neural networks represent collections of mathematical models that emulate the processes observed in biological nervous systems, including the capability to adapt and learn. From the perspective of either empirical accuracy or performance, no modeling method has yet to demonstrate clear superiority. One of the main challenges facing any modeling strategy is the limited risk factor data on which the models can be based. Currently, the commercially available prediction tools depend largely on standard medical care claims (age, sex, diagnoses, procedures, prescriptions, service dates, and cost). Model coefficients are developed using one year s data to predict a second year. With the limited source of risk factor data, model predictive capabilities are limited as well. All predictive models are tools that must be used with good clinical and managerial judgment and other sources of information in making decisions. 13

19 ii. Historical Use of ACG for PM Although a separate predictive model is new to the ACG System as of version 6.0, for years elements of the ACG case-mix grouping algorithms have been successfully adapted by users to apply their own customized predictive models. ACGs work very well for this purpose because, at their core, they focus on the dimensions that help predict high risk, such as: persistence of the conditions, seriousness/severity of the conditions, the co-morbid nature of disease, the likelihood of a negative outcome, and the need for high levels of medical services. Prior work by Starfield et al. (Primary Care, Co-morbidity, and Case Management, presented at the Conference on Health Care Risk Adjustment, Minneapolis, MN, May 2, 2001 and available in the virtual ACG library at: suggests that the components of the ACG System have inherent utility as predictors of high future costs. As is shown in Table 1, the co-occurrence of ADGs and the presence of selected ACGs are superior to prior hospitalizations in predicting high costs in subsequent years. Table 1. Potential Influence of Prior Co-Morbidity and Hospitalization on Future Cost. Measure Percentage of Members Percentage High Cost in Year 2 Percentage High Cost in Year 3 2+ Hospitalizations 0.7% 53.3% 51.5% 12+ ADGs 0.9% 65.6% 66.1% 4+ Major ADGs 0.4% 72.4% 70.1% Selected ACGs 1.7% 60.0% 52.0% Total <3.7% For other successful applications of the ACG System components to predict high-risk cases, see the proceedings of the 2002 International Johns Hopkins ACG Risk Adjustment Conference, Baltimore, MD, April 28, 2002 (available in the virtual ACG library at: 14

20 c. The Nuts & Bolts of the acgpm Model i. Specifying the Predictor Variables The acgpm modeling strategy makes use of the comprehensive array of morbidity metrics that are available within ACGs. The model incorporates the morbidity-based ACGs, selected disease-specific EDCs, and a newly developed diagnostic indicator of the likelihood that someone will be hospitalized in an ensuing year. We have also added an indicator of the level of prior pharmacy use. First, a brief word on the issue of prior-use measures. Prior use is a fallible indicator of future use because it includes many acute conditions that get resolved. The acgpm focuses on individuals with a high morbidity burden and with high-impact chronic conditions that are likely to continue to require significant health care resources. Prior-use measures are also not appropriate as risk factors for risk-adjusted rate setting or profiling as they potentially could provide incentives to overuse resources. That is, providers can readily increase the risk rating of their patient (and potentially reimbursement) simply by ordering more services. Including prior use is appropriate for high-risk case identification since the goal is to identify and potentially intervene among high-cost users. Several alternative prior-use measures were assessed for inclusion in the acgpm model, but levels of previous pharmacy expense proved to be the most powerful resource predictor with the clearest clinical implications and with a minimal addition to the burden of data collection. The acgpm risk factor variables used in the model are as follows: age (seven age groups from infants to 64 years of age), sex, ACGs (three broad morbidity groupings from low to high in addition to selected individual ACGs), a hospital dominant marker (reflecting diagnoses where hospital care was dominant, though care could be provided in a variety of settings), identification of pregnancy, where no delivery has yet occurred, pharmacy expense levels, and EDCs (a limited set that represent high impact and chronic conditions). Specific EDC disease markers were incorporated into the model if they represented common high-cost chronic conditions that were frequent targets for disease management programs, uncommon, but high impact on both cost and health, conditions, 15

21 conditions for which the evidence linking health care to outcomes is strong, complications that potentially signify instability in a chronic illness. (e.g., retinopathy), or conditions that are a major biologic influence on health status (e.g., transplant status, malignancy). The focus of the current acgpm software release is on the non-elderly (i.e., under 65 years of age) population. The specific ACGs and EDCs that were used in building acgpm are documented in Appendix 1. ii. Defining the Model Outputs As noted earlier, the acgpm model offers two types of outputs: a probability score of being a member of the high-risk 1 group next year and a predicted resource index reflecting expected cumulative resource use. The two indicators are intended for different purposes (case selection for the former and cost estimation for the latter) and benefit from somewhat different statistical methodologies. Probability scores range between zero and one. For example, conceptually, an individual with a probability score of.4 has a 40 in 100 chance of being in the high-risk cohort next year. The predicted resource index ranges from zero to roughly 40 with a population mean of 1.0. The index can be readily converted to a predicted dollar amount. These two outputs are repeated for both total costs and for pharmacy costs only. The Johns Hopkins ACG development team chose to use logistic regression (logit) to develop the probability score for a patient becoming a member of the high-risk group next year. This is an important departure from the prevailing strategies for high-risk case identification that usually employ ordinary least squares (OLS) regression based on linear modeling strategies. Logit models are best for predicting events (yes/no occurrences) in this case, being a high user of resources In terms of estimating resource use using multivariate methods, regression is the most effective strategy for estimating dollars. Therefore we used an OLS model for the resource-use prediction component of our model. The acgpm s predicted resource index is presented as a 2 Based upon repeated testing of alternative thresholds, for model development we have defined high risk to represent individuals whose predicted costs for the following year are expected to fall within the top 5% of a plan s members. This choice of high risk definition does not preclude users from adopting other definitions, e.g., the top 1%, ½ or 1% or even top 10% of plan members. 16

22 relative value that can be adjusted to reflect local mean costs. This adjusted scale can be readily converted into dollar amounts comparable to future cost estimates by a simple algebraic process. The process for deriving dollar values is discussed later in this chapter. d. A Framework for Discussing the Performance of Predictive Models There are many different ways to apply the acgpm model, and each organization will have unique data and contextual issues. Furthermore, there are few standard approaches in the literature for evaluating the accuracy of any predictive model for resource use. Thus assessing and reporting the performance of predictive models is not straightforward. We are providing users with several alternative ways to assess performance because ACG users tend to be discerning with regard to appropriate methods and statistical techniques. What follows is a summary of an evaluation that assessed the performance of our models in helping to identify persons who are members of high-risk/high-cost cohorts in the subsequent year. We tested the models using actual data from a very large data set, consisting of over two million lives enrolled in several health plans. We adopted a split half approach, using a random selection of half the observations to build the models and the other half of the observations to validate the model. The performance figures reported in the following section are based on the validation half of the data. Unless otherwise specified, we apply the version of our model that includes pharmacy cost as one of the risk factor inputs. One way to assess predictive models is to ascertain how well they classify cases as actually being a member of a high-risk group in a future period. This yes/no accuracy assessment is similar in many ways to the statistical/epidemiologic approaches that are used to assess the accuracy of diagnostic screening tests or exams. However, this quantitative approach does not tell the full story regarding model performance. Other important questions are, What are the prediction characteristics of the model? and Can it aid in the identification of cases where intervention is possible and where care can be improved? There is no explicit test to determine this result, but our performance assessment attempts to consider this capability as well. Epidemiologists often use sensitivity and specificity to assess the validity of screening tests. These performance indicators are defined as follows: Sensitivity is the percentage of true high-risk cases that are successfully identified: 17

23 Sensitivity = true positives/(true positives + false negatives). Specificity is the percentage of true low-risk cases that are successfully identified: Specificity = true negatives/(true negatives + false positives). Specificity is not especially useful in assessing this or any other predictive model since the focus is on only a very small subset of high-risk persons and a large number of true negatives. Positive predictive value (PPV) represents a potentially more useful alternative to sensitivity. PPV is defined as the probability that someone predicted by the model to have high expected Year 2 resource use does, in fact, have high Year 2 resource use. Mathematically, this is expressed as: PPV = true positives/(true positives + false positives). A PPV score gives information about the likelihood that a person who tests positive (in this case is predicted to be high-risk) actually will be a high-resource user in Year 2. Finally, another widely used measure of the ability of such models to correctly classify patients is the c-statistic. The c-statistic provides an overall measure of model performance and represents the probability that an observation is correctly classified as a true positive or true negative along a continuum of test thresholds (in this case probability score thresholds). The closer the c-statistic is to 1.0, the better the model. For a summary of these and other performance indicators that are often applied to diagnostic tests and predictive models, see Appendix 2. e. Performance of the acgpm Model We sought to assess our model s performance by asking several key questions: How well does acgpm improve upon prior cost alone in identifying high-risk cases? Do the cases identified by the model represent a group meriting intervention? How well does acgpm do in estimating future costs of care? What added predictive value is gained by including the optional pharmacy cost predictor? 18

24 i. How Well Does the acgpm Improve on Prior Cost Alone in Identifying High-risk Cases? The statistical properties of the acgpm total cost model are compared to prior cost only predictions in Table 2. The performance of the acgpm is shown with respect to a series of probability thresholds based on the prediction scores output by the model for all members of a large health plan test population. Table 2 shows the acgpm s accuracy at six different probability score cut-offs within the large health plan test population of about 410,000 persons (under the age of 65 years). By setting the risk threshold low, e.g.,.4 or higher, a higher percentage of cases is included (in the example, about 1.33% of the population). Set the risk threshold higher and select a lower percentage of cases (only 0.10% at a minimum threshold of.9 or higher). Thus certainty, expressed as the PPV, comes at a price. By setting a high threshold, you come close to certainty that every case you select will become high cost in Year 2. Set a lower threshold and there is less certainty but a higher likelihood that you re picking up all the potentially high-cost cases. Even with probability floors as low as.4, the likelihood is still better than chance (via PPV) that an acgpm-predicted high-risk patient will actually turn out to be a high consumer of resources. For comparison purposes, in Table 2 we also show performance statistics associated with using prior cost as the only predictor in an identically sized group. That is, we identified the highest cost individuals from Year 1 on the basis of actual experience. This prior cost cohort was selected to be exactly the same size as the number of individuals identified using the various acgpm probability thresholds (see the proportion of population in column 2 of Table 2. Table 2 indicates that the acgpm consistently outperforms prior cost in predicting actual Year 2 cost. Sensitivity for both the acgpm and the prior cost groups is low given that these represent very small highly targeted groups and thus do not capture many cases. However, as the PPVs suggest, most of the identified cases turn out to be truly high cost in the following year. 19

25 Table 2. Model Performance (Total Cost) at Different acgpm Probability Scores vs. Samesized Prior Cost Cohort Probability Score Threshold Percentage of Population 1.33% 0.89% 0.63% 0.42% 0.25% 0.10% acgpm Prior Cost * Sensitivity PPV Sensitivity PPV Based on a validation sample of approximately 410,000 covered lives. * Prior cost cohorts were chosen to the same size as acgpm high-cost cohorts, i.e., if selecting cases with an acgpm probability score of.7 or higher yielded 100 predicted high-risk patients, the prior cost comparison would be the 100 cases at the highest Year 1 cost. ii. Do the Cases Identified by the Model Represent a Group Meriting Intervention? There is a clear distinction in terms of mean cost between cases identified by the acgpm as potentially high-risk and those not so identified. As shown in Figure 3, predicted high-risk cases proved to be nearly 13 times as expensive in actual Year 2 dollars as those not so identified. Individuals identified as being high-risk for pharmacy service use were almost 14 times as expensive as those not so identified. The model, thus, is identifying cases that are quite distinct in terms of very high resource use. 20

26 Figure 3. Mean Costs of High-risk Versus All Other Cases by Cost Category Dollars $15,000 $10,000 $5,000 $0 $13,787 $1,072 Total $2,532 Type of Cost Top 5% All Other $189 Pharmacy *With an acgpm probability score of 0.4 or higher. The model also preferentially captures cases with chronic conditions for which case management services are often available. As shown in Table 3, the acgpm-identified at-risk group includes a higher percentage of these potentially case manageable chronic conditions than does prior cost. Table 3. Percentage of Selected Chronic Conditions in High-risk Cohorts Identified by acgpm and Prior Cost Approaches. Percentage of Condition High-risk Cohort acgpm Prior Cost Hypertension Low Back Pain Diabetes Ischemic Heart Disease Arthritis Lipoid Metabolism Congestive Heart Failure Asthma COPD Depression Chronic Renal Failure

27 iii. How Well Does acgpm Do in Estimating Future Costs of Care? The acgpm offers a predictive resource index that is a relative value for resource use related to both total and pharmacy costs in Year 2. This relative value can be used for many applications including the estimation of future expenditures for specific subgroups of patients who are targeted for case management. The probability model results and the linear modeling results can work in tandem. The acgpm s probability score output is recommended for selecting patients, while the predictive resource index is recommended for calculating expected costs (or potential cost savings) for population subgroups. The R-squared statistic is commonly used to assess the performance of OLS-based linear models. The R-squared expresses the percentage of variation in the outcome variable that is explained by the model. We believe this is an appropriate evaluative benchmark for cost predictors calculated on a linear basis (in our case the relative weight for resource use) but not for yes/no logistic predictions (such as our risk probability score). When assessing our predicted resource index model, the performance characteristics of the acgpm are comparable to prior cost: the acgpm explains 14% of the variation in total charges compared to 12% for prior cost. iv. What Is the Performance Bonus if Pharmacy Cost Data Are Available as a Risk Factor? When they are available, we encourage the use of pharmacy cost data as a source of riskfactor information 2. The acgpm performance statistics presented in the preceding text are based on models that include pharmacy cost as an input variable. We compare the performance of the acgpm probability score based on models with and without pharmacy cost data in Table 4. While there is added information if pharmacy costs are available, if they are not available, the performance penalty is not high, especially in predicting total costs. 3 The acgpm model includes a simple five category variable based on previous pharmacy cost history. Each person is placed into one of five groupings, from very low to very high. 22

28 Table 4. Performance Characteristics of acgpm with and without Inclusion of Prior Pharmacy Cost as a Risk Factor Model Sensitivity PPV C-Statistic acgpm Total Cost acgpm Total Cost with Pharmacy acgpm Pharmacy Cost acgpm Pharmacy Cost with Pharmacy Model performance data are based on an acgpm probability score threshold of 0.6, representing approximately half of 1% of the population for total cost and the top 2% of the population for pharmacy costs. The inclusion of prior pharmacy costs as a predictor variable for the predicted resource index has little impact on the model s R-squared associated with total cost (.14 up from.11). However, the addition of the pharmacy categorical cost information does have a significant impact on the R-squared associated with prediction of next year s pharmacy use (.34 up from.17). Understandably, it does appear that prior pharmacy cost is an important factor in predicting future pharmacy costs. Users who wish to focus on pharmacy cost would be advised to incorporate prior cost data if they are available. f. Reports The software currently puts two reports into the ACG output stream: and (1) High Risk Individuals and Expected Resource Use by Disease Category and (2) Frequency and Percent Distribution of Probability Scores. Each report is discussed below. i. High Risk Individuals and Expected Resource Use by Disease Category This report (see Figure 4) shows how your population is distributed by EDC category and how they are predicted to consume resources among some of the riskier probability levels. The columns represent a range of probability scores from a minimum value to the highest reported value. The 11 EDCs represent chronic conditions that are often the subject of case and disease management initiatives. For an example of how to use this report see the following section (Application Issues). 23

29 Figure 4. acgpm Report One. ii. Frequency and Percent Distribution of Probability Scores This report (see Figure 5) is intended to provide you with a sense of how your population is distributed according to probability score. The information is intended to help you establish the size of potential cohorts for case/disease management. This distribution will, of course, change within specific EDCs. Figure 5. acgpm Report Two. g. Application Issues i. Using the acgpm Case Management Report As described in the section on model performance, the acgpm identifies a group of patients distinct from those who would have been selected by prior cost alone. Further, this group of patients appears to have a higher percentage of those who are typically case managed. To illustrate how the model might be used to better target individualized case management 24

30 interventions, we provide a sample output report (Table 5) as produced by the acgpm software (if users request this with a control card). For a series of probability thresholds, the table shows how individuals in a large health plan are distributed within selected chronic disease groups and how their acgpm predicted resource use varies for each risk-level cohort. The diseases represent a sampling of some of the conditions for which disease management or case management services are often available within an integrated delivery organization. These chronically ill patients are all projected to use resources well above the plan s average (of 1.00). In a comparison of the three potential case management cohorts defined on the basis of the three alternative acgpm probability score thresholds (i.e.,.4 or higher,.6 or higher,.8 or higher), the projected intensity of resource use generally doubles from the lowest to highest risk group within each disease. Moreover, when resource use in the top acgpm risk category (with a score of.8 or higher) is compared with the cohort of persons with the disease, but who were not identified as being in one of the top tier risk groups (i.e., those with probability score of less than.4), the predicted expenditure variation is dramatic at times almost ten-fold. Persons with chronic renal failure experience very high predicted resource use within all of the probability score cohorts. To a lesser degree, the same is true for congestive heart failure. For conditions such as these, projected resource use appears to be high regardless of the probability level. Aside from employing individualized case management, it may be especially appropriate to employ disease management programs that address the cohort of patients with these diseases. It is evident that persons within each disease group will require a different approach to care management, but the various predictive measures produced by the acgpm system will provide valuable additional information to allow clinical professionals to better design and implement these interventions. All of the remaining diseases that are depicted in Table 5 appear to be appropriate candidates for case management. To have the greatest impact, it would be useful to focus on those individuals at higher predictive risk levels who are currently experiencing relatively low resource use. Our experience suggests that at least 10% of patients will fall into this group. Finally, it is important to consider the comorbidity profiles of these high-risk groups. It is likely that high risk cases are affected by multiple diseases and that the condition reported in 25

31 Table 5 may not be the primary cost driver. Thus patterns of comorbidities should also be assessed in the course of planning case management initiatives. Table 5. The Number of Cases and Associated Predicted acgpm Relative Resource Use by Alternative acgpm Risk Probability Thresholds for Selected Chronic Conditions Number of Cases Mean Predicted Relative Probability Score Category Resource Use Probability Score Category Disease Category Total < Arthritis 17, Asthma 27, Diabetes 16,991 1, Hypertension 50,122 2,064 1, Ischemic Heart Disease 9, Congestive Heart Failure 1, Hyperlipidemia 31,240 1, Low Back Pain 61,980 1, Depression 10, Chronic Renal Failure COPD 6, There are trade-offs involved in setting probability threshold levels for case identification. The higher the minimum probability, the greater the projected resource use, but also the smaller the target group. As noted earlier, higher probabilities also improve the likelihood that cases identified do indeed turn out to be high resource users. The performance implications at a range of thresholds based on the acgpm probability score among a large test population are depicted graphically in Figure 6. 26

32 Figure 6. Trends in Model Performance, by Probability Threshold Performance Scale Probability Threshold Sensitivity Positive Predictive Value The user s choice of a case identification probability threshold relates to the desired application, availability of resources for case management, and prevailing local practices. Depending on the intervention an organization might prefer to maximize sensitivity, while at other times, maximization of predictive value might be the goal. For example, when implementing a customized web-based informational campaign within a health plan, the goal may be to contact the highest 5% or 10% of the persons at-risk. This goal could represent all those persons with a probability score of.1 or higher. Based on the sensitivity of the model at these probability levels, it may be possible to capture over 50% of those persons who will be members of the highest resource use cohort next year. However, the PPVs for this group will be in the.25 range, and thus most of these individuals will not actually be members of the highest resource use group next year. But given the non-intensive nature of the intervention, that is of small concern. Moreover, even the individuals targeted for this program who do not end up in the highest resource group in the subsequent period (the false positives) will still have risk and associated resource use that is far higher than the underlying population. Many types of individualized case management can be quite resource-intensive in their own right, and so there should be a high likelihood that persons targeted for case management are, in fact, truly high risk. Thus for these types of programs, very high PPVs are desirable. Still, many patients who might be appropriate candidates for case management will be excluded from 27

33 selection, given the corresponding low sensitivities. It should be noted that even when cases are included that have probabilities as low as.4, the overall odds of these individuals being very high resource users next year are still much higher than random. In this case about 60% will be high risk next year compared to 5% of the population overall if selected at random. For individualized case management programs, having access to current and predicted resource use data for candidate individuals can be of real utility in setting priorities for inclusion. When there are more persons who would likely benefit than current program resources allow, these data could be used to help make an argument for expanding such resources on the basis of the potential return on investment (ROI). Using the approach described below (see the section entitled Adjusting Relative Weights to Compute Predicted Costs), you can assign a projected cost to each individual within your targeted high-risk group. Ordering this population by the difference between current and projected cost will quickly highlight those individuals for whom case management has the greatest potential. 28

34 APPENDICES Appendix 1. ACGs and EDCs Included in the acgpm ACGs 4220: 4-5 Other ADG Combinations, Age 1-17, 1+ Major ADGs 4330: 4-5 Other ADG Combinations, Age 18-44, 2+ Major ADGs 4420: 4-5 Other ADG Combinations, Age >44, 1 Major ADGs 4430: 4-5 Other ADG Combinations, Age >44, 2+ Major ADGs 4510: 6-9 Other ADG Combinations, Age 1-5, No Major ADGs 4520: 6-9 Other ADG Combinations, Age 1-5, 1+ Major ADGs 4610: 6-9 Other ADG Combinations, Age 6-17, No Major ADGs 4620: 6-9 Other ADG Combinations, Age 6-17, 1+ Major ADGs 4730: 6-9 Other ADG Combinations, Male, Age 18-34, 2+ Major ADGs 4830: 6-9 Other ADG Combinations, Female, Age 18-34, 2+ Major ADGs 4910: 6-9 Other ADG Combinations, Age >34, 0-1 Major ADGs 4920: 6-9 Other ADG Combinations, Age >34, 2 Major ADGs 4930: 6-9 Other ADG Combinations, Age >34, 3 Major ADGs 4940: 6-9 Other ADG Combinations, Age >34, 4+ Major ADGs 5010: 10+ Other ADG Combinations, Age 1-17 No Major ADGs 5020: 10+ Other ADG Combinations, Age 1-17, 1 Major ADGs 5030: 10+ Other ADG Combinations, Age 1-17, 2+ Major ADGs 5040: 10+ Other ADG Combinations, Age 18+, 0-1 Major ADGs 5050: 10+ Other ADG Combinations, Age 18+, 2 Major ADGs 5060: 10+ Other ADG Combinations, Age 18+, 3 Major ADGs 5070: 10+ Other ADG Combinations, Age 18+, 4+ Major ADGs 5320: Infants: 0-5 ADGs, 1+ Major ADGs 5330: Infants: 6+ ADGs, No Major 5340: Infants: 6+ ADGs, 1+ Major ADG Pregnancy w/out Delivery EDCs ADM04: Complications of Mechanical Devices ALL04: Asthma, w/o status asthmaticus ALL05: Asthma, WITH status asthmaticus ALL06: Disorders of the Immune System CAR03: Ischemic heart disease (excluding acute myocardial infarction) CAR04: Congenital heart disease CAR05: Congestive heart failure CAR06: Cardiac valve disorders CAR07: Cardiomyopathy CAR09: Cardiac arrhythmia CAR14: Hypertension, w/o major complications CAR15: Hypertension, WITH major complications END02: Osteoporosis END09: Type 1 Diabetes with major complicating condition 29

35 END08: Type 1 Diabetes w/o major complicating condition END07: Type 2 Diabetes with major complicating condition END06: Type 2 Diabetes w/o major complicating condition EYE13: Diabetic Retinopathy GAS02: Inflammatory bowel disease GAS05: Chronic liver diseases GAS06: Peptic ulcer disease GAS12: Chronic pancreatitis GSI08: Edema GSU11: Peripheral vascular disease GSU13: Aortic aneurysm GSU14: Gastrointestinal Obstruction/Perforation GTC01: Chromosomal anomalies GUR04: Prostatic hypertrophy HEM01: Hemolytic anemia HEM05: Aplastic anemia HEM06: Deep vein thrombosis HEM07: Hemophilia, coagulation Disorder INF04: HIV, AIDS MAL02: Low impact malignant neoplasms MAL03: High impact malignant neoplasms MAL04: Malignant neoplasms, breast MAL06: Malignant neoplasms, ovary MAL07: Malignant neoplasms, esophagus MAL08: Malignant neoplasms, kidney MAL09: Malignant neoplasms, liver and biliary tract MAL10: Malignant neoplasms, lung MAL11: Malignant neoplasms, lymphomas MAL12: Malignant neoplasms, colorectal MAL13: Malignant neoplasms, pancreas MAL14: Malignant neoplasms, prostate MAL15: Malignant neoplasms, stomach MAL16: Acute Leukemias MAL18: Malignant neoplasms, bladder MUS10: Fracture of neck of femur (hip) MUS14: Low back pain NUR05: Cerebrovascular disease NUR07: Seizure disorders NUR08: Multiple sclerosis NUR09: Muscular dystrophy NUR12: Quadriplegia and Paraplegia NUR15: Head Injury NUR16: Spinal Cord Injury/Disorders NUR17: Paralytic Syndromes, Other NUR18: Cerebral Palsy NUR19: Developmental disorders NUT02: Nutritional deficiencies 30

36 PSY01: Anxiety, neuroses PSY03: Tobacco abuse PSY05: Attention deficit disorder PSY07: Schizophrenia and affective psychosis PSY08: Personality disorders PSY09: Depression REC01: Cleft lip and palate REC03: Chronic ulcer of the skin REN01: Chronic renal failure RES03: Cystic fibrosis RES04: COPD RES09: Tracheostomy ACGs Included in the Three Resource Utilization Bands (RUBs) Reference Group ACG RUB Level 1 (Included in Intercept) ACG ACG ACG RUB Level ACG ACG RUB Level

37 Appendix 2. Summary of Performance Measures for Diagnostic Testing* Screening Test + TP (True Positive) FN - (False Negative) True Condition + - Predictive Value + = TP/ TP + FP Predictive Value - = TN/ FN + TN Sensitivity = TP/ TP + FN Specificity = TN/ FP + TN FP (False Positive) TN (True Negative) Receiver Operating Characteristic (ROC) = Function of Sensitivity vs. 1- Specificity at different screening thresholds (C Statistic = area under curve) * In the context of predictive modeling, a + for the true condition reflects a person being a member of the high-risk cohort (i.e., represents with the top 5% of costs in Year 2). The + for the screening test in this case would represent an individual with an acgpm probability score that is about the threshold considered actionable (e.g., a score that would place a person in the top half of 1% of the population. 32

38 Section 3 Improvements to the Johns Hopkins Expanded Diagnosis Clusters This section discusses Release 6.0 improvements to the Johns Hopkins expanded diagnosis clusters (EDCs) methodology, including: refinements of the EDC taxonomy, and enhanced reporting features. New users of the software are encouraged to review Chapter 13, Dino-Clusters : The Johns Hopkins Expanded Diagnosis Clusters (EDCs) of the Version 5.0 Documentation and Application Manual for complete details of the EDC taxonomy, a discussion of the EDC output file, and technical specifications for customizing EDC reports. a. Overview Expanded diagnosis clusters (EDCs) were added to the ACG System in 2001 to provide the ability to partition populations into diagnosis-specific subgroups for a more complete understanding of case mix. Our intent was to condense the unwieldy list of ICD-9 codes into a much smaller number of clinically homogeneous clusters. As was the case for the original diagnosis clusters, our system places an emphasis on commonly occurring conditions treated primarily in ambulatory settings. The EDC system has undergone considerable refinement and now also includes less commonly occurring conditions, many of which may be treated in the hospital. In some instances, the original EDCs have been further subdivided to better reflect the effect of complicated illness, examples of which are presented on Table 1. Table 7. Subdivision of EDCs for Additional Clinical Specificity VERSION 5 EDC REVISED VERSION 6 EDCs ALL02 Asthma ALL04 Asthma, w/o status asthmaticus ALL05 Asthma WITH status asthmaticus CAR02 Hypertension CAR14 Hypertension, w/o major complications CAR15 Hypertension, WITH major complications END01 Diabetes Mellitus END06 Type 2 diabetes w/o major complicating conditions END07 Type 2 diabetes WITH major complicating conditions END08 Type 1 diabetes w/o major complicating conditions END09 Type 1 diabetes WITH major complicating conditions As these examples illustrate, in some instances the main refinement was to split the original category into complicated and uncomplicated subgroups. As a case in point, the asthma 33

39 EDC has been divided on the basis of the presence of status asthmaticus (i.e., an acute exacerbation of asthma). In the cardiology cluster, hypertension has been divided on the basis of the presence of a complicating condition. In other instances, the disease itself was further delineated, such as within the endocrine cluster where diabetes mellitus has been split into four categories first by categorizing patients according to type 1 versus type 2 diabetes, and second according to the presence of a major, complicating comorbidity. (See Table A.1 in the Appendix to this chapter for a complete listing of these complicating conditions.) In addition to splitting existing EDCs, new categories have been added to the current software release. Table 2 provides a summary of new EDCs. In most instances, the additions within a given major EDC have been modest. For example, in the Administrative cluster, categories for transplant status and complications of mechanical devices were added. In the Cardiology cluster, a new category for acute myocardial infarction (CAR11) and Cardiac arrest/shock (CAR12) have been added. The malignancy and neurologic MEDC categories have been substantially expanded. All in all, 40 new EDC categories have been added and affect 16 of the 27 major EDC (MEDC) clusters, and several existing EDCs have been subdivided to provide additional clinical specificity. There are now a total of 236 EDC categories in ACG Release 6.0. For a sense of how these EDCs are distributed in the real world, their frequency of occurrence in a large under-65 population (2 million covered lives) is provided in Appendix A.2. 34

40 Table 8. Supplemental EDCs by MEDC Category MEDC/EDC CATEGORY DESCRIPTION MEDC/EDC CATEGORY DESCRIPTION Administrative Malignancies (continued) ADM03 Transplant status MAL10 Malignant neoplasm, lung ADM04 Complications of mechanical devices MAL11 Malignant neoplasm, lymphomas Allergy MAL12 Malignant neoplasm, colorectal ALL06 Disorders of the immune system MAL13 Malignant neoplasm, pancreas MAL14 Malignant neoplasm, prostate Cardiovascular MAL15 Malignant neoplasm, stomach CAR11 Disorders of lipoid metabolism MAL16 Acute leukemia CAR12 Acute Myocardial Infarction MAL18 Malignant neoplasm, bladder Eye Musculoskeletal EYE13 Diabetic retinopathy MUS16 Amputation status Gastrointestinal- Hepatic Neurologic GAS11 Acute pancreatitis NUR12 Quadriplegia and paraplegia GAS12 Chronic pancreatitis NUR15 Head injury General Surgery NUR16 Spinal cord injury/disorders GSU13 Aortic aneurysm NUR17 Paralytic syndromes, other GSU14 Gastrointestinal obstruction/perforation NUR18 Cerebral palsy Hematologic NUR19 Developmental disorder HEM06 Aplastic anemia Psychosocial HEM07 Deep vein thrombosis PSY09 Depression Infections Renal HEM08 Septicemia REN03 Acute renal failure Malignancies REN04 Nephritis/nephrosis MAL04 Malignant neoplasm, breast Respiratory MAL05 Malignant neoplasm, cervix, uterus RES08 Pulmonary embolism MAL06 Malignant neoplasm, ovary RES09 Tracheostomy Malignant neoplasm, esophagus RES10 Respiratory arrest MAL07 MAL08 Malignant neoplasm, kidney Skin Malignant neoplasm, liver MAL09 and biliary tract SKN18 Benign neoplasm of skin and subcutaneous tissues b. New Reporting Features The EDC reporting capabilities of ACG Release 6.0 have been significantly enhanced to facilitate their easy implementation within your organization. The current automatically generated reports include EDC distributions, tables combining disease-specific EDCs and the morbidity-class, as well as age/sex-adjusted comparison of EDC distributions across populations. 35

41 EDCxRUB Reports A series of reports combining the disease-specific EDC and ACG methodologies provides managers with information that should be useful in better targeting their care management programs. Reports are generated first by major EDC (or MEDC) level to provide a summary overview, and then this set of summary tables is generated for each specific diagnosis cluster. Examples of the new reports are provided in Tables 3 and 4. Table 3: Percent distribution of each co-morbidity level within EDC (Samples) RUB-1 RUB-2 RUB-3 RUB-4 RUB-5 EDC Very Low Low Average High Very High ADM01:General medical exam ADM02:Surgical aftercare ADM03:Transplant status ADM04:Complications of mechanical ALL01:Allergic reactions ALL03:Allergic rhinitis ALL04:Asthma, w/o status asthmati ALL05:Asthma, with status asthmat ALL06:Disorders of the immune sys CAR01:Cardiovascular signs and sy CAR03:Ischemic heart disease (exc CAR04:Congenital heart disease CAR05:Congestive heart failure CAR06:Cardiac valve disorders CAR07:Cardiomyopathy CAR08:Heart murmur CAR09:Cardiac arrhythmia CAR10:Generalized atherosclerosis CAR11:Disorders of lipoid metabol CAR12:Acute myocardial infarction CAR13:Cardiac arrest, shock CAR14:Hypertension, w/o major com CAR15:Hypertension, with major co Each row in the tables represents a separate MEDC (or EDC) category, and the columns array individuals within a particular MEDC (or EDC) into five (from very low to very high) morbidity groupings that we term resource utilization bands (RUBs). RUB categories are based on ACG assignments (see Chapter 8, Calibrating ACGs for Intended Use: ACG Weights, Resource Utilization Bands, and Carve-Outs, in the Release 5.0 Documentation and Application Manual). The first part of this two-part table (Table 3) presents the percentage distribution for each EDCxRUB comorbidity level, and the second (Table 4) presents an estimate of each group s expected relative resource use. Focusing on the EDC for general medical exam, ADM01, you see that the first row of Table 3 shows 19.8% of users with this EDC fell into RUB-1 or a very low resource group, 32.9% fell into the low resource RUB-2, 39.9% fell into 36

42 average resource RUB-3, and so on. Looking at the same row in Table 4, you see that the anticipated resource use for such individuals is ranges from a low of 0.19 to a high of for those in the highest resource group. Resource estimates provided in these tables are based on nationally representative ACG weights built into the software (see the Using the Available Relative Value Weights section for additional details on concurrent weights and RUB assignments included as part of Release 6.0). These tables help to illustrate the variability of costs within disease categories and will be useful to managers for better understanding resource use. Generally it is not necessarily all individuals with selected diseases who are expensive; rather, it is individuals with multiple comorbidities who consume most of the health care resources. Table 4: Estimated Concurrent Resource Use by RUB by EDC (Samples) RUB-1 RUB-2 RUB-3 RUB-4 RUB-5 EDC Very Low Low Average High Very High ADM01:General medical exam ADM02:Surgical aftercare ADM03:Transplant status ADM04:Complications of mechanical ALL01:Allergic reactions ALL03:Allergic rhinitis ALL04:Asthma, w/o status asthmati ALL05:Asthma, with status asthmat ALL06:Disorders of the immune sys CAR01:Cardiovascular signs and sy CAR03:Ischemic heart disease (exc CAR04:Congenital heart disease CAR05:Congestive heart failure CAR06:Cardiac valve disorders CAR07:Cardiomyopathy CAR08:Heart murmur CAR09:Cardiac arrhythmia CAR10:Generalized atherosclerosis CAR11:Disorders of lipoid metabol CAR12:Acute myocardial infarction CAR13:Cardiac arrest, shock CAR14:Hypertension, w/o major com CAR15:Hypertension, with major co Standardized Morbidity Ratios The current release generates a series of age/sex standardized EDC-based morbidity ratio tables to assist with population-level profiling based on a user-defined population stratifier that can be provided by means of the software s input file (see the Installation and Usage Guide for the technical details of implementing this feature). Separate reports are generated for each population group defined by the user. Such information can help practitioners and managers understand which specific conditions within a subgroup of interest are more or less common (beyond statistical chance) than the overall population average. As illustrated in Table 5, summary statistics generated for each group in this case, by MEDC category, include 37

43 observed prevalence rates, age/sex-expected prevalence, standardized morbidity ratio (SMR), as well as low and high indicators for statistical significance at the 95% confidence interval. Table 5. Age/sex-adjusted Comparison of Disease Distributions across Populations ***Observed to Expected Standardized Morbidity Ratio by MEDC*** Population: ALL Number of persons=2,141,852 Major EDC Observed Age-Sex Standardized Prevalence Expected Morbidity 95% per 1000 Prevalence Ratio Confidence Interval Population per 1000 (SMR) (Low) (High) Administrative Allergy Cardiovascular Dental Ears, Nose, Throat Endocrine Eye Female Reproductive Gastrointestinal/Hepatic General Signs and Symptoms General Surgery Genetic Genito-urinary Hematologic Infections Malignancies Musculoskeletal Neurologic Nutrition Psychosocial Reconstructive Renal Respiratory Rheumatologic Skin Toxic Effects Unassigned A Note on Customizing These Reports The technical specifications for creating these tables on other population subgroups and/or based on locally calibrated cost data are provided in Section d. Dino-Cluster Applications/Approaches of Chapter 13 in the Release 5.0 Documentation and Application Guide. Although these tables are built into the software to allow quick implementation, to maximize the usefulness of this type of information it is recommended that they be re-created and calibrated to local data. 38

44 Appendix A.1 Table A.1: List of Complicating Conditions Used to Split Diabetes EDCs ICD-9 Code Description 2501 DIABETES W KETOACIDOSIS* DMII KETO NT ST UNCNTRLD DMI KETO NT ST UNCNTRLD DMII KETOACD UNCONTROLD DMI KETOACD UNCONTROLD 2502 DIAB W HYPEROSMOLAR COMA* DMII HPRSM NT ST UNCNTRL DMI HPRSM NT ST UNCNTRLD DMII HPROSMLR UNCONTROLD DMI HPROSMLR UNCONTROLD 2503 DIABETES WITH COMA NEC* DMII O CM NT ST UNCNTRLD DMI O CM NT ST UNCNTRLD DMII OTH COMA UNCONTROLD DMI OTH COMA UNCONTROLD 2504 DIAB W RENAL MANIFEST* DMII RENL NT ST UNCNTRLD DMI RENL NT ST UNCNTRLD DMII RENAL UNCNTRLD DMI RENAL UNCNTRLD 410 ACUTE MYOCARDIAL INFARCT* 4100 AMI ANTEROLATERAL WALL* AMI ANTEROLATERAL,UNSPEC AMI ANTEROLATERAL, INIT AMI ANTEROLATERAL,SUBSEQ 4101 AMI ANTERIOR WALL NEC* AMI ANTERIOR WALL,UNSPEC AMI ANTERIOR WALL, INIT AMI ANTERIOR WALL,SUBSEQ 4102 AMI INFEROLATERAL WALL* AMI INFEROLATERAL,UNSPEC AMI INFEROLATERAL, INIT AMI INFEROLATERAL,SUBSEQ 4103 AMI INFEROPOSTERIOR WALL* AMI INFEROPOST, UNSPEC AMI INFEROPOST, INITIAL AMI INFEROPOST, SUBSEQ 4104 AMI INFERIOR WALL NEC* AMI INFERIOR WALL,UNSPEC AMI INFERIOR WALL, INIT AMI INFERIOR WALL,SUBSEQ 39

45 ICD-9 Code Description 4105 AMI LATERAL WALL NEC* AMI LATERAL NEC, UNSPEC AMI LATERAL NEC, INITIAL AMI LATERAL NEC, SUBSEQ 4106 TRUE POSTERIOR INFARCT* TRUE POST INFARCT,UNSPEC TRUE POST INFARCT, INIT TRUE POST INFARCT,SUBSEQ 4107 SUBENDOCARDIAL INFARCT* SUBENDO INFARCT, UNSPEC SUBENDO INFARCT, INITIAL SUBENDO INFARCT, SUBSEQ 4108 MYOCARDIAL INFARCT NEC* AMI NEC, UNSPECIFIED AMI NEC, INITIAL AMI NEC, SUBSEQUENT 4109 MYOCARDIAL INFARCT NOS* AMI NOS, UNSPECIFIED AMI NOS, INITIAL AMI NOS, SUBSEQUENT 411 OTH AC ISCHEMIC HRT DIS* 4110 POST MI SYNDROME 4111 INTERMED CORONARY SYND 4118 AC ISCHEMIC HRT DIS NEC* CORONARY OCCLSN W/O MI AC ISCHEMIC HRT DIS NEC 412 OLD MYOCARDIAL INFARCT 413 ANGINA PECTORIS* 4130 ANGINA DECUBITUS 4131 PRINZMETAL ANGINA 4139 ANGINA PECTORIS NEC/NOS 414 OTH CHR ISCHEMIC HRT DIS* 4140 CORONARY ATHEROSCLEROSIS* COR ATH UNSP VSL NTV/GFT CRNRY ATHRSCL NATVE VSSL CRN ATH ATLG VN BPS GRFT CRN ATH NONATLG BLG GRFT COR ATH ARTRY BYPAS GRFT COR ATH BYPASS GRAFT NOS 4141 ANEURYSM OF HEART* ANEURYSM, HEART (WALL) CORONARY VESSEL ANEURYSM ANEURYSM OF HEART NEC 4148 CHR ISCHEMIC HRT DIS NEC 40

46 ICD-9 Code Description 4149 CHR ISCHEMIC HRT DIS NOS 581 NEPHROTIC SYNDROME* 5810 NEPHROTIC SYN, PROLIFER 5811 EPIMEMBRANOUS NEPHRITIS 5812 MEMBRANOPROLIF NEPHROSIS 5813 MINIMAL CHANGE NEPHROSIS 5818 NEPHROTIC SYN W OTH LES* NEPHROTIC SYN IN OTH DIS NEPHROTIC SYNDROME NEC 5819 NEPHROTIC SYNDROME NOS 582 CHRONIC NEPHRITIS* 5820 CHR PROLIFERAT NEPHRITIS 5821 CHR MEMBRANOUS NEPHRITIS 5822 CHR MEMBRANOPROLIF NEPHR 5824 CHR RAPID PROGR NEPHRIT 5828 CHR NEPHRITIS W OTH LES* CHR NEPHRITIS IN OTH DIS CHRONIC NEPHRITIS NEC 5829 CHRONIC NEPHRITIS NOS 583 NEPHRITIS NOS* 5830 PROLIFERAT NEPHRITIS NOS 5831 MEMBRANOUS NEPHRITIS NOS 5832 MEMBRANOPROLIF NEPHR NOS 5834 RAPIDLY PROG NEPHRIT NOS 5836 RENAL CORT NECROSIS NOS 5837 NEPHR NOS/MEDULL NECROS 5838 NEPHRITIS NOS W OTH LES* NEPHRITIS NOS IN OTH DIS NEPHRITIS NEC 5839 NEPHRITIS NOS 584 ACUTE RENAL FAILURE* 5845 LOWER NEPHRON NEPHROSIS 5846 AC RENAL FAIL, CORT NECR 5847 AC REN FAIL, MEDULL NECR 5848 AC RENAL FAILURE NEC 5849 ACUTE RENAL FAILURE NOS 585 CHRONIC RENAL FAILURE 586 RENAL FAILURE NOS V56 DIALYSIS ENCOUNTER* V560 RENAL DIALYSIS ENCOUNTER V561 FT/ADJ XTRCORP DIAL CATH V562 FIT/ADJ PERIT DIAL CATH V568 DIALYSIS ENCOUNTER, NEC 41

47 Appendix A.2 Table A.2: Major Expanded Diagnosis Clusters and their Component Expanded Diagnosis Clusters Number and Prevelance per Thousand of Major Expanded Diagnosis Clusters and their Component Expanded Diagnosis Clusters EDC Description No. No. Persons Persons per 1000 Population ADM Administrative ADM01 General medical exam ADM02 Surgical aftercare ADM03 Transplant status ADM04 Complications of mechanical devices ALL Allergy ALL01 Allergic reactions ALL03 Allergic rhinitis ALL04 Asthma, w/o status asthmaticus ALL05 Asthma, with status asthmaticus ALL06 Disorders of the immune system CAR Cardiovascular CAR01 Cardiovascular signs and symptoms CAR03 Ischemic heart disease (excluding acute myocard CAR04 Congenital heart disease CAR05 Congestive heart failure CAR06 Cardiac valve disorders CAR07 Cardiomyopathy CAR08 Heart murmur CAR09 Cardiac arrhythmia CAR10 Generalized atherosclerosis CAR11 Disorders of lipoid metabolism CAR12 Acute myocardial infarction CAR13 Cardiac arrest, shock CAR14 Hypertension, w/o major complications CAR15 Hypertension, with major complications DEN Dental DEN01 Disorders of mouth DEN02 Disorders of teeth DEN03 Gingivitis DEN04 Stomatitis EAR Ears, Nose, Throat EAR01 Otitis media EAR02 Tinnitus EAR03 Temporomandibular joint disease EAR04 Foreign body in ears, nose, or throat EAR05 Deviated nasal septum EAR06 Otitis externa EAR07 Wax in ear EAR08 Deafness, hearing loss EAR09 Chronic pharyngitis and tonsillitis EAR10 Epistaxis EAR11 Acute upper respiratory tract infection END Endocrine END02 Osteoporosis END03 Short stature END04 Thyroid disease END05 Other endocrine disorders END06 Type 2 diabetes, w/o complication END07 Type 2 diabetes w/complications END08 Type 1 diabetes, w/o complication END09 Type 1 diabetes w/complications EYE Eye EYE01 Ophthalmic signs and symptoms EYE02 Blindness EYE03 Retinal disorders (excluding diabetic retinopat EYE04 Disorders of the eyelid and lacrimal duct EYE05 Refractive errors

48 EYE06 Cataract, aphakia EYE07 Conjunctivitis, keratitis EYE08 Glaucoma EYE09 Infections of eyelid EYE10 Foreign body in eye EYE11 Strabismus, amblyopia EYE12 Traumatic injuries of eye EYE13 Diabetic retinopathy FRE Female Reproductive FRE01 Pregnancy and delivery, uncomplicated FRE02 Female genital symptoms FRE03 Endometriosis FRE04 Pregnancy and delivery with complications FRE05 Female infertility FRE06 Abnormal pap smear FRE07 Ovarian cyst FRE08 Vaginitis, vulvitis, cervicitis FRE09 Menstrual disorders FRE10 Contraception FRE11 Menopausal symptoms FRE12 Utero-vaginal prolapse GAS Gastrointestinal/Hepatic GAS01 Gastrointestinal signs and symptoms GAS02 Inflammatory bowel disease GAS03 Constipation GAS04 Acute hepatitis GAS05 Chronic liver disease GAS06 Peptic ulcer disease GAS07 Diarrhea GAS08 Gastroesophageal reflux GAS09 Irritable bowel syndrome GAS10 Diverticular disease of colon GAS11 Acute pancreatitis GAS12 Chronic pancreatitis GSI General Signs and Symptoms GSI01 Nonspecific signs and symptoms GSI02 Chest pain GSI03 Fever GSI04 Syncope GSI05 Nausea, vomiting GSI06 Debility and undue fatigue GSI07 Lymphadenopathy GSI08 Edema GSU General Surgery GSU01 Anorectal conditions GSU02 Appendicitis GSU03 Benign and unspecified neoplasm GSU04 Cholelithiasis, cholecystitis GSU05 External abdominal hernias, hydroceles GSU06 Chronic cystic disease of the breast GSU07 Other breast disorders GSU08 Varicose veins of lower extremities GSU09 Nonfungal infections of skin and subcutaneous t GSU10 Abdominal pain GSU11 Peripheral vascular disease GSU12 Burns--1st degree GSU13 Aortic aneurysm GSU14 Gastrointestinal obstruction/perforation GTC Genetic GTC01 Chromosomal anomalies GUR Genito-urinary GUR01 Vesicoureteral reflux GUR02 Undescended testes GUR03 Hypospadias, other penile anomalies GUR04 Prostatic hypertrophy GUR05 Stricture of urethra GUR06 Urinary symptoms GUR07 Other male genital disease GUR08 Urinary tract infections GUR09 Renal calculi GUR10 Prostatitis HEM Hematologic HEM01 Hemolytic anemia

49 HEM02 Iron deficiency, other deficiency anemias HEM03 Thrombophlebitis HEM04 Neonatal jaundice HEM05 Aplastic anemia HEM06 Deep vein thrombosis HEM07 Hemophilia, coagulation disorder INF Infections INF01 Tuberculosis infection INF02 Fungal infections INF03 Infectious mononucleosis INF04 HIV, AIDS INF05 Sexually transmitted diseases INF06 Viral syndromes INF07 Lyme disease INF08 Septicemia MAL Malignancies MAL01 Malignant neoplasms of the skin MAL02 Low impact malignant neoplasms MAL03 High impact malignant neoplasms MAL04 Malignant neoplasms, breast MAL05 Malignant neoplasms, cervix, uterus MAL06 Malignant neoplasms, ovary MAL07 Malignant neoplasms, esophagus MAL08 Malignant neoplasms, kidney MAL09 Malignant neoplasms, liver and biliary tract MAL10 Malignant neoplasms, lung MAL11 Malignant neoplasms, lymphomas MAL12 Malignant neoplasms, colorectal MAL13 Malignant neoplasms, pancreas MAL14 Malignant neoplasms, prostate MAL15 Malignant neoplasms, stomach MAL16 Acute leukemia MAL18 Malignant neoplasms, bladder MUS Musculoskeletal MUS01 Musculoskeletal signs and symptoms MUS02 Acute sprains and strains MUS03 Degenerative joint disease MUS04 Fractures (excluding digits) MUS05 Torticollis MUS06 Kyphoscoliosis MUS07 Congenital hip dislocation MUS08 Fractures and dislocations/digits only MUS09 Joint disorders, trauma related MUS10 Fracture of neck of femur (hip) MUS11 Congenital anomalies of limbs, hands, and feet MUS12 Acquired foot deformities MUS13 Cervical pain syndromes MUS14 Low back pain MUS15 Bursitis, synovitis, tenosynovitis MUS16 Amputation status NUR Neurologic NUR01 Neurologic signs and symptoms NUR02 Headaches NUR03 Peripheral neuropathy, neuritis NUR04 Vertiginous syndromes NUR05 Cerebrovascular disease NUR06 Parkinson's disease NUR07 Seizure disorder NUR08 Multiple sclerosis NUR09 Muscular dystrophy NUR10 Sleep problems NUR11 Dementia and delirium NUR12 Quadriplegia and paraplegia NUR15 Head injury NUR16 Spinal cord injury/disorders NUR17 Paralytic syndromes, other NUR18 Cerebral palsy NUR19 Developmental disorder NUT Nutrition NUT01 Failure to thrive NUT02 Nutritional deficiencies NUT03 Obesity PSY Psychosocial PSY01 Anxiety, neuroses

50 PSY02 Substance use PSY03 Tobacco abuse PSY04 Behavior problems PSY05 Attention deficit disorder PSY06 Family and social problems PSY07 Schizophrenia and affective psychosis PSY08 Personality disorders PSY09 Depression REC Reconstructive REC01 Cleft lip and palate REC02 Lacerations REC03 Chronic ulcer of the skin REC04 Burns--2nd and 3rd degree REN Renal REN01 Chronic renal failure REN02 Fluid/electrolyte disturbances REN03 Acute renal failure REN04 Nephritis, nephrosis RES Respiratory RES01 Respiratory signs and symptoms RES02 Acute lower respiratory tract infection RES03 Cystic fibrosis RES04 Emphysema, chronic bronchitis, COPD RES05 Cough RES06 Sleep apnea RES07 Sinusitis RES08 Pulmonary embolism RES09 Tracheostomy RES10 Respiratory arrest RHU Rheumatologic RHU01 Autoimmune and connective tissue diseases RHU02 Gout RHU03 Arthropathy RHU04 Raynaud's syndrome SKN Skin SKN01 Contusions and abrasions SKN02 Dermatitis and eczema SKN03 Keloid SKN04 Acne SKN05 Disorders of sebaceous glands SKN06 Sebaceous cyst SKN07 Viral warts and molluscum contagiosum SKN08 Other inflammatory conditions of skin SKN09 Exanthems SKN10 Skin keratoses SKN11 Dermatophytoses SKN12 Psoriasis SKN13 Disease of hair and hair follicles SKN14 Pigmented nevus SKN15 Scabies and pediculosis SKN16 Diseases of nail SKN17 Other skin disorders SKN18 Benign neoplasm of skin and subcutaneous tissue TOX Toxic Effects TOX01 Toxic effects of nonmedicinal agents TOX02 Adverse effects of medicinal agents UDC Unassigned UDC00 Unassigned diagnosis code

51

52 Section 4 Using the Software-provided ACG Relative Value Weights With Release 6.0 of the ACG software, weights are for the first time being made available as part of the ACG output stream. Weights is the term that we have traditionally used to represent measures of the level of resource use that are associated with an ACG assignment. Essentially they represent an average resource use expectation for a particular ACG category and are generally based on local data. Weights can be expressed as actual dollars expected to be spent over a period of time or as relative values (the ratio of expected use in that ACG to an overall population mean). This section discusses the use of new internal relative value weights and the conversion of these scores to dollar amounts. Readers are especially encouraged to review the section on the rescaling process before using these weights. For an extensive discussion on the computation of weights and calibration of local data, see Chapter 8, Calibrating ACGs for Intended Use: ACG Weights, Resource Utilization Bands, and Carveouts, in the Release 5.0 Documentation and Application Manual). a. Concurrent ACG Weights A fixed set of concurrent ACG weights is now available as part of the software output file (see Installation and Usage Guide for instructions on how to turn on this option). These are relative weights, i.e., relative to a population mean, and are standardized to a mean of 1.0. The software-supplied weights may be considered a national reference or benchmark for comparisons with locally calibrated ACG weights. However, in some instances (e.g., for those with limited or no cost data), these weights may also be used as a reasonable proxy for local cost data. Table 2 at the end of this section is a complete listing of ACGs and their corresponding nationally representative concurrent ACG weight. (See the additional discussion below about the importance of rescaling so that dollars are not over- or under-predicted). Our experience indicates that concurrent or retrospective ACG weights, especially when expressed as relative values, have remarkable stability. Where differences in ACG weights across plans are present, it is almost universally attributable to differences in covered services reflected by different benefit levels. The new software-provided concurrent weights were developed from a nationally representative database comprising approximately two million lives with comprehensive benefit coverage. 47

53 Ideally, ACG weights should be calculated from plan-specific local data to account most accurately for benefit levels and area practice patterns. The reference population (on which the weights are developed) should be as similar as possible to the assessment population to which the weights are applied. However, and as noted, in the absence of local cost data the softwareprovided internal weights may prove useful for calculating reasonably representative profiling statistics (see Chapter 12, ACG Risk Adjustment and Provider Profiling, in the Release 5.0 Documentation and Application Manual). b. Converting Scores to Dollars Both the ACG concurrent weights and the new acgpm Predicted Risk Index (PRI, see Section 2 of this document) are expressed as relative values, where the mean is centered at 1.0. Individuals with scores higher than 1.0 are more expensive than average whereas those with scores less than 1.0 are less expensive than average. Such relative indices can easily be converted to dollar amounts by multiplying by the underlying mean of the population to which the risk adjustment values will be applied. Before converting scores to dollar amounts, it is important to rescale the data so as to account for differences between the reference population (in this case the Johns Hopkins nationally representative database comprising over two million covered lives) and the population to which the weights are applied (e.g., your population of interest). Rescaling is necessary to assure that the underlying mean of the predictions is 1.0. A similar process is undertaken when you use your own reference population, when it has somewhat different characteristics or circumstances (e.g., it is from a previous time period, or benefit coverage is somewhat different). Unless rescaling is done, resource use (or payments) may be over- or under-predicted. Table 1 below and the accompanying discussion provide a simplified example for a population with only 12 members. c. The Rescaling Process Step 1: Compute population mean weight. Compute a separate grand mean for each of the weights (either concurrent ACG weights or the acgpm PRI) generated for your population 48

54 (the observations represent individuals). The mean for this example is shown in Table 1 at the bottom of column B. Step 2: Apply weighting factor. Divide each individual weight by the rescaling factor (i.e., the mean) that you computed in Step 1. The result is the rescaled relative weight (column C). Step 3: Compute population mean cost. For the same population on which the weights were based, compute the mean cost for the current data year. For this example, the mean cost was $1, Step 4: Compute cost. Multiply the rescaled relative weights generated for each member of the population (Column C) by the average population cost generated from Step 3 to calculate an estimated individual cost (column D). Table 1: Estimating Costs in a Sample of Cases A B C D Relative Rescaled Estimated Observation Weight Weight Cost $ $ $ $ $ $ $ $ $1, $1, $2, $5, Mean $1, The rescaling factor functions as a summary case-mix index for understanding how the rating population (e.g., your local population) compares to the development data (JHU s nationally representative database). The interpretation of this factor is analogous to how one interprets both relative weights and profiling indicators. If the rescaling factor is greater than

55 (as it was in the example), then your population is sicker; if the factor is less than 1.0, then your population is healthier than the reference population. d. Adjustments for Inflation If you are going to use the scores for predicting future expenditures it may be appropriate to inflation-adjust these values. Based on Bureau of Labor Statistics results, for the calendar year 2002 medical care costs rose by approximately 4.7% over the previous year (see In the preceding example, if you were going to apply this inflation adjustment, you would multiply the mean cost computed in Step 3 by to reflect inflation. For this example, the inflation-adjusted mean cost for the next year would have been $1, instead of $1, Depending on the local situation, it may also be appropriate to modify future cost expectations for other actuarial factors such as changes in benefit structure of costsharing provisions. Please note that the above discussion was meant to offer general instructional guidance on the rescaling of relative values and inflation adjustment. Given that no two analytic or actuarial applications are exactly alike, and given the potentially major impact such a process may have on the management or financial applications within your organization, it is essential that you seek and follow advice from experienced statistical or actuarial specialists before finalizing the general processes described above. 50

56 Table 2: Relative Concurrent PMPY Weights ACG ACG Label Relative Weight 0100 Acute Minor, Age Acute Minor, Age 2 to Acute Minor, Age > Acute Major Likely to Recur, w/o Allergies Likely to Recur, with Allergies Asthma Chronic Medical, Unstable Chronic Medical, Stable Chronic Specialty Eye/Dental Chronic Specialty, Unstable Psychosocial, w/o Psych Unstable Psychosocial, with Psych Unstable, w/o Psych Stable Psychosocial, with Psych Unstable, w/ Psych Stable Preventive/Administrative Pregnancy: 0-1 ADGs Pregnancy: 0-1 ADGs, delivered Pregnancy: 0-1 ADGs, not delivered Pregnancy: 2-3 ADGs, no Major ADGs Pregnancy: 2-3 ADGs, no Major ADGs, delivered Pregnancy: 2-3 ADGs, no Major ADGs, not delivered Pregnancy: 2-3 ADGs, 1+ Major ADGs Pregnancy: 2-3 ADGs, 1+ Major ADGs, delivered Pregnancy: 2-3 ADGs, 1+ Major ADGs, not delivered Pregnancy: 4-5 ADGs, no Major ADGs Pregnancy: 4-5 ADGs, no Major ADGs, delivered Pregnancy: 4-5 ADGs, no Major ADGs, not delivered Pregnancy: 4-5 ADGs, 1+ Major ADGs Pregnancy: 4-5 ADGs, 1+ Major ADGs, delivered Pregnancy: 4-5 ADGs, 1+ Major ADGs, not delivered Pregnancy: 6+ ADGs, no Major ADGs Pregnancy: 6+ ADGs, no Major ADGs, delivered Pregnancy: 6+ ADGs, no Major ADGs, not delivered Pregnancy: 6+ ADGs, 1+ Major ADGs Pregnancy: 6+ ADGs, 1+ Major ADGs, delivered Pregnancy: 6+ ADGs, 1+ Major ADGs, not delivered Acute Minor and Acute Major Acute Minor and Likely to Recur, Age Acute Minor and Likely to Recur, Age 2 to Acute Minor and Likely to Recur, Age > 5, w/o Allergy Acute Minor and Likely to Recur, Age > 5, with Allergy

57 ACG ACG Label Relative Weight 2300 Acute Minor and Chronic Medical: Stable Acute Minor and Eye/Dental Acute Minor and Psychosocial, w/o Psych Unstable Acute Minor and Psychosocial, with Psych Unstable, w/o Psych Stable Acute Minor and Psychosocial, with Psych Unstable and Psych Stable Acute Minor and Likely to Recur Acute Minor/Acute Major/Likely to Recur, Age Acute Minor/Acute Major/Likely to Recur, Age 2 to Acute Minor/Acute Major/Likely to Recur, Age 6 to Acute Minor/Acute Major/Likely to Recur, Age > 11, w/o Allergy Acute Minor/Acute Major/Likely to Recur, Age > 11, with Allergy Acute Minor/Likely to Recur/Eye & Dental Acute Minor/Likely to Recur/Psychosocial Acute Minor/Acute Major/Likely Recur/Eye & Dental Acute Minor/Acute Major/Likely Recur/Psychosocial Other ADG Combinations, Age < Other ADG Combinations, Males Age 18 to Other ADG Combinations, Females Age 18 to Other ADG Combinations, Age > Other ADG Combinations, Age < 18, no Major ADGs Other ADG Combinations, Age < 18, 1+ Major ADGs Other ADG Combinations, Age 18 to 44, no Major ADGs Other ADG Combinations, Age 18 to 44, 1+ Major ADGs Other ADG Combinations, Age 18 to 44, 2+ Major ADGs Other ADG Combinations, Age > 44, no Major ADGs Other ADG Combinations, Age > 44, 1+ Major ADGs Other ADG Combinations, Age > 44, 2+ Major ADGs Other ADG Combinations, Age < 6, no Major ADGs Other ADG Combinations, Age < 6, 1+ Major ADGs Other ADG Combinations, Age 6 to 17, no Major ADGs Other ADG Combinations, Age 6 to 17, 1+ Major ADGs Other ADG Combinations, Males, Age 18 to 34, no Major ADGs Other ADG Combinations, Males, Age 18 to 34, 1+ Major ADGs Other ADG Combinations, Males, Age 18 to 34, 2+ Major ADGs Other ADG Combinations, Females, Age 18 to 34, no Major ADGs Other ADG Combinations, Females, Age 18 to 34, 1+ Major ADGs Other ADG Combinations, Females, Age 18 to 34, 2+ Major ADGs Other ADG Combinations, Age > 34, 0-1 Major ADGs Other ADG Combinations, Age > 34, 2 Major ADGs Other ADG Combinations, Age > 34, 3 Major ADGs Other ADG Combinations, Age > 34, 4+ Major ADGs Other ADG Combinations, Age 1 to 17, no Major ADGs Other ADG Combinations, Age 1 to 17, 1 Major ADGs Other ADG Combinations, Age 1 to 17, 2 Major ADGs

58 ACG ACG Label Relative Weight Other ADG Combinations, Age > 17, 0-1 Major ADGs Other ADG Combinations, Age > 17, 2 Major ADGs Other ADG Combinations, Age > 17, 3 Major ADGs Other ADG Combinations, Age > 17, 4+ Major ADGs No Diagnosis or Only Unclassified Diagnosis & Non-Users (1 input file) No Diagnosis or Only Unclassified Diagnosis (2 input files) Non-Users (2 input files) Infants: 0-5 ADGs, no Major ADGs Infants: 0-5 ADGs, no Major ADGs, low birth weight Infants: 0-5 ADGs, no Major ADGs, normal birth weight Infants: 0-5 ADGs, 1+ Major ADGs Infants: 0-5 ADGs, 1+ Major ADGs, low birth weight Infants: 0-5 ADGs, 1+ Major ADGs, normal birth weight Infants: 6+ ADGs, no Major ADGs Infants: 6+ ADGs, no Major ADGs, low birth weight Infants: 6+ ADGs, no Major ADGs, normal birth weight Infants: 6+ ADGs, 1+ Major ADGs Infants: 6+ ADGs, 1+ Major ADGs, low birth weight Infants: 6+ ADGs, 1+ Major ADGs, normal birth weight Invalid Age or Date of Birth

59

60 Section 5 Selecting the Right Tool from the Expanding ACG Tool Box The preceding sections of this Release Document and the many sections of the comprehensive Version 5.0 Documentation and Application Manual offer significant levels of detail on each of the alternative applications of the various ACG measures. But even with this large mount of material to help guide your way, we recognize that a simple overview is needed to suggest which tool you should select from the Version 6.0 ACG tool box for each application within your organization. Targeted at both new and old users alike, this section offers a quick overview of the myriad ACG applications and suggests how the various components of the ACG tool box might be combined to maximize their usefulness to you. In a succinct fashion, this section also attempts to summarize some material that is presented elsewhere in our documentation. Where possible, linkages to more detailed discussion are offered to readers. a. One Size Does Not Fit All For over a decade the Johns Hopkins ACG risk-adjustment/case-mix methodology has been applied by many hundreds of users to meet an extremely diverse range of health care management and organizational needs. The ACG System represents a suite of tools that have been used to support basic and complex applications in finance, administration, care delivery, and evaluative research. These applications have been both real-time (concurrent) and forwardlooking (prospective). They may involve simple spreadsheet calculations or complex multivariable statistical models. No other risk adjustment method has been used for so many purposes in so many places, while at the same time showing such high levels of quantitative and qualitative success. The flexibility offered by our tool box means that we recognize that one size does not fit all. This also means that a bit of custom tailoring may be needed to get the best fit within your organization. The current ACG release represents a major expansion of the array of tools available within the Johns Hopkins ACG technology. As described earlier in this document, the acgpm model included in this software release produces two new predictive measures: a probability score indicating the likelihood that an individual will be a member of your very high risk cohort next year, and a Predicted Resource Index (PRI) that reflects the likely amounts of 55

61 resource use by persons next year relative to the other individuals in your population. In addition to these two brand new prospective measures of risk, this release also provides you with the option of using software-supplied concurrent weights for ACG cells (using the same reference population we used to develop acgpm) and provides a fixed set of ACG-based RUB categories. With more power and functionality come more options for users to navigate. The purpose of this section is to provide some help in making decisions about using the ACG System to support your individual requirements. b. Describing a Population s Health The ACG System is designed as a tool for understanding and explaining population health. The System s various diagnosis-based risk assessment markers provide a useful means for comparing the morbidity of different subpopulation of interest to you. Simple descriptive analyses like those shown in the following sample tables compare the distribution of morbidity across selected populations groupings. These are offered as models for how you may wish to apply our system to describe the morbidity characteristics of those cared for by your organization. Table 1: Comparison of ADG Distribution Across Two Enrollee Groups ADG Description Total Group 1 Group 2 1 Time Limited: Minor 14.7% 14.8% 14.4% 2 Time Limited: Minor -Primary Infections 32.2% 33.2% 27.4% 3 Time Limited: Major 5.5% 4.0% 12.3% 4 Time Limited: Major-Primary Infections 6.1% 5.1% 10.6% 5 Allergies 3.6% 3.6% 3.3% 6 Asthma 4.4% 4.2% 5.0% 7 Likely to Recur: Discrete 8.6% 6.6% 17.2% 8 Likely to Recur: Discrete-Infections 20.7% 22.0% 14.9% 9 Likely to Recur: Progressive 2.0% 0.8% 7.7% 10 Chronic Medical: Stable 12.9% 7.4% 37.1% 11 Chronic Medical: UnStable 8.6% 4.0% 28.8% 12 Chronic Specialty: Stable-Ortho 0.9% 0.5% 2.8% 13 Chronic Specialty: Stable-ENT 0.7% 0.6% 1.4% 14 Chronic Specialty: Stable-Eye 2.6% 2.0% 5.3% 15 No Longer in Use 0.0% 0.0% 0.0% 16 Chronic Specialty: UnStable-Ortho 0.8% 0.4% 2.4% 56

62 ADG Description Total Group 1 Group 2 continued 17 Chronic Specialty: UnStable-ENT 0.0% 0.0% 0.1% 18 Chronic Specialty: UnStable-Eye 1.6% 0.8% 5.2% 19 No Longer in Use 0.0% 0.0% 0.0% 20 Dermatologic 4.5% 4.4% 5.0% 21 Injuries/Adverse Effects: Minor 10.8% 10.2% 13.7% 22 Injuries/Adverse Effects: Major 9.3% 8.1% 14.3% 23 Psychosocial: Time Limited, Minor 3.5% 3.0% 5.5% 24 Psychosoc:Recur or Persist: Stable 9.8% 7.4% 20.3% 25 Psychosoc:Recur or Persist: UnStable 5.8% 2.5% 20.1% 26 Signs/Symptoms: Minor 16.9% 15.3% 24.4% 27 Signs/Symptoms: Uncertain 17.5% 14.1% 32.3% 28 Signs/Symptoms: Major 14.8% 11.6% 28.9% 29 Discretionary 5.8% 4.8% 10.4% 30 See and Reassure 1.8% 1.3% 3.8% 31 Prevention/Administrative 43.5% 46.7% 29.5% 32 Malignancy 1.0% 0.3% 4.0% 33 Pregnancy 2.2% 2.6% 0.3% 34 Dental 1.4% 1.4% 1.7% Table 1 illustrates how ADGs, the building blocks of the ACG system, can quickly demonstrate differences in types of morbidity categories across sub-groupings within your organization. In this example, the case-mix profile of Group 2 tends to be more complex than that of Group 1, with the prevalence of the chronic medical and psychosocial ADGs being especially high. An advantage of ADGs is they can quickly identify clinically meaningful morbidity trends that may be obscured at the disease-specific or relative morbidity index levels. As discussed in the Overview of New Features in ACG Release 6.0 section of this document, the ACG software will automatically assign a six-level (Low to High) simplified morbidity category we term RUBs (for Resource Utilization Bands). The six RUBs are formed by combining the ACG mutually exclusive cells that measure overall morbidity burden. Utilizing the Release 6.0 RUB categories, Table 2 demonstrates how a simple RUBbased analysis highlights differences in the distribution of morbidity of the Group 1 and Group 2 exemplary sub-populations. Confirming the impression drawn from Table 1, the Group 2 population clusters in the bands associated with higher overall morbidity burdens. 57

63 Table 2: Percentage Distribution of Two Sub-groups, by RUB Categories RUB Category Total Group 1 Group Non-users Healthy Users Low Morbidity Moderate High Very High As discussed earlier in this document, through use of disease-specific EDCs a standardized morbidity ratio report is now included as part of the standard ACG print file. 3 Based on the Major subheadings of Expanded Diagnosis Clusters, this report presents MEDC level disease prevalence of a sub-population of interest after taking into account the age and gender mix of the group relative to the underlying population. Thus, this report will assist users in isolating statistically significant (demographically adjusted) disease category differences within a sub-population of interest. The diagnostic/morbidity distribution reports outlined here should be useful for many clinically oriented applications within your organization. These could include population clinical needs assessments and targeting where disease management or outreach programs might be developed. 3 See Refinements to the Johns Hopkins Expanded Diagnosis Clusters (EDCs) in the Version 6.0 Release Notes and Chapter 13, Dino-Clusters: The Johns Hopkins Expanded Diagnosis Clusters (EDCs) in the Version 5.0 Documentation and Application Manual for additional details on interpreting this table and how to generate this table on other desired population groups. 58

64 Table 3. Observed to Expected Standardized Morbidity Ratio (SMR) by MEDC Population: AW Number of persons= Observed Age/Gender Approximate Prevalence Expected 95 percent per 1000 per 1000 confidence interval Major EDC Population Population SMR (low) (high) Administrative Allergy Cardiovascular Dental Ears, Nose, Throat Endocrine Eye Female Reproductive Gastrointestinal/Hepatic General Signs and Symptoms General Surgery Genetic Genito-urinary Hematologic Infections Malignancies Musculoskeletal Neurologic Nutrition Psychosocial Reconstructive Renal Respiratory Rheumatologic Skin Toxic Effects Unassigned c. Profiling Resource Use One of the most popular uses of the ACG software is to set risk-adjusted resource consumption norms for sub-groups of patients/members within an organization. These norms are compared to actual resource use in order to profile provider efficiency and to help suggest where over-use and under-use may be a problem. Profiling applications are very amenable to simple actuarial cell strategies for risk adjustment. Most ACG users apply the ACG mutually exclusive cells for this purpose while others have chosen to combine ACGs and use RUBs for these applications. The simpler RUB method is sometimes selected when the population s numbers are small or when the need to communicate the inner-workings of the methods to a wide audience of providers is critical. If a user has historical claims data (or other similar data sources), it is generally preferable to calculate expected resource use values for each ACG (or RUB) for each resource measure of interest (e.g., total cost, hospital use, specialist referrals, pharmacy) based on actual 59

65 patterns of practice within your organization. If such data are unavailable or inadequate, then the relative weights supplied as part of this release can be used as a proxy. 4 Table 4 presents a summary of the most common profiling statistics: 5 1) the actual to group average resource use (unadjusted efficiency ratio); 2) the expected to plan average (the case-mix index or morbidity factor); and 3) the actual to expected average resource use (efficiency ratios). The first is a measure of how the profiling group compares to the average population. The second, the morbidity factor provides an indication of how sick the profiling population is compared to the average population. The last statistic, the observed to expected ratio ( O/E Ratio ) provides an indication of how many health care resources were consumed by this group compared to how many resources they would have consumed had they utilized the average resource use of the population based on their case-mix characteristics. All three of these statistics are expressed as relative values with the average or normative value centered at 1.0. Scores greater than 1.0 indicate higher than average whereas those less than 1.0 indicate lower than average. Tests of statistical significance can be developed to assess outlier status. Clearly the use of risk adjustment provides a dramatically different basis for assessing the performance of the three profiled sites. Profiles such as those summarized above are a useful tool for evaluating performance and allocating resources within a wide range of ACG users. The most common applications include: financial exchange between MCO and providers, assessing provider efficiency, resource planning, evaluating access to care, and detecting fraud, waste, and abuse. In the absence of local resource data that can be used to determine local weights, the concurrent weights available within the ACG System can be used to develop summary measures of case-mix for comparisons between groups. 4 See Chapter 8, Calibrating ACGs for Intended Use: ACG Weights, Resource Utilization Bands, and Carve-outs, of the Version 5.0 Documentation and Application Manual and the Using the Available Relative Value Weights section of this document for a detailed discussion of relevant methodologic issues related to weight calculation. 5 For further details see Chapter 12, ACG Risk Adjustment and Provider Profiling, in the Version 5.0 Documentation and Application Manual. 60

66 Table 4. Comparison of Observed to Expected Visits and Calculation of Three Profiling Ratios SITE A SITE B SITE C 1) Actual Visits per Person (Observed) 2) Plan Average ) Actual to Group Average* (Unadjusted Efficiency Ratio) 4) Number of Expected Visits** ) Expected to Plan Average*** (Morbidity Factor) 6) Observed to Expected Ratio**** (Adjusted Efficiency Ratio) * Row 1 divided by Row 2 ** Expected based on ACG characteristics at each site *** Row 4 divided by Row 2 **** Row 1 divided by Row 4 d. Disease Management and Case Management Applications As discussed previously, concurrent ACG / RUB morbidity information can be combined with EDCs to control for morbidity differences across a given disease-specific group of interest (e.g., diabetics enrolled in a disease management program). EDCs will be useful in portraying the disease characteristics of a population of interest. Within disease management programs, if significant differences in expected resource consumption exist across the morbidity sub-classes, this analytic approach should be quite useful in better targeting interventions towards sub-groups at higher risk. Along these lines, two new tables (see Tables 5 and 6 for examples) are now part of the print file that the ACG software produces. Each row of these tables represents persons falling into EDC (or MEDC) disease-specific categories; the columns array these individuals into RUB co-morbidity categories according to their ACG assignment. Table 5 presents the percentage distribution for a series of selected EDCs across the five RUB categories. Table 6 presents the expected relative resource use within each RUB. This table illustrates co-morbidity s profound influence on resource use within individual disease groups. The ACG-based RUBs do a very good job of explaining variations in resource use within specific diseases. The ACG software 61

67 automatically generates these reports based on nationally representative weights, but such tables are likely to become even more useful when calibrated to local cost and practice patterns. 6 Table 5: Percentage Distribution of Each Co-morbidity Level within an EDC (Samples) RUB-1 RUB-2 RUB-3 RUB-4 RUB-5 EDC Very Low Low Average High Very High ADM01:General medical exam ADM02:Surgical aftercare ADM03:Transplant status ADM04:Complications of mechanical ALL01:Allergic reactions ALL03:Allergic rhinitis ALL04:Asthma, w/o status asthmati ALL05:Asthma, with status asthmat ALL06:Disorders of the immune sys CAR01:Cardiovascular signs and sy CAR03:Ischemic heart disease (exc CAR04:Congenital heart disease CAR05:Congestive heart failure CAR06:Cardiac valve disorders CAR07:Cardiomyopathy CAR08:Heart murmur CAR09:Cardiac arrhythmia CAR10:Generalized atherosclerosis CAR11:Disorders of lipoid metabol CAR12:Acute myocardial infarction CAR13:Cardiac arrest, shock CAR14:Hypertension, w/o major com CAR15:Hypertension, with major co As discussed elsewhere, EDCs are very useful for many purposes. If users so choose, they can develop their own reports, and the EDCs that define the rows in Tables 5 and 6 could be replaced by episodes of illness categories that an organization may obtain from other sources. ACG-based RUBs are equally effective in explaining variations in resource use within episodes of care. 6 See Chapter 13, Dino-Clusters: The Johns Hopkins Expanded Diagnosis Clusters (EDCs), in the Version 5.0 Documentation and Application Guide for detailed instructions on how to create these tables calibrated to your own data. 62

68 Table 6: Estimated Concurrent Resource Use by RUB by MEDC (Samples) RUB-1 RUB-2 RUB-3 RUB-4 RUB-5 EDC Very Low Low Average High Very High ADM01:General medical exam ADM02:Surgical aftercare ADM03:Transplant status ADM04:Complications of mechanical ALL01:Allergic reactions ALL03:Allergic rhinitis ALL04:Asthma, w/o status asthmati ALL05:Asthma, with status asthmat ALL06:Disorders of the immune sys CAR01:Cardiovascular signs and sy CAR03:Ischemic heart disease (exc CAR04:Congenital heart disease CAR05:Congestive heart failure CAR06:Cardiac valve disorders CAR07:Cardiomyopathy CAR08:Heart murmur CAR09:Cardiac arrhythmia CAR10:Generalized atherosclerosis CAR11:Disorders of lipoid metabol CAR12:Acute myocardial infarction CAR13:Cardiac arrest, shock CAR14:Hypertension, w/o major com CAR15:Hypertension, with major co e. High-risk Case Identification for Case Management The various components of the acgpm the new ACG predictive modeling module represent a real advance for users wishing to establish or augment care management programs within their organization. Furthermore, existing ACG measures have many applications in this domain as well. There are many ways to adapt the ACG suite of tools in the pursuit of improved patient care. This sub-section provides a summary and overview of some of the recommended approaches that an organization may wish to consider in the care-management and quality improvement (QI) domains. As discussed in some detail earlier in this document, the new acgpm risk measurement tools provide information at the individual patient level to help identify persons who potentially would be well served by special attention from the organization s care management infrastructure. This high-risk case identification process could be used to target a person for interventions such as a referral to a case-manager, special communication with the patient s physician, structured disease management programs, or educational outreach. As part of the new acgpm module the Release 6.0 software now includes a report that provides a disease-specific (based on selected individual and aggregated EDCs) distribution of 63

69 risk probability scores and average expected resource use for different risk cohorts. This latter report, shown here as Table 7, will be potentially be useful in helping to frame a strategy for targeting various risk cohorts within disease management programs. Table 7. Number of Cases and acgpm Predicted Relative Resource Use, by Risk Probability Thresholds for Selected Chronic Conditions Number of Cases Predicted Relative Resource Use Probability Score Category Probability Score Category Disease Category (EDC) Total < Arthritis 17, Asthma 27, Diabetes 16,991 1, Hypertension 50,122 2,064 1, Ischemic Heart Disease 9, Congestive Heart Failure 1, Hyperlipidemia 31,240 1, Low Back Pain 61,980 1, Depression 10, Chronic Renal Failure COPD 6, The acgpm s probability score was fine-tuned to identify persons who will likely be the ones in your organization who would most benefit from special attention. To capitalize on this new method, an organization will want to develop periodic reports of members with high acgpm scores who also meet other organizational criteria, such as enrollment with certain providers, falling into certain eligibility categories, residing in certain geographic areas, or meeting previous patterns of utilization. After these other stratifiers are taken into consideration as appropriate, a case finding report should list all in-scope individuals arrayed from highest to lowest based on the overall acgpm high-risk probability score within your organization. In addition to running the report automatically generated by the software, users are encouraged to develop their own individual risk summary reports on each potential case over a certain threshold (say the top 1% of individuals). This target group can be winnowed further by 64

70 case managers on the basis of various sources of information available from the ACG software and elsewhere. These additional data might include primary care provider information, service history, history of prior inclusion in care management programs, and results from any ongoing surveys (such as health-risk appraisals). Please see the The ACG Predictive Model: Helping to Manage Care for Persons at Risk for High Future Cost section of this document for a comprehensive discussion of the new acgpm module and its applications. f. Capitation, Actuarial Underwriting, and Rate Setting ACGs have been successful for so long because they do a good job at capturing the complex interplay of co-morbidities that explain the impact of case-mix on resource use. This factor clearly distinguishes ACGs from other risk-adjustment strategies that treat diseases individually, as if they each have a completely independent effect on health. The validity of the ACG view of illness has been borne out through the successful application of this riskadjustment strategy across a range of applications for over a decade. i. Using ACGs as Actuarial Cells The ACG System has made it possible to accomplish risk adjustment with fairly simple and straightforward analytic strategies. For example, ACGs can readily be used as actuarial cells, which have long been the primary actuarial method for both capitation rate setting and underwriting. Actuarial cells represent a fixed number of discrete categories into which individuals are placed based on their expected use of resources. ACGs are very well suited for assigning individuals into these types of actuarial cells. There are a number of advantages associated with using an actuarial cell-based approach to risk adjustment for capitation and underwriting: Simplicity. Once the population has been classified into around 100 ACG cells, it is possible to risk-adjust the population by using a spreadsheet. Some users have chosen to simplify this approach even further by collapsing the ACGs into smaller homogeneous groupings, resource utilization bands (RUBs). Even when grouped into RUBs, studies indicate that ACGs retain much of their explanatory power. 65

71 Less prone to gaming or manipulation. Particularly in applications involving rate setting, there could be incentives to game risk-adjustment strategies to increase payment. Unlike some other disease-specific risk adjusters, aggressive efforts to capture additional diagnostic codes on the part of providers will have a more limited impact on ACG assignments. Where code creep associated with general increases in completeness and accuracy of coding exists, the simplicity of the ACG system makes it very easy to identify this trend and to implement appropriate action, such as re-calibration of weights. Stability. The conceptual elegance and underlying simplicity of ACGs have made the system very stable over long periods. The underlying clinical truth captured by ACGs does not change dramatically with each new data set and each new application. Ease of making local calibrations. It is very easy to recalibrate ACG-based actuarial cells to reflect local differences in patterns of practice, benefit structure, and provider fees. Especially for capitation and rate-setting tasks, we encourage users to calibrate the ACG output to reflect the unique nature of the local cost structure. The same simplicity that makes it possible to risk-adjust using a spreadsheet makes it equally possible to accomplish recalibration using the same types of simple tools. The ultimate testimony to the value of ACGs used as the basis of actuarial cells is the fact that for almost a decade they have been used to facilitate the exchange of many billions of dollars within numerous private and public health plans in both the United States and Canada. ii. ACGs in Multivariate Models Multivariate regression for risk adjustment has been used for many years by some of the more sophisticated users of ACGs. If additional risk descriptors are available beyond diagnosis, age, and sex, this approach has the potential for improved predictive models. The strength of regression-based strategies is the ease with which additional risk factor information can be incorporated and thereby introduce better control for the effects of case-mix. This ease is also a potential drawback since regression may introduce some assumptions and statistical pitfalls that can be troublesome without seasoned analytical support. Their inherent 66

72 complexity makes them difficult to calibrate to local cost patterns, and regression models are also potentially easier to game because more factors can be manipulated. Finally, while it is possible to introduce a wide range of variables that improve the model s explanatory power, this explanatory power is often confined to the data set and time period on which the model is based. The model s results may end up differing significantly from year to year depending on the interrelations of the myriad risk factors that have been included, a phenomenon referred to as overfitting. To address some of these analytic challenges, the acgpm provided as part of this release represents a regression-based strategy that can be applied for prospective financial applications. As discussed, one acgpm output, the probability score, has been specifically tailored for case and disease-management applications. The other acgpm output, the predicted resource index (PRI), assigns a relative value that can be readily converted to dollars. This PRI output is most relevant for financial risk-adjustment applications and can be considered a substitute for ACG cells for prospective rate setting or payment. One important caveat is worth noting here. Prior pharmacy cost has been made an optional risk factor variable in the new acgpm. Although it is useful for calculating the most accurate predictions for future costs, we do NOT recommend that models using the optional pharmacy cost predictor be applied to capitation rate setting. Instead, we suggest that the acgpm model relying only on ICD input variables be used for such a purpose. We take this position for the same reason we believe that episode groupers that rely on procedure codes (such as CPT) and Rx-groupers based on use of specific medications (as defined by NDC codes) should not be used for rate-setting purposes or efficiency profiles. Risk factor variables of this type, which are directly defined by the providers clinical practices, are potentially intertwined with patterns of over use or under use. Risk-adjusted rates based on these factors may, in a circular manner (termed endogoneity by the economists), lead to setting rates that are inappropriate either too high or too low. Moreover, when risk factors are determined by such drug use (or procedural) delivery patterns, providers who practice efficiently could potentially be penalized for their efficiency. This circularity issue is not a major concern when only diagnostic information (not linked to specific types or settings of service) is used as the main source of information on risk factors. 67

73 iii. To Regress or Not to Regress: That Is the Question One of the key decision points in using risk adjustment for financial applications is whether to use a simple actuarial cell approach or a more complex multivariate model. If you have been applying ACG-based actuarial cells successfully for some time, there may be little incentive to change since ACGs alone remain a highly effective case-mix adjustment tool. "If it ain t broke, don t fix it." If you are just starting out in your selection of methods, you will need to balance the stability and ease of use of ACG-based actuarial cells against the potentially enhanced ability to explain variations in resource use by applying regression modeling strategies. If you have access to additional well-validated risk factor data and if you have previous experience using regression models within your organization, then you should consider using regression. In regression strategies, ACGs, ADGs, and EDCs remain valuable as distinct risk factors to be supplemented by additional data. Although EDCs are useful for identifying individuals with specific high impact diseases, it is important to note that they do not account for burden of co-morbidity as do ACGs. Therefore, we do not generally recommend that EDCs be used as the only means of controlling for case-mix in regression analysis. The ultimate choice of risk-adjustment approach depends on the specific application, and it is prudent to compare both actuarial cell and regression approaches over a span of several years before making a final decision. For multivariate models, the R-squared statistic is often used as an indicator of performance. In fact, extreme caution is recommended when evaluating models based on R-squared values. The R-squared statistic is very sensitive to outlier individuals. Aside from considering measures of model fit, such as the R-squared value, you should consider whether the results are reasonably stable over time. It is also advisable to simulate the degree to which each approach results in over- or under-payment to key segments of your population. iv. Concurrent versus Prospective Applications The time frame used for most rate setting and other financial analyses is a prospective or predictive one. That is, this year s diagnostic information is used to determine risk factors and expected resource consumption in some future period. Thus the weights associated with each risk factor are calibrated to that future period. But this is not the only temporal approach that organizations can use for rate setting. Some ACG users have implemented concurrent rating processes for financial exchanges. In such cases, this year s expected resource use among the 68

74 benchmark population is attached to each ACG cell as a relative value rather than next year s resource use. While we do encourage experienced actuaries and financial analysts to learn more about the advantages and challenges of these innovative concurrent approaches, we do not recommend that organizations apply concurrent approaches to payment without first simulating the impact that these methods might have on the rate-setting process. A real-world example of a concurrent approach to rate setting is one being implemented in Minnesota Medicaid where plan-level payments are based on concurrent ACG-adjusted profiles of the plan. Under this scenario, payment to a health plan is the same for each individual enrollee within a particular plan; however, the amount paid is case-mix adjusted by the plan s overall morbidity burden (relative to an average, across the population, of 1.0). This approach assumes that the morbidity burden of large groups (i.e., any individual health plan) is fairly stable and that the group s overall morbidity does not change much by the addition/exit of any one individual. For additional discussion on this and related issues related to risk adjustment as applied to financial exchanges, we encourage readers to review our chapter incorporated into Charles Wrightson s recently published book Financial Strategy for Managed Care Organizations: Rate Setting, Risk Adjustment, and Competitive Advantage (see for ordering details). Our chapter is available online at Readers are also encouraged to review the ACG bibliography at that site for a variety of articles illustrating ACGs used for capitation. g. In Closing As part of our ongoing commitment to furthering the international state-of-the-art of riskadjustment methodology and supporting ACG users worldwide, we will continue to perform evaluation, research, and development. We will look forward to sharing the results of this work with our user-base via white papers, our web site, peer-reviewed articles, and in-person presentations. After you have carefully reviewed the documentation supplied with this software release, we would welcome your inquiries on any topic of relevance to your use of ACGs within your organization. (Contact us at askacg@jhsph.edu.) We thank you for using ACGs and for 69

75 helping us to work toward meeting the Johns Hopkins University s ultimate goal of improving the quality, efficiency, and equity of health care across the United States and around the globe. 70

76 Section 6 Installation and Usage Intended for old and new users alike, this chapter is written for the programmer/analyst who will be using the software. This chapter begins with an overview of the technical enhancements new to Release 6.0 and is required reading for all users of the software. The remainder of the chapter is divided into the following main sections: installing the software, using the software, required components of the input files and how to pass data to the software, and output files. Readers are referred to the relevant sections of the ACG Version 6.0 Release Notes and the Version 5.0 Documentation and Application Manual for additional details on calculating or using ACG weights, discussions of the built-in reporting features and how to customize these reports, and other relevant criteria for implementing ACG technologies within your organization. a. For Users Already Familiar with the ACG Software Although the functionality of Release 6.0 has been greatly enhanced, these improvements only slightly affect the installation and use of the software. This section highlights the few important changes so that current ACG System users will be able to get up and running quickly. A full set of current installation and usage instructions follows so that new (and current) users will have all the information that they need readily at hand. What is needed to take advantage of ACG System enhancements involves two new control card keywords and four names that have been added to the OUTREC control card. The control card keywords allow for additional (and optional) data to be passed to the grouper from the input file; the four names accommodate additional (and optional) output fields produced by the software. 71

77 i. New Control Card Keywords Two new record layout control card keywords permit the input of two new data fields: 1. POP is used for a group membership identifier such as underwriting group or PCP assignment; and 2. PCOST is used for pharmacy cost data. The group identifier, called POP, serves as a stratifier for producing age/sex-adjusted prevalence rates and standard morbidity ratio reports using the Johns Hopkins Expanded Diagnosis Clusters (EDCs) typology. 7 The optional pharmacy cost information, called PCOST, is a useful adjunct that improves performance of the acgpm (see description of the HRCI card below). If pharmacy cost information is available, we recommend its inclusion. ii. New OUTREC Features Four new names have been added to the control card that controls the output file. OUTREC additions are as follows: 1. CWT, to output a set of fixed concurrent ACG weights based on our nationally representative database (written to the output file as ##.###). 2. RUB, to output six resource use levels (Resource Utilization Bands, or RUBs for short) expressed as an ordinal number with values between zero and five as follows: 0 = nonusers 1 = healthy users 2 = low morbidity 3 = moderate morbidity 4 = high morbidity 5 = very high morbidity 3. HOS, to output a Boolean indicator (e.g., a value of zero or one) for the presence of conditions likely to lead to a hospitalization. 4. HRCI, to output the four scores for predictive modeling (written to the output file as ###.###) in the following order: 7 Note: Please refer to Section 2 of this Addenda Additions and Refinements to the Expanded Diagnosis Cluster (EDC) Methodology (including new reports) and ICD-9 Coding Updates and Chapter 13, Dino-Clusters : The Johns Hopkins Expanded Diagnosis Clusters (EDCs), of the Version 5.0 Documentation and Application Manual for technical specifications necessary to customize these tables to your application. 72

78 a. Total cost predicted resource index an estimate for Year 2 total expenditures (including pharmacy charges) expressed as a relative weight; b. Pharmacy cost predicted resource index an estimate for Year 2 predicted pharmacy expenditures also expressed as a relative weight; c. Probability of being in the high total cost cohort a probability score with values between zero and one, indicating the likelihood that a person will have high cost in the subsequent time period; and d. Probability of being in the high pharmacy cost cohort a probability score indicating the likelihood that a person will have high pharmacy cost in the subsequent time period. Also note that users may include the POP and PCOST control cards in the OUTREC statement if they are interested in having this information appear in the output (as well as the input) file produced by the software. iii. Processing Speed and Space Requirements Unlike prior versions of the software that allocated memory dynamically for each individual, the expanded reporting features and modeling components of Release 6.0 require more processing time. Processing time depends on a variety of factors including but not limited to CPU speed, disk read-write speed, available memory and disk space, as well as the size of the input files. As a general guideline the software can be expected to take two to three times longer than previous versions. For groups that are fewer than 100,000 the increase in processing speed is negligible, but for very large populations the software may take substantially longer than prior versions to process similarly sized input files. Note that processing time may be reduced if duplicate diagnoses cards are removed from the input data stream. The expanded reports and model-building component of Release 6.0 now also necessitate the use of temporary files. While the ultimate output file(s) produced by the software will be similar in size to those produced by prior versions of the software (with the addition of space needed to accommodate new OUTREC fields), sufficient disk space must be available for the writing of temporary files. At a minimum the software will need 1 MB of available disk space for general overhead. On top of this, an estimated additional 600 bytes per person for member identifier, age, and sex fields plus an additional 51 bytes per each unique ICD per person are 73

79 required. To simplify things somewhat, as a general guideline Release 6.0 requires temporary disk space approximately five to six times the size of the input file. Because of the modeling component of the software, there is now for the first time a limit to the maximum number of people that can be processed in one run. We estimate the upper limit to be approximately 3.5 million individuals. The limit is imposed not on the maximum number of individuals, but rather on the maximum allowable temporary file size, which has been set at 2 GB in most operating systems. If you have membership in excess of 3.5 million, please contact your software distributor for further guidance on how best to divide your input file into smaller population subgroupings. iv. The Print File The number of fixed reports has increased in this release, and so the size of the software s print file can be expected to be larger. The size of the print file will be at least partially determined by the number of levels provided in the population stratifier control card (e.g., POP discussed in the preceding paragraphs) because these reports are generated for each individual stratifier. If hundreds of stratifiers are used, the software will generate hundreds of pages of printed reports. Even if hundreds of stratifiers are included, the print file can be easily managed by loading it into text processing software and globally changing the print font to Courier 8. With this simple adjustment, reports should be legible and page breaks should appear in logical places. Extracts of this file can easily be extracted by using the cut and paste feature to separate files as desired. v. Other The input file or files no longer need to be sorted. Consequently, the NOSORT card is not needed (and is no longer provided). The age reference date must be provided on the DOB card and must be of the form CCYYMMDD. Previously this information could be left blank, and the system date would be used, or it could be provided as YYMMDD. The format of the EDC output file has changed, so current users must change their programs to read data from these files. Although the EDCs are still written from columns one through five, the unique member ID does not begin until column 12 so as to allow for future expanded EDC categories. 74

80 b. Installing the Software The ACG grouper software is supplied on a diskette or CD. Installation involves copying all files from this disk to your hard disk; other installation steps depend on the platform. If you are upgrading from an earlier release and want to maintain the prior release for comparison purposes, then rename the old executable file before installing the new version. This can be done by using the DOS rename command or the UNIX mv command. Consult your system documentation for more information on renaming files. Alternatively, you could create separate directories for each version of the software, although you should be sure to use the appropriate version. The version of the software used at any one time is listed in the heading of the print file output. i. UNIX Platforms The file supplied on the disk is named ACGGROUP. Copy this file to your UNIX partition. You may then need to change the file mode to allow execute permission. This is usually done with the chmod command, e.g., chmod +x acggroup. The software can then be invoked by merely entering acggroup controlcardfilename. By pressing the break key combination for your system (e.g., Ctrl-Break or Ctrl-C), the system can be halted. That completes the UNIX installation. ii. PC Platforms The file supplied on the disk is named ACGGROUP.EXE. Copy this file to your hard disk for the PC-DOS installation. 75

81 c. Using the Software Figure 1. Opening Screen for ACG Version 6.0. After copying the file to the appropriate subdirectory, to access the software (acggroup.exe) simply double click 8 on the file listing or icon using Windows Explorer or File Manager. Figure 1 is a screen capture of what should appear. The software prompts for the controlcardfilename. The controlcardfilename specifies the location of a file containing a series of ACG control cards that communicate to the software a) the location of the input and output files and what fields to read from the input file, b) which ACG branching options are to be used, and c) what ACG-based risk assessment variables are to be written to the output file(s). (See Table 1. More detail to follow.) After you type in the full filename and hit return, the software executes and a series of progress bars will appear at the bottom of the screen indicating the percentage of data processed. When 100% of the data is processed, the window automatically closes, and files created by the software will reside in the appropriate directories (as indicated by the control card). If a problem is found with any of the control cards, then an error message is written to the screen and the user must press enter to halt execution of the program (at which time viewing the print file may help users to better ascertain where the problem resides and/or at what point the software stopped executing). 8 Alternatively, type the following from the command line: ACGGROUP CONTROLCARDFILENAME (where CONTROLCARDFILENAME is the filename that contains all of the control cards) 76