- Email: [email protected]
Development of a Chronic Disease Indicator Score Using a Veterans Affairs Medical Center Medication Database
J Clin Epidemiol Vol. 52, No. 6, pp. 551–557, 1999 Copyright © 1999 Elsevier Science Inc. All rights reserved.
0895-4356/99/$–see front matter PII S0895-4356(99)00029-3
Development of a Chronic Disease Indicator Score Using a Veterans Affairs Medical Center Medication Database Daniel C. Malone,* Sarah J. Billups, Robert J. Valuck, and Barry L. Carter for the IMROVE Investigators University of Colorado Health Sciences Center, School of Pharmacy, Denver, Colorado ABSTRACT. Objective: Develop a chronic disease index that approximates the number of chronic diseases a patient has using a medication database. Methods: An expert panel determined whether specific medication classes could be indicative of a chronic disease. Those classes identified were incorporated into a computer program and then used to screen the medication records of 246 randomly selected patients to estimate the number of chronic diseases present in each patient. This number was designated as the chronic disease index (CDI). The CDI was then validated against chart review. The CDI and a measure of disease severity, the chronic disease score (CDS) also were compared. The sensitivity and specificity of the computer program was analyzed for seven common chronic diseases. Results: The expert panel designated 54 drug classes containing medications used to treat chronic diseases. The CDI correlated moderately with the number of chronic diseases found via chart review (r 5 0.65; P 5 0.001) and highly with the CDS (r 5 0.81; P 5 0.001). The index predicted the presence of three common diseases with a sensitivity of $75%, and of six common diseases with a specificity of $75%. Conclusions: The CDI correlates moderately well with the actual number of chronic disease states present. This tool may be useful for researchers when trying to identify patients with specific diseases and also for risk adjustment. J CLIN EPIDEMIOL 52;6:551–557, 1999. © 1999 Elsevier Science Inc. KEY WORDS. Chronic disease, clinical pharmacy information systems, health status indicators, prescriptions, drug, severity of illness index
BACKGROUND Assessment of disease prevalence in large patient populations can be a formidable task. Direct methods of assessing disease prevalence include the use of computerized codes such as the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM), diagnostic-related groups (DRGs), or the Clinical Classifications for Health Policy Research (CCBHR) system developed by the Agency for Health Care Policy and Research (AHCPR) [1]. These coding schemes are often limited by the primary purpose for which the codes were developed. The accuracy of the coding is somewhat suspect if reimbursement varies depending upon which ICD-9-CM code is recorded for a given patient visit [2]. Completeness of diagnostic codes comes into question if the recording system limits the number of coding slots [2]. Additionally, coding may reflect only the conditions treated during a specific hospital or *Address for correspondence: Daniel C. Malone, Ph.D., University of Colorado Health Sciences Center, 4200 East Ninth Avenue, C238, Denver, CO 80262. Tel: (303) 315-3868; Fax: (303) 315-4630; E-mail: [email protected] Accepted for publication on 12 February 1999.
clinic visit, while other important chronic conditions may not be recorded. Prescription drug databases can be used as an indirect method for assessing disease prevalence. These databases have the advantage of generally being complete, accurate, and reliable. Accuracy and completeness are generally high because prescription records are subject to legal requirements. Also, financial reimbursement is often dependent upon timely, accurate recording of prescription information. Finally, unlike many segments of health care, prescription records are stored with similar or identical data elements, including at least the drug, strength, quantity dispensed, and days supply. In the Veterans Affairs health care system many patients get all or nearly all of their medications from a Veterans Affairs Medical Center (VAMC), so a complete medication profile is generally available. In managed care settings, the presence and severity of disease has been estimated using prescription databases [3,4]. These studies illustrate the usefulness of electronic prescription data. Although these studies have demonstrated the relationship between certain medications and health care expenditures, the classes of medications were too limited to be used without modification. Therefore, we
D. C. Malone et al.
552
set out to develop a scale that would utilize all available prescription data. This index was originally developed for the IMROVE study as one component of a tool to detect patients at high risk for ethical medication-related problems [5]. The purpose of the IMROVE study was to assess the impact of clinical pharmacists on the outcomes of patients at high-risk for medication-related problems. Six criteria were used to identify this high-risk population, one of which related to the number of chronic diseases present [6]. OBJECTIVE The purpose of this research was to develop an index that would estimate the presence and number of chronic diseases based upon data in a medication database from ambulatory patients. Secondly, we evaluated the tool’s ability to identify specific chronic diseases. METHODS There are various proprietary and non-proprietary methods to classify medications (AHFS®, First Data Bank®, Medispan®). For this study, we chose the VA classification system which is also used by the United States Pharmacopoeia Drug Information Handbook.1 This system uses a fivealpha-numeric character code to classify medications. The first two characters are alphabetic characters that classify the item into broad drug categories (such as “AB” for antimicrobials and “CV” for cardiovascular agents). The next three digits numerically classify items into specific drug classes (such as antivirals, penicillins, beta-blockers, etc.). Each drug is classified under one and only one code; thus no cross-referencing occurs. The 1995 system lists a total of 423 drug and non-drug classes to describe specific products. For purposes of this study, we excluded all non-drug products (bandages, syringes, needles), large-volume parenteral products, medications that are not administered to outpatients (anesthetics, thrombolytics, diagnostic agents, etc), and over-the-counter medications. Thus, only the remaining 263 drug classes, which represent ethical drugs, were included. We assembled an expert panel of clinicians to assess the likelihood that each drug class (of the 263 analyzed) indicates the presence of a chronic disease. The expert panel consisted of three individual raters (one internal medicine physician and two clinical pharmacists who practice in primary care) and one consensus group. The consensus group consisted of five clinical pharmacists, each of whom works in a different ambulatory care setting. The group’s responses were reached by mutual consensus and were treated as one
Programming code is available from the authors that links the VA classification system to the American Hospital Formulary System (AHFS). 1
rater. Thus, there were four independent raters on the expert panel. Their responses were all weighted equally. We asked each of these four raters to answer the following question for each drug class: “Given that you know a patient is taking a drug from this class, what is the likelihood that the patient has a chronic disease?” The clinicians could respond with one of three answers for each drug class: (1) it probably indicates the presence of a chronic disease, (2) it possibly indicates the presence of a chronic disease, or (3) it is unlikely to indicate the presence of a chronic disease. If the rater denoted that the drug class either (1) probably, or (2) possibly indicates the presence of a chronic disease, he or she was asked to list the specific chronic disease(s) that drug class might indicate. Clinicians were asked only to assess the probability of the presence of a chronic condition, not to grade disease severity. This study is only concerned with those medication classes indicative of a chronic disease. A drug class was considered indicative of a chronic disease if at least three of the four raters rated it as “it probably indicates a chronic disease.” Drug classes considered indicative of a chronic disease were incorporated into a SAS® (Cary, NC) computer program. Considering some diseases often require treatment with two or more medication classes, medication classes have been grouped together for purposes of calculating the CDI, thus overestimation of the number of chronic disease states by the CDI is avoided. For example, oral hypoglycemic agents are grouped together with insulin, so a patient taking a medication from both these classes would only receive a score of 1. Specific groupings are illustrated in Table 1. The total score for the CDI is intended to estimate the number of chronic diseases present. The CDI does not indicate the severity of the disease for a given patient, only that a particular disease is likely to be present. Validation To ascertain the validity of those drug classes identified as representative of chronic disease, we conducted a review for a random sample of 246 patient medical records from a VAMC in the Rocky Mountain region. This VAMC is divided into two primary care firms for ambulatory patients: A and B. Data for this analysis utilized data from patients in firm A. On April 1, 1997, 2803 patients were assigned to firm A. The number of chronic diseases per patient documented in the medical records was determined by two independent reviewers and compared to the number predicted by the CDI. A second validation analysis correlated the CDI with Von Korff et al.’s CDS [3] using the Pearson correlation coefficient. Finally, an analysis was conducted to determine the sensitivity and specificity of the program to identify chronic conditions. For each disease present in at least 25 patients (per chart review), the chronic disease(s) predicted by the index were compared against the actual chronic diseases listed in the patient’s medical chart.
Chronic Disease Index
553
TABLE 1. Drug classes rated by expert panel of indicating presence of a chronic disease
Drug class code
Drug class
Rater agreement (%)
Disease states indicated Alcoholism Chronic UTI HIV, Parkinson’s, Herpes Malignant neoplasm Malignant neoplasm Malignant neoplasm Malignant neoplasm Pheochromocytoma, hypertension Afib, heart valve, coagulation d/o Anemia, neutropenia Migraine Epilepsy, affective d/o, pain Parkinsonism Depression, pain, migraine Depression Depression Schizophrenia, psychosis, hiccups Schizophrenia, psychosis Bipolar d/o, depression Obesity, depression, ADD, narcolepsy CHF, Afib, SVT Hypertension, CHF, SVT, migraines, tremor, angina Hypertenson, BPH, pain Hypertension, angina, SVT Angina Arrhythmias Hyperlipidemia Hypertension Hypertension, migraines, CHF, tobacco dependence Hypertension, fluid retention, CHF Fluid retention, CHF, renal disease Hypertension, fluid retention, ascites, CHF, hyperaldosteronism Hypertension, CHF, DM Hypertension, CHF, DM Pancreatic insufficiency RA, asthma, COPD Corticoadrenal insufficiency DM DM Hypothyroidism Thyrotoxicosis, Grave’s Organ transplant Renal failure RA, psoriasis, pemphigus Glaucoma Glaucoma Glaucoma Allergic rhinitis Asthma, COPD Asthma, COPD Asthma, COPD Asthma, COPD Asthma Hemorrhoids, noninfectious gastroenteritis/colitis, Crohns, IBD
AD100 AM550 AM800 AN100 AN200 AN500 AN900 AU200 BL100 BL400 CN105 CN400 CN500 CN601 CN602 CN609 CN701 CN709 CN750 CN802 CV050 CV100 CV150 CV200 CV250 CV300 CV350 CV400 CV490 CV701 CV702 CV704
Alcohol deterrents Methenamine salts Antivirals Antineoplastics, alkylating agents Antineoplastics, antibiotics Antineoplastics, hormones Antineoplastics, other Sympathomimetics Anticoagulants Blood formation products Antimigraine agents Anticonvulsants Antiparkinson’s agents Tricyclic antidepressants MAOIs Antidepressants, other Phenothiazine/related antipsychotics Antipsychotics, other Lithium salts Amphetamine-like stimulants Digitalis glycosides Beta-blockers/related Alpha-blockers/related Calcium channel blockers Antianginals Antiarrhythmics Antilipemic agents Antihypertensive combinations Antihypertensives, other Thiazides/related diuretics Loop diuretics Potassium sparing/combo diuretics
100 100 75 75 75 75 75 75 75 100 75 100 100 75 100 75 100 100 100 100 100 75 75 100 100 100 75 75 75 75 75 75
CV800 CV805 GA500 HS051 HS052 HS501 HS502 HS851 HS852 IM600 IR200 MS106 OP101 OP102 OP103 RE101 RE102 RE103 RE104 RE105 RE109 RS100
ACE inhibitor Angiotensin II antagonists Digestants Glucocorticoids Mineralocorticoids Insulin Oral hypoglycemic agents Thyroid supplements Antithyroid agents Immune suppressants Peritoneal dialysis solutions Gold compounds, antirheumatic Beta-blockers, topical ophthalmic Miotics, topical ophthalmic Adrenergics, topical ophthalmic Anti-inflammatories, inhalation Bronchodilators, sympathomimetic Inhaled Bronchodilators, sympathomimetic oral Bronchodilators, xanthine-derivatives Bronchodilators, anticholinergics Antiasthma, other anti-inflammatories, rectal
100 100 75 75 75 100 100 100 75 75 100 100 75 75 75 75 100 100 100 100 75 75
D. C. Malone et al.
554
Positive predictive values and negative predictive values of the CDI were also calculated for these seven disease states. RESULTS Selection of Drug Classes Indicative of a Chronic Disease Fifty-four (21%) of the 263 total drug classes evaluated were categorized as probably indicative of a chronic disease. Table 2 summarizes the rate of agreement between the four reviewers as to whether a drug probably does or probably does not indicate the presence of a chronic disease. Twenty-five of these 54 classes had 100% agreement by the raters, while the remaining 29 classes had 75% agreement. Table 1 lists the 54 drug classes that were classified as indicative of a chronic disease, the level of agreement that led to this classification, and the diseases the expert panel believed each drug class indicated. Diseases listed by the expert panel for the 54 drug classes in Table 1 were considered detectable by the CDI. The CDI can theoretically detect 52 different chronic diseases.
There were several chronic diseases noted upon chart review that were “undetectable” by the CDI. Three chronic diseases were detected by chart review in at least 10% (25/ 246) of the patients, but were considered undetectable by the CDI. These include degenerative joint disease (57 patients; 23%), gastro-esophageal reflux (33 patients; 13%), and peptic ulcer disease (24 patients; 10%). The majority of the expert panel did not consider medications used to treat these diseases indicative of chronic disease, and thus they were not included in the CDI. We reviewed these patients manually to see what proportion could have been detected by the CDI if additional medication classes were added. The results of this analysis are presented in Table 4. Greater than two-thirds of patients with gastro-esophageal reflux or peptic ulcer disease were receiving pharmacological treatment for their condition, but less than half of the patients with degenerative joint disease were receiving drug therapy that was dispensed from the VAMC. Therefore, it is possible that adding the medications used to treat these conditions would improve the sensitivity for the CDL although this may be at the risk of reduced specificity.
Validation: Chart Review First, we compared the total number of chronic diseases found via chart review against the total number predicted via the CDI. The distribution of CDI scores and the actual distribution of chronic diseases as determined by chart review are presented in Figure 1. The shapes of these two histograms are similar with one exception: the CDI predicted that patients have zero chronic diseases with much greater frequency than was found by chart review. The correlation between these two methods was 0.65 (P , 0.0001), indicating that a moderate relationship exists between the number of chronic diseases actually present (according to chart review) and the number predicted by the CDI (Table 3). The CDI predicts 43% of the variance in number of chronic diseases actually present.
Validation: CDI and CDS The second method used to validate the CDI was to compare it to a CDS developed by Von Korff et al. [3,7]. The objective of the CDS is to assess health status using automated pharmacy data. The CDS has been evaluated for its ability to predict total ambulatory visits, hospitalization, and mortality, and was found to correlate positively with all these measures [4]. The CDS weights certain drug classes higher than others, while the CDI weights all indicative drug classes equally. Thus the range of possible scores for the CDS is somewhat wider than that for the CDI. The sample of patients used for this validation confirmed this difference. CDS scores ranged from 0 to 15, while CDI scores ranged from 0 to 12
TABLE 2. Agreement between raters
The drug class probably indicates a chronic disease state The drug class probably does not indicate a chronic disease state
Percent agreement between raters
Number of drug classes meeting this level of agreement
Percentage
100 75
25 29
10 11
100 75
110 39
42 15
Drug classes with ,75% agreement as to whether or not they indicate a chronic disease state
60
23
263
100a
Total aPercentages
were rounded, thus arithmetic total is .100%.
Chronic Disease Index
555
FIGURE 1. Frequency of chronic diseases within a Veterans Affairs Medical Center as determined by medical chart and computer algorithm (n 5 246)(solid bars, medical chart; open bars, CDI.).
(see Table 3). Values for the two instruments correlated quite strongly, with a correlation coefficient of 0.81 (P , 0.0001). CDI as an Indicator of Specific Disease States Our secondary objective was to assess whether the CDI could be used to identify patients with specific disease states. To accomplish this, we compared the index’s findings against the diseases actually documented in the medical record for conditions that were documented in at least 10% of the patients. Seven chronic diseases met this criteria and were used to assess the sensitivity and specificity of the CDI. Results of these analyses are presented in Table 5. The CDI predicted three of the seven disease states with a sensitivity of greater than or equal to 75% and six of the seven with a specificity of greater than or equal to 75%. As noted above, there were several diseases that occurred frequently in this population that were not detected by the computer algorithm. DISCUSSION The CDI correlates well with the construct it was designed to measure, the actual number of chronic diseases in a population of patients. Because the index relies on a patient
obtaining a prescription for a specific medication to suggest the presence of disease, some chronic diseases, such as those not treated with drug therapy, are missed. As a result, the index has greater specificity and moderate sensitivity, causing it to sometimes underestimate the number of diseases a patient actually has. The same pattern of high specificity and moderate sensitivity was noted when the index was applied to predict specific chronic diseases. Several measures have been developed to estimate the presence and/or severity of acute and chronic disease. However, most of these measures are intended to quantify risk for short-term outcomes for hospitalized patients [8]. Fewer measures have been developed to quantify risk for longer periods of a time (such as 1 year) and can therefore be applied to ambulatory patient populations [3,9,10]. All three of these “longer-term” measures of disease severity utilize diagnostic information as a critical element. Two of these rely upon diagnostic codes, [9,10] which carry the disadvantage of possible reimbursement bias and are also at risk for being incomplete or even unavailable in many settings [2]. The third measure utilizes pharmacy data from a large HMO setting to create the CDS [3]. In this study, the CDS correlated highly with the CDI although the two instruments measure somewhat different constructs. Both these measures offer the advantage of a reliable database that is
TABLE 3. Correlation of CDI with actual number of chronic diseases per chart review,
and with CDS
CDI score Chart review CDS Score *P , 0.0001.
Mean 6 SD
Median (range)
Pearson correlation coefficient
2.96 6 2.67 3.27 6 1.81 4.00 6 3.40
2 (0–12) 3 (0–11) 3 (0–15)
— 0.65* 0.81*
D. C. Malone et al.
556
TABLE 4. Chronic conditions frequently ($10%) detected in medical chart but not by the CDI
Disease state
What medication, if any, was the patient taking for this condition? (n)a
Number of patients affected
Degenerative joint disease
52
Gastro-esophageal reflux disease
33
Peptic ulcer disease
24
Acetaminophen (12) Acetaminophen/codeine (4) Ibuprofen (4) Other NSAIDb (4) Acetaminophen/oxycodone (1) No medication (27) Histamine2-receptor antagonist (13) Proton-pump inhibitor (10) Antacids (2) No medication (10) Histamine2-receptor antagonist (7) Proton-pump inhibitor (9) Antacids (2) No medication (7)
aObservations
must have included at least 120 days worth of medication to be counted. anti-inflammatory drug. Note: Some patients received .1 medication for the indication of interest. bNon-steroidal
readily available in many health care organizations. The CDI may be preferable in certain circumstances because it utilizes a nonproprietary drug classification system and includes a larger number of drug classes. However, the CDS has been validated against several important outcome measures, including ambulatory care office visits, ambulatory costs, and mortality. Such validation studies are currently underway for the CDI. Application of the CDI Probably the most widespread application of measures of disease status is for risk adjustment. Risk adjustment of large patient populations is useful to predict many health-carerelated outcomes. Outcomes often of interest include (1) morbidity and mortality; (2) complications of disease or medical care; (3) physical and psychosocial functional sta-
tus; (4) quality of life; (5) costs of medical care; (6) use of specific health-care services; and (7) satisfaction with medical care [8]. The CDI could be used alone or in combination with other measures for risk adjustment and thus to predict some or perhaps all of the outcomes listed above. The index has already been applied in a research setting where it is one of six criteria being used to identify patients at high risk for drug-related problems (poor disease control, poor adherence, adverse drug reactions, and drug–drug interactions) in the IMPROVE Study [5]. Limitations When evaluating the CDI’s usefulness in predicting the prevalence of chronic disease, certain limitations must be taken into consideration. The validation of this index was conducted at a VAMC. The population of patients within
TABLE 5. CDI as an indicator of selected disease statesa
Disease state Alcoholism Coronary artery disease Chronic obstructive pulmonary disease Depression Diabetes mellitus Dyslipidemia Hypertension aGold
Number of patients predicted to have disease (CDI)
Number of patients with actual disease (medical chart)
Sensitivity
5 98
33 58
41 57 43 44 147
26 33 47 61 139
Specificity
Positive predictive value
Negative predictive value
15% 83%
100% 73%
1.0 0.49
0.88 0.93
73% 70% 91% 61% 86%
90% 84% 100% 88% 75%
0.46 0.40 1.0 0.84 0.82
0.97 0.95 0.98 0.88 0.81
standard was the medical chart. Diseases occurring in 10% or more of the population according to the medical chart.
Chronic Disease Index
the VA is probably much different than those within managed care organizations. Further validation is needed before this tool should be adopted in non-VAMC settings. The method by which chronic diseases were identified is open to criticism. This study had the expert panel examine each medication class and then determine if all medications in this class were used to treat chronic conditions. Another approach would be to identify those chronic diseases that are routinely treated with medications. Differences may arise depending upon the method chosen. Another limitation is that this tool only evaluates those diseases for which a prescription medication is prescribed and dispensed. Thus a diabetic patient who is adequately controlled with diet therapy will not be detected using this method. Although this limitation reduces the sensitivity of the instrument, it also ensures that only conditions considered severe enough (by the physician and the patient) to require pharmacotherapy will be included. Another limitation is that certain medication classes that may indicate a chronic disease were not rated as such by the expert panel. As a result, commonly used medications such as anti-gout agents, acetaminophen, histamine2-antagonists, and proton-pump inhibitors were not considered indicative of chronic disease despite the finding that many patients in this study (approximately 25%) had chronic conditions which are usually treated with these medications. A problem in using the CDI may arise when patients obtain medications outside of the health care system being studied or purchase over-the-counter items. It also is important to note that some drug classes are used to treat more than one disease state, thus the CDI is likely to underestimate the number of diseases present when these drug classes are used. This phenomenon makes predictions of specific chronic diseases nonspecific. Another source of error when interpreting the results of this study is the potential for incomplete information in the medical record. Although the medical record was used as the gold standard for this study, the “problem lists” in the records sometimes contained incomplete and/or unclear information. CONCLUSIONS The results of this study suggest that certain drug classes can be used as indicators of chronic disease. The CDI had a moderately strong correlation with the actual number of chronic diseases present, and a high correlation with other automated measures. The CDI tended to underestimate the actual number of disease states present, suggesting that the computer program is conservative in its estimation of
557
chronic disease prevalence. The CDI demonstrated the ability to detect seven specific diseases with moderate sensitivity and good-to-excellent specificity. Thus, this tool may be useful to estimate prevalence of these diseases in a large population. The authors acknowledge: Sune F. Isaksen and Jacob Jonassen for conducting the medical chart review; other IMPOVE co-investigators: Charles D. Sintek, M.S. and Debra J. Barnette, Pharm.D.; the guidance and support from the Veterans’ Affairs, Pharmacia & Upjohn Steering Committee: Jannet Carmichael, Pharm.D., Donald L. Kendzierski, Pharm.D., Julio Lopez, Pharm.D., Andy Muniz, M.D., Art Schuna, M.S., David Solomon, Charles E. Weber, R.Ph., J. Rod Barnes, David K. Burnham, David W. Johnson, R.Ph., George I. Wortman, R.Ph.; the expert panel members: Melissa Foss, Pharm.D., Larry A. Bourg, M.D., J. Mark Ruscin, Pharm.D., and Joseph J. Saseen, Pharm.D.; and anonymous reviewers for their insightful comments. This study is supported in part by Pharmacia & Upjohn and the ACCP Merck and Company Pharmacoeconomics Fellowship.
References 1. Elixhauser A, McCarthey E. Clinical Classifications for Health Policy Research, Version 2: Hospital Inpatient Statistics (AHCPR Publication No. 96-0017, 1996). Healthcare Cost and Utilization Project (HCUIP-3) Research Note 1, Agency for Health Care Policy and Research. Rockville, MD: U.S. Public Health Service; 1996. 2. Iezzoni LI. Assessing quality using administrative data. Ann Intern Med 1997; 127(8): 666–674. 3. Von Korff M, Wagner EH, Saunders K. A chronic disease score from automated pharmacy data. J Clin Epidemiol 1992; 45: 197–203. 4. Johnson RE, Hornbrook MC, Nichols GA. Replicating the chronic disease score (CDS) from automated pharmacy data. J Clin Epidemiol 1994; 47(10): 1191–1199. 5. Carter BL, Malone DC, Valuck RJ, Barnette DJ, Sintek CD, Billups SJ. The IMROVE study: Background and study design. Am J Health Syst Pharm 1998; 55: 62–67. 6. Koecheler JA, Abramowitz PW, Swim SE, Daniels CE. Indicators for the selection of ambulatory patients who warrant pharmacist monitoring. Am J Hosp Pharm 1989; 46: 729– 732. 7. Clark DO, VonKorff M, Saunders K, Baluch WM, Simon GE. A chronic disease score with empirically derived weights. Med Care 1995; 33: 783–795. 8. Iezzoni LI. Risk Adjustment for Measuring Healthcare Outcomes, 2d Edition. Chicago, IL: Health Administration Press; 1997. 9. Ash A, Porell F, Gruenberg L, Sawitz E, Beiser A. Adjusting Medicare capitation payments using prior hospitalization data. Health Care Financing Review 1989; 10(4): 17–29. 10. Starfield B, Weiner J, Mumford W, Steinwachs D. Ambulatory care groups: A categorization of diagnoses for research and management. Health Services Research 1991; 26(l): 53–74.
0895-4356/99/$–see front matter PII S0895-4356(99)00029-3
Development of a Chronic Disease Indicator Score Using a Veterans Affairs Medical Center Medication Database Daniel C. Malone,* Sarah J. Billups, Robert J. Valuck, and Barry L. Carter for the IMROVE Investigators University of Colorado Health Sciences Center, School of Pharmacy, Denver, Colorado ABSTRACT. Objective: Develop a chronic disease index that approximates the number of chronic diseases a patient has using a medication database. Methods: An expert panel determined whether specific medication classes could be indicative of a chronic disease. Those classes identified were incorporated into a computer program and then used to screen the medication records of 246 randomly selected patients to estimate the number of chronic diseases present in each patient. This number was designated as the chronic disease index (CDI). The CDI was then validated against chart review. The CDI and a measure of disease severity, the chronic disease score (CDS) also were compared. The sensitivity and specificity of the computer program was analyzed for seven common chronic diseases. Results: The expert panel designated 54 drug classes containing medications used to treat chronic diseases. The CDI correlated moderately with the number of chronic diseases found via chart review (r 5 0.65; P 5 0.001) and highly with the CDS (r 5 0.81; P 5 0.001). The index predicted the presence of three common diseases with a sensitivity of $75%, and of six common diseases with a specificity of $75%. Conclusions: The CDI correlates moderately well with the actual number of chronic disease states present. This tool may be useful for researchers when trying to identify patients with specific diseases and also for risk adjustment. J CLIN EPIDEMIOL 52;6:551–557, 1999. © 1999 Elsevier Science Inc. KEY WORDS. Chronic disease, clinical pharmacy information systems, health status indicators, prescriptions, drug, severity of illness index
BACKGROUND Assessment of disease prevalence in large patient populations can be a formidable task. Direct methods of assessing disease prevalence include the use of computerized codes such as the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM), diagnostic-related groups (DRGs), or the Clinical Classifications for Health Policy Research (CCBHR) system developed by the Agency for Health Care Policy and Research (AHCPR) [1]. These coding schemes are often limited by the primary purpose for which the codes were developed. The accuracy of the coding is somewhat suspect if reimbursement varies depending upon which ICD-9-CM code is recorded for a given patient visit [2]. Completeness of diagnostic codes comes into question if the recording system limits the number of coding slots [2]. Additionally, coding may reflect only the conditions treated during a specific hospital or *Address for correspondence: Daniel C. Malone, Ph.D., University of Colorado Health Sciences Center, 4200 East Ninth Avenue, C238, Denver, CO 80262. Tel: (303) 315-3868; Fax: (303) 315-4630; E-mail: [email protected] Accepted for publication on 12 February 1999.
clinic visit, while other important chronic conditions may not be recorded. Prescription drug databases can be used as an indirect method for assessing disease prevalence. These databases have the advantage of generally being complete, accurate, and reliable. Accuracy and completeness are generally high because prescription records are subject to legal requirements. Also, financial reimbursement is often dependent upon timely, accurate recording of prescription information. Finally, unlike many segments of health care, prescription records are stored with similar or identical data elements, including at least the drug, strength, quantity dispensed, and days supply. In the Veterans Affairs health care system many patients get all or nearly all of their medications from a Veterans Affairs Medical Center (VAMC), so a complete medication profile is generally available. In managed care settings, the presence and severity of disease has been estimated using prescription databases [3,4]. These studies illustrate the usefulness of electronic prescription data. Although these studies have demonstrated the relationship between certain medications and health care expenditures, the classes of medications were too limited to be used without modification. Therefore, we
D. C. Malone et al.
552
set out to develop a scale that would utilize all available prescription data. This index was originally developed for the IMROVE study as one component of a tool to detect patients at high risk for ethical medication-related problems [5]. The purpose of the IMROVE study was to assess the impact of clinical pharmacists on the outcomes of patients at high-risk for medication-related problems. Six criteria were used to identify this high-risk population, one of which related to the number of chronic diseases present [6]. OBJECTIVE The purpose of this research was to develop an index that would estimate the presence and number of chronic diseases based upon data in a medication database from ambulatory patients. Secondly, we evaluated the tool’s ability to identify specific chronic diseases. METHODS There are various proprietary and non-proprietary methods to classify medications (AHFS®, First Data Bank®, Medispan®). For this study, we chose the VA classification system which is also used by the United States Pharmacopoeia Drug Information Handbook.1 This system uses a fivealpha-numeric character code to classify medications. The first two characters are alphabetic characters that classify the item into broad drug categories (such as “AB” for antimicrobials and “CV” for cardiovascular agents). The next three digits numerically classify items into specific drug classes (such as antivirals, penicillins, beta-blockers, etc.). Each drug is classified under one and only one code; thus no cross-referencing occurs. The 1995 system lists a total of 423 drug and non-drug classes to describe specific products. For purposes of this study, we excluded all non-drug products (bandages, syringes, needles), large-volume parenteral products, medications that are not administered to outpatients (anesthetics, thrombolytics, diagnostic agents, etc), and over-the-counter medications. Thus, only the remaining 263 drug classes, which represent ethical drugs, were included. We assembled an expert panel of clinicians to assess the likelihood that each drug class (of the 263 analyzed) indicates the presence of a chronic disease. The expert panel consisted of three individual raters (one internal medicine physician and two clinical pharmacists who practice in primary care) and one consensus group. The consensus group consisted of five clinical pharmacists, each of whom works in a different ambulatory care setting. The group’s responses were reached by mutual consensus and were treated as one
Programming code is available from the authors that links the VA classification system to the American Hospital Formulary System (AHFS). 1
rater. Thus, there were four independent raters on the expert panel. Their responses were all weighted equally. We asked each of these four raters to answer the following question for each drug class: “Given that you know a patient is taking a drug from this class, what is the likelihood that the patient has a chronic disease?” The clinicians could respond with one of three answers for each drug class: (1) it probably indicates the presence of a chronic disease, (2) it possibly indicates the presence of a chronic disease, or (3) it is unlikely to indicate the presence of a chronic disease. If the rater denoted that the drug class either (1) probably, or (2) possibly indicates the presence of a chronic disease, he or she was asked to list the specific chronic disease(s) that drug class might indicate. Clinicians were asked only to assess the probability of the presence of a chronic condition, not to grade disease severity. This study is only concerned with those medication classes indicative of a chronic disease. A drug class was considered indicative of a chronic disease if at least three of the four raters rated it as “it probably indicates a chronic disease.” Drug classes considered indicative of a chronic disease were incorporated into a SAS® (Cary, NC) computer program. Considering some diseases often require treatment with two or more medication classes, medication classes have been grouped together for purposes of calculating the CDI, thus overestimation of the number of chronic disease states by the CDI is avoided. For example, oral hypoglycemic agents are grouped together with insulin, so a patient taking a medication from both these classes would only receive a score of 1. Specific groupings are illustrated in Table 1. The total score for the CDI is intended to estimate the number of chronic diseases present. The CDI does not indicate the severity of the disease for a given patient, only that a particular disease is likely to be present. Validation To ascertain the validity of those drug classes identified as representative of chronic disease, we conducted a review for a random sample of 246 patient medical records from a VAMC in the Rocky Mountain region. This VAMC is divided into two primary care firms for ambulatory patients: A and B. Data for this analysis utilized data from patients in firm A. On April 1, 1997, 2803 patients were assigned to firm A. The number of chronic diseases per patient documented in the medical records was determined by two independent reviewers and compared to the number predicted by the CDI. A second validation analysis correlated the CDI with Von Korff et al.’s CDS [3] using the Pearson correlation coefficient. Finally, an analysis was conducted to determine the sensitivity and specificity of the program to identify chronic conditions. For each disease present in at least 25 patients (per chart review), the chronic disease(s) predicted by the index were compared against the actual chronic diseases listed in the patient’s medical chart.
Chronic Disease Index
553
TABLE 1. Drug classes rated by expert panel of indicating presence of a chronic disease
Drug class code
Drug class
Rater agreement (%)
Disease states indicated Alcoholism Chronic UTI HIV, Parkinson’s, Herpes Malignant neoplasm Malignant neoplasm Malignant neoplasm Malignant neoplasm Pheochromocytoma, hypertension Afib, heart valve, coagulation d/o Anemia, neutropenia Migraine Epilepsy, affective d/o, pain Parkinsonism Depression, pain, migraine Depression Depression Schizophrenia, psychosis, hiccups Schizophrenia, psychosis Bipolar d/o, depression Obesity, depression, ADD, narcolepsy CHF, Afib, SVT Hypertension, CHF, SVT, migraines, tremor, angina Hypertenson, BPH, pain Hypertension, angina, SVT Angina Arrhythmias Hyperlipidemia Hypertension Hypertension, migraines, CHF, tobacco dependence Hypertension, fluid retention, CHF Fluid retention, CHF, renal disease Hypertension, fluid retention, ascites, CHF, hyperaldosteronism Hypertension, CHF, DM Hypertension, CHF, DM Pancreatic insufficiency RA, asthma, COPD Corticoadrenal insufficiency DM DM Hypothyroidism Thyrotoxicosis, Grave’s Organ transplant Renal failure RA, psoriasis, pemphigus Glaucoma Glaucoma Glaucoma Allergic rhinitis Asthma, COPD Asthma, COPD Asthma, COPD Asthma, COPD Asthma Hemorrhoids, noninfectious gastroenteritis/colitis, Crohns, IBD
AD100 AM550 AM800 AN100 AN200 AN500 AN900 AU200 BL100 BL400 CN105 CN400 CN500 CN601 CN602 CN609 CN701 CN709 CN750 CN802 CV050 CV100 CV150 CV200 CV250 CV300 CV350 CV400 CV490 CV701 CV702 CV704
Alcohol deterrents Methenamine salts Antivirals Antineoplastics, alkylating agents Antineoplastics, antibiotics Antineoplastics, hormones Antineoplastics, other Sympathomimetics Anticoagulants Blood formation products Antimigraine agents Anticonvulsants Antiparkinson’s agents Tricyclic antidepressants MAOIs Antidepressants, other Phenothiazine/related antipsychotics Antipsychotics, other Lithium salts Amphetamine-like stimulants Digitalis glycosides Beta-blockers/related Alpha-blockers/related Calcium channel blockers Antianginals Antiarrhythmics Antilipemic agents Antihypertensive combinations Antihypertensives, other Thiazides/related diuretics Loop diuretics Potassium sparing/combo diuretics
100 100 75 75 75 75 75 75 75 100 75 100 100 75 100 75 100 100 100 100 100 75 75 100 100 100 75 75 75 75 75 75
CV800 CV805 GA500 HS051 HS052 HS501 HS502 HS851 HS852 IM600 IR200 MS106 OP101 OP102 OP103 RE101 RE102 RE103 RE104 RE105 RE109 RS100
ACE inhibitor Angiotensin II antagonists Digestants Glucocorticoids Mineralocorticoids Insulin Oral hypoglycemic agents Thyroid supplements Antithyroid agents Immune suppressants Peritoneal dialysis solutions Gold compounds, antirheumatic Beta-blockers, topical ophthalmic Miotics, topical ophthalmic Adrenergics, topical ophthalmic Anti-inflammatories, inhalation Bronchodilators, sympathomimetic Inhaled Bronchodilators, sympathomimetic oral Bronchodilators, xanthine-derivatives Bronchodilators, anticholinergics Antiasthma, other anti-inflammatories, rectal
100 100 75 75 75 100 100 100 75 75 100 100 75 75 75 75 100 100 100 100 75 75
D. C. Malone et al.
554
Positive predictive values and negative predictive values of the CDI were also calculated for these seven disease states. RESULTS Selection of Drug Classes Indicative of a Chronic Disease Fifty-four (21%) of the 263 total drug classes evaluated were categorized as probably indicative of a chronic disease. Table 2 summarizes the rate of agreement between the four reviewers as to whether a drug probably does or probably does not indicate the presence of a chronic disease. Twenty-five of these 54 classes had 100% agreement by the raters, while the remaining 29 classes had 75% agreement. Table 1 lists the 54 drug classes that were classified as indicative of a chronic disease, the level of agreement that led to this classification, and the diseases the expert panel believed each drug class indicated. Diseases listed by the expert panel for the 54 drug classes in Table 1 were considered detectable by the CDI. The CDI can theoretically detect 52 different chronic diseases.
There were several chronic diseases noted upon chart review that were “undetectable” by the CDI. Three chronic diseases were detected by chart review in at least 10% (25/ 246) of the patients, but were considered undetectable by the CDI. These include degenerative joint disease (57 patients; 23%), gastro-esophageal reflux (33 patients; 13%), and peptic ulcer disease (24 patients; 10%). The majority of the expert panel did not consider medications used to treat these diseases indicative of chronic disease, and thus they were not included in the CDI. We reviewed these patients manually to see what proportion could have been detected by the CDI if additional medication classes were added. The results of this analysis are presented in Table 4. Greater than two-thirds of patients with gastro-esophageal reflux or peptic ulcer disease were receiving pharmacological treatment for their condition, but less than half of the patients with degenerative joint disease were receiving drug therapy that was dispensed from the VAMC. Therefore, it is possible that adding the medications used to treat these conditions would improve the sensitivity for the CDL although this may be at the risk of reduced specificity.
Validation: Chart Review First, we compared the total number of chronic diseases found via chart review against the total number predicted via the CDI. The distribution of CDI scores and the actual distribution of chronic diseases as determined by chart review are presented in Figure 1. The shapes of these two histograms are similar with one exception: the CDI predicted that patients have zero chronic diseases with much greater frequency than was found by chart review. The correlation between these two methods was 0.65 (P , 0.0001), indicating that a moderate relationship exists between the number of chronic diseases actually present (according to chart review) and the number predicted by the CDI (Table 3). The CDI predicts 43% of the variance in number of chronic diseases actually present.
Validation: CDI and CDS The second method used to validate the CDI was to compare it to a CDS developed by Von Korff et al. [3,7]. The objective of the CDS is to assess health status using automated pharmacy data. The CDS has been evaluated for its ability to predict total ambulatory visits, hospitalization, and mortality, and was found to correlate positively with all these measures [4]. The CDS weights certain drug classes higher than others, while the CDI weights all indicative drug classes equally. Thus the range of possible scores for the CDS is somewhat wider than that for the CDI. The sample of patients used for this validation confirmed this difference. CDS scores ranged from 0 to 15, while CDI scores ranged from 0 to 12
TABLE 2. Agreement between raters
The drug class probably indicates a chronic disease state The drug class probably does not indicate a chronic disease state
Percent agreement between raters
Number of drug classes meeting this level of agreement
Percentage
100 75
25 29
10 11
100 75
110 39
42 15
Drug classes with ,75% agreement as to whether or not they indicate a chronic disease state
60
23
263
100a
Total aPercentages
were rounded, thus arithmetic total is .100%.
Chronic Disease Index
555
FIGURE 1. Frequency of chronic diseases within a Veterans Affairs Medical Center as determined by medical chart and computer algorithm (n 5 246)(solid bars, medical chart; open bars, CDI.).
(see Table 3). Values for the two instruments correlated quite strongly, with a correlation coefficient of 0.81 (P , 0.0001). CDI as an Indicator of Specific Disease States Our secondary objective was to assess whether the CDI could be used to identify patients with specific disease states. To accomplish this, we compared the index’s findings against the diseases actually documented in the medical record for conditions that were documented in at least 10% of the patients. Seven chronic diseases met this criteria and were used to assess the sensitivity and specificity of the CDI. Results of these analyses are presented in Table 5. The CDI predicted three of the seven disease states with a sensitivity of greater than or equal to 75% and six of the seven with a specificity of greater than or equal to 75%. As noted above, there were several diseases that occurred frequently in this population that were not detected by the computer algorithm. DISCUSSION The CDI correlates well with the construct it was designed to measure, the actual number of chronic diseases in a population of patients. Because the index relies on a patient
obtaining a prescription for a specific medication to suggest the presence of disease, some chronic diseases, such as those not treated with drug therapy, are missed. As a result, the index has greater specificity and moderate sensitivity, causing it to sometimes underestimate the number of diseases a patient actually has. The same pattern of high specificity and moderate sensitivity was noted when the index was applied to predict specific chronic diseases. Several measures have been developed to estimate the presence and/or severity of acute and chronic disease. However, most of these measures are intended to quantify risk for short-term outcomes for hospitalized patients [8]. Fewer measures have been developed to quantify risk for longer periods of a time (such as 1 year) and can therefore be applied to ambulatory patient populations [3,9,10]. All three of these “longer-term” measures of disease severity utilize diagnostic information as a critical element. Two of these rely upon diagnostic codes, [9,10] which carry the disadvantage of possible reimbursement bias and are also at risk for being incomplete or even unavailable in many settings [2]. The third measure utilizes pharmacy data from a large HMO setting to create the CDS [3]. In this study, the CDS correlated highly with the CDI although the two instruments measure somewhat different constructs. Both these measures offer the advantage of a reliable database that is
TABLE 3. Correlation of CDI with actual number of chronic diseases per chart review,
and with CDS
CDI score Chart review CDS Score *P , 0.0001.
Mean 6 SD
Median (range)
Pearson correlation coefficient
2.96 6 2.67 3.27 6 1.81 4.00 6 3.40
2 (0–12) 3 (0–11) 3 (0–15)
— 0.65* 0.81*
D. C. Malone et al.
556
TABLE 4. Chronic conditions frequently ($10%) detected in medical chart but not by the CDI
Disease state
What medication, if any, was the patient taking for this condition? (n)a
Number of patients affected
Degenerative joint disease
52
Gastro-esophageal reflux disease
33
Peptic ulcer disease
24
Acetaminophen (12) Acetaminophen/codeine (4) Ibuprofen (4) Other NSAIDb (4) Acetaminophen/oxycodone (1) No medication (27) Histamine2-receptor antagonist (13) Proton-pump inhibitor (10) Antacids (2) No medication (10) Histamine2-receptor antagonist (7) Proton-pump inhibitor (9) Antacids (2) No medication (7)
aObservations
must have included at least 120 days worth of medication to be counted. anti-inflammatory drug. Note: Some patients received .1 medication for the indication of interest. bNon-steroidal
readily available in many health care organizations. The CDI may be preferable in certain circumstances because it utilizes a nonproprietary drug classification system and includes a larger number of drug classes. However, the CDS has been validated against several important outcome measures, including ambulatory care office visits, ambulatory costs, and mortality. Such validation studies are currently underway for the CDI. Application of the CDI Probably the most widespread application of measures of disease status is for risk adjustment. Risk adjustment of large patient populations is useful to predict many health-carerelated outcomes. Outcomes often of interest include (1) morbidity and mortality; (2) complications of disease or medical care; (3) physical and psychosocial functional sta-
tus; (4) quality of life; (5) costs of medical care; (6) use of specific health-care services; and (7) satisfaction with medical care [8]. The CDI could be used alone or in combination with other measures for risk adjustment and thus to predict some or perhaps all of the outcomes listed above. The index has already been applied in a research setting where it is one of six criteria being used to identify patients at high risk for drug-related problems (poor disease control, poor adherence, adverse drug reactions, and drug–drug interactions) in the IMPROVE Study [5]. Limitations When evaluating the CDI’s usefulness in predicting the prevalence of chronic disease, certain limitations must be taken into consideration. The validation of this index was conducted at a VAMC. The population of patients within
TABLE 5. CDI as an indicator of selected disease statesa
Disease state Alcoholism Coronary artery disease Chronic obstructive pulmonary disease Depression Diabetes mellitus Dyslipidemia Hypertension aGold
Number of patients predicted to have disease (CDI)
Number of patients with actual disease (medical chart)
Sensitivity
5 98
33 58
41 57 43 44 147
26 33 47 61 139
Specificity
Positive predictive value
Negative predictive value
15% 83%
100% 73%
1.0 0.49
0.88 0.93
73% 70% 91% 61% 86%
90% 84% 100% 88% 75%
0.46 0.40 1.0 0.84 0.82
0.97 0.95 0.98 0.88 0.81
standard was the medical chart. Diseases occurring in 10% or more of the population according to the medical chart.
Chronic Disease Index
the VA is probably much different than those within managed care organizations. Further validation is needed before this tool should be adopted in non-VAMC settings. The method by which chronic diseases were identified is open to criticism. This study had the expert panel examine each medication class and then determine if all medications in this class were used to treat chronic conditions. Another approach would be to identify those chronic diseases that are routinely treated with medications. Differences may arise depending upon the method chosen. Another limitation is that this tool only evaluates those diseases for which a prescription medication is prescribed and dispensed. Thus a diabetic patient who is adequately controlled with diet therapy will not be detected using this method. Although this limitation reduces the sensitivity of the instrument, it also ensures that only conditions considered severe enough (by the physician and the patient) to require pharmacotherapy will be included. Another limitation is that certain medication classes that may indicate a chronic disease were not rated as such by the expert panel. As a result, commonly used medications such as anti-gout agents, acetaminophen, histamine2-antagonists, and proton-pump inhibitors were not considered indicative of chronic disease despite the finding that many patients in this study (approximately 25%) had chronic conditions which are usually treated with these medications. A problem in using the CDI may arise when patients obtain medications outside of the health care system being studied or purchase over-the-counter items. It also is important to note that some drug classes are used to treat more than one disease state, thus the CDI is likely to underestimate the number of diseases present when these drug classes are used. This phenomenon makes predictions of specific chronic diseases nonspecific. Another source of error when interpreting the results of this study is the potential for incomplete information in the medical record. Although the medical record was used as the gold standard for this study, the “problem lists” in the records sometimes contained incomplete and/or unclear information. CONCLUSIONS The results of this study suggest that certain drug classes can be used as indicators of chronic disease. The CDI had a moderately strong correlation with the actual number of chronic diseases present, and a high correlation with other automated measures. The CDI tended to underestimate the actual number of disease states present, suggesting that the computer program is conservative in its estimation of
557
chronic disease prevalence. The CDI demonstrated the ability to detect seven specific diseases with moderate sensitivity and good-to-excellent specificity. Thus, this tool may be useful to estimate prevalence of these diseases in a large population. The authors acknowledge: Sune F. Isaksen and Jacob Jonassen for conducting the medical chart review; other IMPOVE co-investigators: Charles D. Sintek, M.S. and Debra J. Barnette, Pharm.D.; the guidance and support from the Veterans’ Affairs, Pharmacia & Upjohn Steering Committee: Jannet Carmichael, Pharm.D., Donald L. Kendzierski, Pharm.D., Julio Lopez, Pharm.D., Andy Muniz, M.D., Art Schuna, M.S., David Solomon, Charles E. Weber, R.Ph., J. Rod Barnes, David K. Burnham, David W. Johnson, R.Ph., George I. Wortman, R.Ph.; the expert panel members: Melissa Foss, Pharm.D., Larry A. Bourg, M.D., J. Mark Ruscin, Pharm.D., and Joseph J. Saseen, Pharm.D.; and anonymous reviewers for their insightful comments. This study is supported in part by Pharmacia & Upjohn and the ACCP Merck and Company Pharmacoeconomics Fellowship.
References 1. Elixhauser A, McCarthey E. Clinical Classifications for Health Policy Research, Version 2: Hospital Inpatient Statistics (AHCPR Publication No. 96-0017, 1996). Healthcare Cost and Utilization Project (HCUIP-3) Research Note 1, Agency for Health Care Policy and Research. Rockville, MD: U.S. Public Health Service; 1996. 2. Iezzoni LI. Assessing quality using administrative data. Ann Intern Med 1997; 127(8): 666–674. 3. Von Korff M, Wagner EH, Saunders K. A chronic disease score from automated pharmacy data. J Clin Epidemiol 1992; 45: 197–203. 4. Johnson RE, Hornbrook MC, Nichols GA. Replicating the chronic disease score (CDS) from automated pharmacy data. J Clin Epidemiol 1994; 47(10): 1191–1199. 5. Carter BL, Malone DC, Valuck RJ, Barnette DJ, Sintek CD, Billups SJ. The IMROVE study: Background and study design. Am J Health Syst Pharm 1998; 55: 62–67. 6. Koecheler JA, Abramowitz PW, Swim SE, Daniels CE. Indicators for the selection of ambulatory patients who warrant pharmacist monitoring. Am J Hosp Pharm 1989; 46: 729– 732. 7. Clark DO, VonKorff M, Saunders K, Baluch WM, Simon GE. A chronic disease score with empirically derived weights. Med Care 1995; 33: 783–795. 8. Iezzoni LI. Risk Adjustment for Measuring Healthcare Outcomes, 2d Edition. Chicago, IL: Health Administration Press; 1997. 9. Ash A, Porell F, Gruenberg L, Sawitz E, Beiser A. Adjusting Medicare capitation payments using prior hospitalization data. Health Care Financing Review 1989; 10(4): 17–29. 10. Starfield B, Weiner J, Mumford W, Steinwachs D. Ambulatory care groups: A categorization of diagnoses for research and management. Health Services Research 1991; 26(l): 53–74.