Download An algorithm for the identification of undiagnosed COPD case

RESEARCH
An Algorithm for the Identification of Undiagnosed
COPD Cases Using Administrative Claims Data
DOUGLAS W. MAPEL, MD, MPH; FLOYD J. FROST, PhD; JUDITH S. HURLEY, MS; HANS PETERSEN, MS;
MELISSA ROBERTS, MS; JENO P. MARTON, MD; and HEMAL SHAH, PharmD
ABSTRACT
BACKGROUND: Chronic obstructive pulmonary disease (COPD) is a major cause of
death in the United States, but most persons who have airflow obstruction have
never been diagnosed with lung disease. This undiagnosed COPD negatively
affects health status, and COPD patients may have increased health care utilization
several years before the initial diagnosis of COPD is made.
OBJECTIVE: To investigate whether utilization patterns derived from analysis of
administrative claims data using a discriminant function algorithm could be used
to identify undiagnosed COPD patients.
METHODS: Each patient who had a new diagnosis of COPD during the study period
(N = 2,129) was matched to as many as 3 control subjects by age and gender.
Controls were assigned an index date that was identical to that of the corresponding
case, and then all health care utilization for cases and controls for the 24 months
prior to the initial COPD diagnosis was compared using logistic regression models.
Factors that were significantly associated with COPD were then entered into a
discriminant function algorithm. This algorithm was then validated using a separate
patient population.
RESULTS: In the main model, 19 utilization characteristics were significantly
associated with preclinical COPD, although most of the power of the discriminant
function algorithm was concentrated in a few of these factors. The main model
was able to identify COPD patients in the validation population of adult subjects
aged 40 years and older (N = 41,428), with a sensitivity of 60.5% and specificity
of 82.1%, even without having information on the history of tobacco use for the
majority of the group. Models developed and tested on only 12 months of utilization data performed similarly.
CONCLUSION: Discriminant function algorithms based on health care utilization
data can be developed that have sufficient positive predictive value to be used
as screening tools to identify individuals at risk for having undiagnosed COPD.
KEYWORDS: Chronic obstructive pulmonary disease, Epidemiology, Health care
utilization; Diagnosis
J Manag Care Pharm. 2006;12(6):458-65
Authors
DOUGLAS W. MAPEL, MD, MPH, is medical director, Lovelace Clinic
Foundation, Albuquerque, New Mexico. FLOYD J. FROST, PhD, is senior
scientist; JUDITH S. HURLEY, MS, is associate scientist; HANS PETERSEN,
MS, is adjunct senior scientist; and MELISSA ROBERTS, MS, is associate
scientist , Lovelace Respiratory Research Institute, Albuquerque, New Mexico.
JENO P. MARTON, MD, is director, U.S. Outcomes Research Group, Pfizer
Global Pharmaceuticals, Inc., New York, New York; HEMAL SHAH, PharmD,
is director, Health Economics & Outcomes Research, Boehringer Ingelheim
Pharmaceuticals, Inc., Ridgefield, Connecticut.
AUTHOR CORRESPONDENCE: Douglas W. Mapel, MD, MPH, Medical
Director, Lovelace Clinic Foundation, 2309 Renard Pl. SE, Suite 103,
Albuquerque, NM 87106-4264. Tel: (505) 262-7857; Fax: (505) 262-7598;
E-mail: [email protected]
Copyright© 2006, Academy of Managed Care Pharmacy. All rights reserved.
458 Journal of Managed Care Pharmacy
JMCP
July/August 2006
Vol. 12, No. 6
A
lthough chronic obstructive pulmonary disease
(COPD) is currently the fourth leading cause of death
in the United States and a rapidly growing concern
worldwide, its exact prevalence is difficult to estimate, and a
large number of cases are likely to be undiagnosed and untreated.
The 2002 National Health Interview Survey found that among
adults over age 65, 4.8% reported having a diagnosis of
emphysema and 5.4% reported having a diagnosis of chronic
bronchitis.1 The true prevalence of COPD is probably even
higher because it is an insidious disease and many patients are
not aware or ignore that they have a problem until a substantial
amount of their lung function is lost.2,3
The Third National Health and Nutrition Survey conducted
comprehensive health assessments on 20,050 adult individuals,
including spirometry on more than 82% of the cohort, selected
from across the United States from 1988 to 1994; it is as close to a
population-based survey of lung function as has ever been done in
this country.4 This survey found that 63% of adults with expiratory
airflow obstruction were not aware that they had a lung disease,
even though this undiagnosed obstruction was associated with
chronic respiratory symptoms and physical impairment.5-7
A growing body of evidence from projects such as the Lung
Health Study has shown that early interventions in COPD
can have a substantial impact on chronic symptoms and
survival.8,9 It is very important that health care providers identify
people with this disease and implement indicated interventions. However, obstacles such as the technical difficulties of
conducting in-office spirometry have made routine screening
for COPD impractical.10-13 It is possible that a more targeted
approach that identifies persons at risk for having undiagnosed
COPD and then refers them on for diagnostic spirometry could
be more efficient and effective.14,15
We have shown in our previous studies that COPD patients
have a higher risk of having comorbid conditions such as heart
disease and that they also have substantially increased health
care costs related to these diseases.16-18 If patterns of health care
utilization characteristic of undiagnosed COPD can be identified,
then it is possible that the information found in administrative
data, such as demographic information, health care claims,
and pharmacy records, could be used as an efficient means of
identifying persons at risk.
The goal of this project was to develop an algorithm that
identifies persons at risk for COPD based on health care
utilization information found in a managed care administrative
claims database. The overriding hypothesis for this effort was
that a set of common health care utilization characteristics
www.amcp.org
An Algorithm for the Identification of Undiagnosed COPD Cases Using Administrative Claims Data
could identify persons who are later diagnosed with COPD with
statistically significant and clinically useful sensitivity and
positive predictive value. To examine this hypothesis, we first
captured the patterns of health care utilization found among
COPD patients in the 2 years prior to their first COPD diagnosis
and then compared these with age- and gender-matched nonCOPD patients. Factors associated with preclinical COPD were
then entered into a discriminate function algorithm program to
determine which factors could be used to discern whether or
not a given individual was likely to have COPD. The algorithm
was then applied to a validation population to see if it
performed with adequate sensitivity and specificity to make it a
practical tool for screening purposes.
■■ Methods
Study Site
This study was conducted among members of the Lovelace
Health Plan (LHP), a staff- and network-model health maintenance organization (HMO) serving New Mexico. Lovelace
Health Plan is the insurance component of LovelaceSandia
Health Systems (LSHS), which also operates a network of
primary care clinics, specialty centers, and hospitals. LHP
served approximately 240,000 health plan members in 2001,
including members of the commercial plan (approximately 700
employer groups) and managed Medicare and Medicaid plans.
LSHS also serves 70,000 to 80,000 fee-for-service clients each
year. As ascertained by self-report on membership surveys, LHP
health plan members are 38.7% Hispanic, 55.8% non-Hispanic
white, 2.1% Native American, and 3.4% other racial designations.
Algorithm Development and Validation Cohorts
All Lovelace Health Plan members who were aged 40 years or
older during calendar year 2001 were randomly assigned to
either the algorithm development group or the algorithm
validation group. The algorithm development group was used
to identify utilization characteristics associated with COPD and
to develop the discriminant function algorithm. The algorithm
validation group was used to test the operational characteristics
of the algorithm and to prove its practical usefulness.
COPD patients in both the algorithm development and
validation groups were identified using medical and pharmacy
claims records. Any patient with one or more claims records
associated with an International Classification of Diseases,
Ninth Revision (ICD-9) code of 491.x (chronic bronchitis),
492.x (emphysema), or 496 (COPD) were designated as a
COPD case. We have examined the validity of these codes and
this system for identifying COPD by medical record abstraction
in previous projects and found them to be accurate, with
more than 95% of these patients having documented clinical
evidence to support the diagnosis.18
For inclusion in the algorithm development group, we also
required that COPD patients had received their first COPD
www.amcp.org
TABLE 1
Definitions Used in Models and Algorithms
Medical Encounter History
Dx (ICD-9) or CPT Code
Tobacco use
305.1 ≤ Dx <305.20 or Dx = V15.82
Edema
782.3
Asphyxia
799.0
Respiratory symptoms
786 ≤ Dx ≤ 786.40 or Dx = 786.52
Arterial circulatory disease
440 ≤ Dx ≤ 447.9
Hypertension
401 ≤ Dx <406 (excluding 401.9)
Valve disease
One of 424.1, 424.2, 424.3
Aortic aneurysm
One of 441.1 - 441.4, 441.9
Ischemic heart disease
410 ≤ Dx <415
Pulmonary heart disease
415 ≤ Dx <417
Heart failure
428 ≤ Dx ≤429
Atherosclerosis
440 ≤ Dx <441
Hematuria
599.7
Peptic ulcer
531 ≤ Dx <534
Asthma
493 ≤ Dx <494
Chest x-ray
71010, 71020
Airflow test
94009, 94010, 94060
Any pneumonia Dx
480 ≤ Dx <488
Any bronchitis Dx
466 ≤ Dx <467 or 490 ≤ Dx <491
Any respiratory infection Dx
079.99 or 465.9
Prescription History*
AHFS (one of)
Any pulmonary Rx
120808, 121200, 480000, 480800, 680400
Any antibiotic Rx
81204, 81206, 81212, 81224
Any cardiovascular Rx
201204, 241200, 401200
Any “other” Rx
100000, 880800, 920000
Any psychotherapeutic Rx
281608
Respiratory Exacerbation History†
One mild exacerbation
More than one mild exacerbation
Moderate exacerbation
Severe exacerbation
* American Hospital Formulary Service (AHFS) classes chosen were those where
more than 30 cases had a prescription of interest and where ratio of cases/controls
was at least 0.75.
† Exacerbations are defined as prescription fills for respiratory drugs or antibiotics
that are associated with either an outpatient visit (mild exacerbation), an
emergency department visit (moderate exacerbation), or a hospitalization
(severe exacerbation) for a respiratory condition.
CPT = current procedural terminology; Dx = diagnosis; ICD-9 = International
Classification of Diseases, Ninth Revision; Rx = prescription.
diagnosis (index diagnosis) between 1997 and 2001. To be
considered a new diagnosis, there could be no diagnosis codes
for COPD appearing prior to the index diagnosis dating back to
the patient’s initial membership in the health plan, or as of
Vol. 12, No. 6
July/August 2006
JMCP
Journal of Managed Care Pharmacy 459
An Algorithm for the Identification of Undiagnosed COPD Cases Using Administrative Claims Data
TABLE 2
Chest x-ray
71010
71020
Chest x-ray, frontal and lateral
Airflow test
94009
Peak flow
94010
Spirometry
94060
Pulmonary
Cardiovascular
Other
Psychotherapeutic
Chest x-ray, frontal
Bronchospasm evaluation
120808
Antimuscarinics/antispasmodics
121200
Sympathomimetic (adrenergic) agents
480000
Antitussives, expects, and mucolytic agents
480800
Antitussives
680400
Antibiotics
(matched to within 5 years, older or younger) and sex-matched
controls who did not have a diagnosis of COPD in their claims
records. Because of the advanced age of some COPD patients, it
was not possible to obtain a 3:1 match for each case.
Nevertheless, we were able to match 3:1 for 75% of the cohort
(n = 1,602), 2:1 for 22% of the cohort (n = 461), 1:1 for 3% of
the cohort (n=62), and only 4 patients could not be matched at
all. The final total for the control cohort was 5,790 patients.
CPT Codes and AHFS Descriptions
Used in Models and Algorithms
81204
Adrenals
Antifungal antibiotics
81206
Cephalosporins
81212
Macrolides
81224
Tetracyclines
520400
Anti-infectives
520404
Antibiotics
520406
Antivirals
520408
Sulfonamides
520412
Miscellaneous anti-infectives
840400
Anti-infectives
840404
Antibiotics
840406
Antivirals
840408
Antifungals
201204
Anticoagulants
241200
Vasodilating agents
401200
Replacement preparations
100000
Antineoplastic agents
880800
Vitamin B complex
920000
Unclassified therapeutic agents
281608
Tranquilizers
AHFS = American Hospital Formulary Service; CPT = current procedural terminology.
January 1, 1990. We excluded any patient who also had a
diagnosis of cancer (ICD-9 140-208; n = 260) or who had another
chronic lung disease not associated with COPD (494.x, 405.x,
and 500-519.x). Exceptions were skin cancers, excluding
melanoma (173.x); breast cancer (174.x); prostate cancer
(185.x); and benign neoplasms (210–239). These were permitted
since they tend to be indolent tumors with little effect on lung
function. Asthma (493.x) was also permitted because of the
common overlap between COPD and asthma. A total of 2,129
COPD cases meeting all inclusion and exclusion criteria were
identified for the algorithm development.
For each COPD case, we attempted to find 3 age-matched
460 Journal of Managed Care Pharmacy
JMCP
July/August 2006
Vol. 12, No. 6
Algorithm Development
To identify factors associated with preclinical COPD, we
captured all hospitalizations, outpatient encounters, and outpatient pharmacy prescription fills for 2 years prior to each
patient’s COPD diagnosis and during the same time period for
their matched controls. The 60 days prior to the date of initial
COPD diagnosis were excluded because utilization during this
time period is likely to be biased toward events that led to
making the diagnosis. Because of the vast number of ICD-9,
current procedural terminology (CPT), and American Hospital
Formulary Service (AHFS) codes used in this database, it was
necessary to condense some of the codes down to one descriptive
term that could then be entered into the logistic regression
model. Tables 1 and 2 list the diagnostic categories, related SAS
variables, and the associated ranges for the ICD-9, CPT, or
AHFS codes that were used in our models.
We then identified the utilization characteristics that were
most strongly associated with COPD using forward step-wise
conditional logistic regression equations (SAS for Windows 8.2;
Cary, NC). Factors that made a significant contribution to the
logistic regression models were then put into the discriminant
function algorithm6 (STEPDISC procedure in SAS) and run in
the administrative claims for the algorithm development
population. Ultimately, only those factors that contributed an
R2 value of 0.0015 or greater to the final algorithm were kept.
An additional algorithm was also developed that was based on
1 year of utilization data.
Identification and Characterization of Exacerbations
Due to the profound impact that exacerbations have on overall
utilization in COPD and the possibility that repeated exacerbations
may lead to the diagnosis, we examined whether identifying exacerbation events could improve the overall performance of the
algorithm. Exacerbations were defined as any inpatient or outpatient
encounter with a primary diagnosis of a respiratory system disease
(ICD-9 codes 462.x-519.x) that was also associated with a prescription fill for an antibiotic or respiratory medication.
Analyses: Algorithm Validation Phase
The algorithm was then applied to the validation group’s 1998
and 1999 claims records to test its sensitivity, specificity, and
positive predictive value as compared with the clinical diagnosis.
Two-by-two tables were created, with the claims diagnosis
www.amcp.org
An Algorithm for the Identification of Undiagnosed COPD Cases Using Administrative Claims Data
considered to be “gold standard” for comparison with the algorithms selection results. Sensitivities, specificities, and positive
predictive values were calculated using the STEPDISC program.
Medical Record Review
As an additional validation measure and to help estimate the
practical usefulness of the algorithm, we abstracted the medical
records of 200 patients from the validation group that the
algorithm had identified as likely to have COPD but who had
never had the clinical diagnosis. Also, to help understand why
the algorithm failed to identify some COPD patients, we
abstracted the records of another 200 patients from the validation
group who had a clinical diagnosis of COPD but who the
algorithm did not identify as at-risk patients. All records were
selected at random and abstracted by an experienced abstractor
using a standardized instrument. Specific clinical information
that would support the diagnosis of COPD included documentation of chronic respiratory complaints (e.g., dyspnea, cough,
wheezing, or more than 2 bouts of bronchitis within 12 months),
spirometry showing airflow obstruction, chest radiographs with
changes consistent with COPD, or a history of cigarette smoking.
Documentation of 2 of these findings was considered to be a
likely COPD case.
■■ Results
In a stratified comparison of utilization, prescription, and
exacerbation factors between the COPD cases and controls,
several areas of increased utilization can easily be identified
(Table 3). As expected, tobacco use is a significant factor, but,
unfortunately, the “V” codes identifying tobacco use are not
used routinely in this health system. The relatively low use of
respiratory medications (38.1%) among the COPD cases was
not unexpected since this is a time period prior to the first
diagnosis of COPD. Respiratory symptoms, episodes of
bronchitis, and use of chest radiographs were common among
the COPD patients, as were other diseases associated with
smoking, such as cardiovascular disease.
Results from the main logistic regression model are depicted
in Figure 1. Note that a history of tobacco use makes a large
contribution to the model, even though only a small proportion
of patients in the COPD group (13.4%) were ever given this
code. Inclusion of exacerbations did not have a significant
association with COPD after inclusion of the other clinical
factors in the model, so our system for identifying exacerbations
was not entered in the final algorithm. Having multiple visits or
pharmacy fills was no more predictive of having COPD than
just having 1 pharmacy fill, so we did not use indicators of high
utilization in any specific area as separate predictors in the
model. The predictive ability of the logistic regression model
was relatively good (percent concordance: 71.1) as were the
model fit statistics (Wald chi-square 968, 30 degrees of freedom;
P <0.001).
www.amcp.org
TABLE 3
Comparison of Patient Characteristics
and Health Care Utilization History
Among COPD and Control Groups*
COPD Dx
(n = 2,129)
Characteristic, % (n)
Age (mean [SE])†
66.0
Male
48.0 (1,021)
[0.3]
No COPD Dx
(n = 5,790)
65.3
[0.2]
47.6 (2,754)
Medical Encounter History
Tobacco use
13.4
(286)
Edema
7.3
(156)
Asphyxia
1.6
(35)
Respiratory symptoms
2.8 (161)
2.3 (134)
0.3
(17)
32.2
(685)
Arterial circulatory
6.0
(128)
2.1 (124)
Hypertension
9.0
(192)
4.4 (253)
Aortic/pulmonic valve disease
2.7
(58)
Ischemia
13.7 (792)
1.3
(78)
16.3
(346)
Cor pulmonale
1.6
(35)
7.9 (459)
Heart failure
8.0
(170)
Atherosclerosis
2.2
(47)
0.8
Hematuria
4.2
(89)
2.3 (130)
Peptic ulcer
3.5
(74)
Asthma
13.9
(296)
3.0 (173)
Chest x-ray
41.9
(893)
18.1 (1,047)
Airflow test
8.7
(186)
2.0 (120)
Any pneumonia diagnosis
9.9
(210)
3.6 (209)
Any bronchitis diagnosis
27.1
(576)
11.3 (656)
Any respiratory infection diagnosis
16.6
(345)
11.7 (679)
Any pulmonary Rx
38.1
(811)
16.0 (925)
Any antibiotic Rx
41.2
(878)
22.8 (1,322)
Any other Rx
13.1
(280)
1.9
(40)
15.1
(322)
0.3
(16)
2.3 (132)
1.6
(45)
(91)
Prescription History
Any psychotherapeutic Rx
Any cardiovascular Rx
6.2 (356)
1.0
(60)
8.1 (471)
Respiratory Exacerbation History‡
One mild exacerbation
15.9
(338)
8.5 (494)
More than one mild exacerbation
7.5
(159)
2.5 (145)
Moderate exacerbation
2.6
(56)
Severe exacerbation
1.4
(30)
None of above medical
history categories
19.0
(405)
1.1
(66)
0.4
(24)
48.5 (2,805)
* Medical history captured over 2-year period (January 1, 1998, through
December 31, 1999); events within 60 days of COPD diagnosis date for cases
have not been included.
† Age is calculated by taking the year of COPD diagnosis date (for controls,
diagnosis date of associated case) and subtracting the birth year.
‡ Exacerbations are defined as prescription fills for respiratory drugs or antibiotics
that are associated with either an outpatient visit (mild exacerbation), an
emergency department visit (moderate exacerbation), or a hospitalization
(severe exacerbation) for a respiratory condition.
COPD = chronic obstructive pulmonary disease; Dx = diagnosis, Rx = prescription.
Vol. 12, No. 6
July/August 2006
JMCP
Journal of Managed Care Pharmacy 461
An Algorithm for the Identification of Undiagnosed COPD Cases Using Administrative Claims Data
Results of the Logistic Regression Model Illustrating the Adjusted Odds Ratio (OR)
and 95% Confidence Limits (CLs) for the Patient and Utilization Factors That
Were Included in the Subsequent Discriminant Function Algorithms
Any Respiratory Infection Dx
History of Aortic/Pulmonary Valve Disease
Any Cardiovascular Rx
One Mild Exacerbation
Severe Exacerbation
History of Aortic Aneurysm
More Than One Mild Exacerbation
Moderate Exacerbation
Male
Any Pneumonia Dx
Any Antibiotic Rx
History of Hematuria
History of Artherosclerosis
Airflow Test
History of Respiratory Symptoms
Any Psychotherapeutic Rx
History of Ischemic Heart Disease
Any Bronchitis Dx
History of Hypertension
Any "Other" Rx
Chest X-ray
History of Arterial Circulatory Dxs
History of Peptic Ulcer
History of Heart Failure
Any Pulmonary Rx
History of Edema
History of Asphyxia
History of Asthma
History of Pulmonary Heart Disease
History of Tobacco Use
OR Estimates (Wald 95% CL)
FIGURE 1
DX = diagnosis; RX = prescription.
Only the 19 factors that were significantly associated with
COPD in the logistic regression model were entered into the
main discriminant function algorithm (Table 4). When applied
to the algorithm development population, the sensitivity of the
model was 44.7% and specificity 85.8%. Exclusion of factors
that contributed only 10 or less to the F value (i.e., the last
6 factors) had very little effect on the model’s sensitivity and
specificity (44.5% and 85.5%, respectively).
In a second model developed using only data from the
12 months prior to the diagnosis, nine of the 10 leading factors
included in the model were also among the 10 leading factors
in the original model (Table 5). When applied to the algorithm
development population, this algorithm’s sensitivity was 42.8%
and sensitivity 84.6%. When this model was retested after
excluding the variable for tobacco use, the sensitivity fell slightly
to 41.0% and specificity to 83.9%.
Algorithm Validation
The algorithm was then applied to the validation population
with 2 years of utilization data (Table 6a). Of a total population
of 41,428 adults aged 40 years and older, 2,240 out of 3,704
462 Journal of Managed Care Pharmacy
JMCP
July/August 2006
COPD patients (60.5%) were correctly identified, with a
positive predictive value of 25%. When the algorithm was
applied only to persons aged 65 years and older, the sensitivity
improved to 64% and positive predictive value to 38% (Table
6b). When the model was applied to the validation population
with only 12 months of cumulative utilization data, its
performance was only slightly less than that seen in the 2-year
population (Tables 7a and 7b).
When we excluded tobacco use from the algorithm and
reapplied it to the validation population, the effects were very
minor. When applied to the validation population with only
12 months of utilization data and restricted to persons aged 65
years and older, the sensitivity fell only from 60.5% to 59.9%,
and the positive predictive value declined from 38.6% to
37.6%. Although it would most likely be advantageous to have
tobacco use history on all patients, the algorithm is able to
identify more than half of the COPD patients in this group even
without any tobacco history information.
Results From the Medical Record Review
Of 200 patients who were identified by the algorithm as likely
Vol. 12, No. 6
www.amcp.org
An Algorithm for the Identification of Undiagnosed COPD Cases Using Administrative Claims Data
TABLE 4
Ranking of Independent Variables Used in
Discriminant Analyses Using 24 Months of
Health Care Utilization Data Prior to the
First COPD Diagnosis*
Factor
F Value
TABLE 5
P>F
Ranking of Independent Variables Used in
Discriminant Analyses Using 12 Months of
Health Care Utilization Data Prior to the
First COPD Diagnosis
Factor
F Value
P>F
1.
Chest x-ray
509.9
<0.001
1.
Respiratory Rx
388.3
<0.001
2.
Respiratory Rx
270.3
<0.001
2.
Chest x-ray
198.6
<0.001
3.
Tobacco use
211.2
<0.001
3.
Asthma Dx
101.4
<0.001
4.
Asthma Dx
101.4
<0.001
4.
Tobacco use
89.9
<0.001
5.
Heart failure Dx
64.9
<0.001
5.
Heart failure Dx
56.5
<0.001
6.
Bronchitis Dx
35.1
<0.001
6.
Bronchitis Dx
26.5
<0.001
7.
Edema Dx
27.4
<0.001
7.
Hypertension Dx
24.7
<0.001
8.
Other Rx
24.2
<0.001
8.
Other Rx
21.2
<0.001
9.
Ischemic heart Dx
18.0
<0.001
9.
Respiratory symptom Dx
15.4
<0.001
10.
Pulmonary heart Dx
14.9
<0.001
10.
Edema Dx
14.4
<0.001
11.
Hypertension Dx
13.7
<0.001
11.
Ischemic heart Dx
13.3
<0.001
12.
Respiratory symptom Dx
13.3
<0.001
12.
Antibiotic Rx
9.8
0.002
13.
Peripheral artery Dx
11.5
<0.001
13.
Pulmonary heart Dx
9.5
0.002
14.
Peptic ulcer Dx
10.4
0.001
14.
Male
6.2
0.006
15.
Respiratory infection Dx
9.7
0.002
15.
Peripheral artery Dx
5.5
0.013
16.
Male
7.4
0.006
16.
Peptic ulcer Dx
5.0
0.019
17.
Antibiotic Rx
7.4
0.006
18.
Spirometry
5.5
0.019
* P values were calculated for F values of one sample with 15 degrees of freedom.
COPD = chronic obstructive pulmonary disease; Dx = diagnosis, Rx = prescription.
19.
Asphyxia Dx
4.6
0.031
* P values were calculated for F values of one sample with 18 degrees of freedom.
COPD = chronic obstructive pulmonary disease; Dx = diagnosis, Rx = prescription.
to have COPD but who did not have a clinical diagnosis, 55
(27.5%) had at least 2 types of evidence supporting a diagnosis
of COPD in their medical records. These tended to be persons
who were smokers and who were treated one or more times for
respiratory infections. Conversely, of 200 COPD patients who
the algorithm said did not have COPD, 69 (34.5%) did not have
at least 2 types of evidence supporting the diagnosis. These
tended to be persons who did not appear to be very compliant
with medication regimens or follow-up to treatment.
■■ Discussion
Our study shows that it is possible to create a predictive
algorithm that uses routinely collected administrative data to
identify persons who may have preclinical or undiagnosed
COPD. The algorithm works even with very little or no
information on tobacco use in the database. We believe that this
algorithm could be used as part of an efficient and effective
secondary health intervention system to identify persons with
possible undiagnosed COPD and refer them for appropriate
work-up and treatment.
We examined a variety of factors that affected the performance
www.amcp.org
of the algorithm. The biggest limitation of this system is that the
sensitivity depends mostly on the patient having health care
utilization for COPD or other tobacco-related conditions.
A substantial proportion of the COPD population in the present
study did not have such utilization. For example, antibiotic use, the
most common factor among the COPD patients, was found in only
41.2% of the COPD patients in the 2 years prior to diagnosis.
Improved sensitivity allows one to identify more patients
who have COPD or who are at risk for the disease. As patient
age increases and overall health care utilization also increases,
the sensitivity of the algorithm improves. Nevertheless, the
algorithm’s sensitivity among persons aged 40 to 49 years in the
validation group was 49.9%. Thus, this algorithm can efficiently
identify many persons with relatively early disease and help get
their COPD diagnosed and treated before severe lung disease
and permanent disability have set in.
One argument against early case finding in COPD is the lack
of interventions proven to change the course of the disease. The
Lung Health Study has shown that COPD patients who are
provided smoking cessation counseling and who manage to
abstain from cigarettes for at least 5 years have significantly
improved probability of survival.19 To date, no pharmacologic
intervention has been proven to improve survival or slow the
accelerated airflow obstruction that is characteristic of COPD.
Vol. 12, No. 6
July/August 2006
JMCP
Journal of Managed Care Pharmacy 463
An Algorithm for the Identification of Undiagnosed COPD Cases Using Administrative Claims Data
TABLE 6A
Application of the Algorithm to the
Validation Group Based on 24 Months
of Health Care Utilization Data for All
Adults Aged 40 Years and Older
TABLE 7A
Application of the Algorithm to the
Validation Group Based on 12 Months
of Health Care Utilization Data for All
Adults Aged 40 Years and Older
Clinical Diagnosis
From Database
Algorithm
Result
Clinical Diagnosis
From Database
COPD
No COPD
Total
COPD
2,240
6,770
9,010
No COPD
1,464
30,954
32,418
Total
3,704
37,724
41,428
Sensitivity = 2,240/3,704 = 60.5%.
Specificity = 30,954/37,724 = 82.1%.
Positive predictive value = 2,240/9,010 = 24.9%.
Negative predictive value = 30,954/32,418 = 95.5%.
COPD = chronic obstructive pulmonary disease.
TABLE 6B
Algorithm
Result
COPD
No COPD
Total
COPD
2,334
7,828
10,162
No COPD
1,782
36,691
38,473
Total
4,116
44,519
48,635
Sensitivity = 2,334/4,116 = 56.7%.
Specificity = 36,691/44,519 = 82.4%.
Positive predictive value = 2,334/10,162 = 23.0%.
Negative predictive value = 36,691/38,473 = 95.4%.
COPD = chronic obstructive pulmonary disease.
Application of the Algorithm to the
Validation Group Based on 24 Months
of Health Care Utilization Data for
Adults Aged 65 Years and Older Only
TABLE 7B
Application of the Algorithm to the
Validation Group Based on 12 Months
of Health Care Utilization Data for
Adults Aged 65 Years and Older Only
Clinical Diagnosis
From Database
COPD
Algorithm
Result
No COPD
Total
Clinical Diagnosis
From Database
COPD
No COPD
Total
1,490
2,387
3,877
841
6,034
6,875
2,331
8,421
10,752
Sensitivity = 1,490/2,331 = 63.9%.
Specificity = 6,034/8,421 = 71.7%.
Positive predictive value = 1,490/3,877 = 38.4%.
Negative predictive value = 6,034/6,875 = 87.8%.
COPD = chronic obstructive pulmonary disease.
Algorithm
Result
JMCP
No COPD
Total
4,021
COPD
1,551
2,470
No COPD
1,011
7,038
8,049
Total
2,562
9,508
12,070
Sensitivity = 1,551/2,562 = 60.5%.
Specificity = 7,038/9,508 = 74.0%.
Positive predictive value = 1,551/4,021 = 38.6%.
Negative predictive value = 7,038/8,049 = 87.4%.
COPD = chronic obstructive pulmonary disease.
Limitations
There are limitations to this study that should be noted before
application to other cohorts. The clinical characteristics of the
Lovelace Health Plan COPD population are likely to be at least
slightly different from those found elsewhere, and the practice
habits of Lovelace Health System physicians are also likely to be
different. Because the algorithm is based on specific utilization
characteristics, this algorithm is likely to perform somewhat
differently in other health care systems. The positive predictive
value of any test depends on the prevalence of the target disease
in the study population, so cohorts that have a lower prevalence
of COPD than ours can be expected to have a lower positive
predictive value. The algorithm had a specificity in the range of
57% to 64% which indicates that the majority of patients
without COPD were appropriately classified, yet a substantial
proportion of patients without disease would be classified as
being at risk. Our tests of the various factors affecting the
464 Journal of Managed Care Pharmacy
COPD
July/August 2006
Vol. 12, No. 6
algorithm’s performance suggest that it is sufficiently robust for
application in other managed care systems; however, further
validation of this algorithm in other systems is warranted.
A sensitivity of 40% to 64% and specificity of 71% to 87%
would generally not be considered adequate to support the
routine use of the algorithm as a screening test. However, we do
not suggest that this algorithm be applied in the same way that
most clinical screening tests are applied. This algorithm should
be viewed simply as a tool that can efficiently identify a large
number of persons who have an increased risk of having a
debilitative and progressive respiratory condition. Tests with
positive predictive values in the 20% to 50% range can be very
effectively applied in early screening programs, but whether or
not this level of efficiency is adequate depends on judgments
about the impact of the disease, the usefulness of early
intervention, and the costs of screening. It is likely that addition
of a second screening test to this algorithm, particularly
www.amcp.org
An Algorithm for the Identification of Undiagnosed COPD Cases Using Administrative Claims Data
information about the history of tobacco use, could significantly
improve the overall positive predictive value. Application of this
algorithm will certainly not replace current recommendations
that all adults who have smoked 1 pack of cigarettes per day
for 10 years or who have other risk factors for COPD
have spirometry performed to document evidence of airflow
obstruction.8,9 Application of the algorithm can, however, help
identify those persons who are suffering the effects of
COPD earlier and direct them toward appropriate therapy.
■■ Conclusion
We believe that this algorithm works well enough for practical
application but, ultimately, the success of the algorithm will
depend on how well it works when applied as part of a program
for identifying and managing persons at risk for undiagnosed
COPD. Application of the algorithm to a database is relatively
simple and only 12 to 24 months of continuous data are needed.
Of the 9,010 persons in the validation group that the algorithm
identified as being likely to have COPD, 2,240 were confirmed
cases (PPV = 25%). Hence, it may be useful to further screen
persons through the use of respiratory-symptom questionnaires
and history of tobacco use before referring them for spirometry.
We also note that we did not match on race because race and
ethnicity are not variables in the database that was used in the
present study.
There are many practical issues that still must be considered,
such as how to appropriately approach patients who have been
identified as being at risk, how to combine the results of the
algorithm with information about tobacco history, and how best
to communicate this information to the patient’s primary care
provider. COPD is a growing problem, especially among
women, and to reverse this trend, we must find innovative ways
to identify patients at risk for the disease and at earlier stages.
Most managed care systems routinely collect the data on which
this algorithm is based, so we strongly recommend that they
consider using this approach as part of a program to improve COPD
care for their patients.
DISCLOSURES
Funding for this research was provided by Boehringer Ingelheim Pharmaceuticals,
Inc., and Pfizer Global Pharmaceuticals, Inc., and was obtained by author Douglas
W. Mapel on behalf of the Lovelace Clinic Foundation, where he is employed.
Mapel reported potential conflicts of interest, including support for the research
provided by grants from Boehringer Ingelheim and Pfizer, service on the speakers
bureaus for both of these companies, receipt of paid honoraria for giving lectures
on COPD, and service as a professional consultant to Pfizer on COPD-related
topics; he reported similar relationships with GlaxoSmithKline. Authors Jeno P.
Marton and Hemal Shah are employed by Pfizer and Boehringer, respectively. The
other authors reported no relationships with any companies that could be viewed
as sources of bias or potential conflicts of interest.
Mapel served as principal author of the study. Study concept and design
were contributed by Mapel, Marton, Shah, and author Floyd J. Frost. Data
collection was the work of authors Judith S. Hurley, Hans Petersen, and
Melissa Roberts; data interpretation was the work of all authors. Drafting of
the manuscript was primarily the work of Mapel, with input from Petersen
and Roberts; revision of the manuscript was the work of Mapel, with input
from the coauthors.
www.amcp.org
REFERENCES
1. Lethbridge-Cejku M, Schiller JS, Bernadel L. Summary health statistics for
U.S. adults: National Health Interview Survey, 2002. National Center for
Health Statistics. Vital Health Stat 10. 2004;222;1-151.
2. Halbert RJ, Isonaka S, George D, Iqbal A. Interpreting COPD prevalence
estimates: what is the true burden of disease? Chest. 2003;123:1684-92.
3. Mannino DM, Homa DM, Akinbami LJ, Ford ES, Redd SC. Chronic
obstructive pulmonary disease surveillance—United States, 1971-2000.
MMWR Surveill Summ. 2002;51:1-16.
4. National Center for Health Statistics. Plan and operation of the Third
National Health and Nutrition Examination Survey, 1988-1994 (NHANES III).
Hyattsville, MD: U.S. Department of Health and Human Services, Public
Health Service, CDC; 1994.
5. Mannino DM, Gagnon RC, Petty TL, Lydick E. Obstructive lung disease
and low lung function in adults in the United States: data from the National
Health and Nutrition Examination Survey, 1988-1994. Arch Intern Med.
2000;160:1683-89.
6. Soriano JB, Davis KJ, Coleman B, Visick G, Mannino D, Pride NB. The
proportional Venn diagram of obstructive lung disease. Chest. 2003;124:
474-81.
7. Coultas DB, Mapel D, Gagnon R, Lydick E. The health impact of undiagnosed airflow obstruction in a national sample of United States adults. Am J
Respir Crit Care Med. 2001;164:372-77.
8. Anthonisen NR, Skeans MA, Wise RA, Manfreda J, Kanner RE, Connett JE.
The effects of a smoking cessation intervention on 14.5-year mortality: a randomized clinical trial. Ann Intern Med. 2005;142:233-39.
9. Petty TL. Scope of the COPD problem in North America. Early studies of
prevalence and NHANES III data: basis for early identification and intervention. Chest. 2000;117(suppl):326S-331S.
10. Ferguson GT, Enright PL, Buist AS, Higgins MW. Office spirometry for
lung health assessment in adults: a consensus statement from the National
Lung Health Education Program. Chest. 2000;117:1146-61.
11. Man SFP, McAlister FA, Anthonisen NR, Sin DD. Contemporary management of chronic obstructive pulmonary disease. JAMA. 2003;290:2313-16.
12. Petty TL. Benefits of and barriers to the widespread use of spirometry.
Curr Opin Pulm Med. 2005;11:115-20.
13. von Weel C. Underdiagnosis of asthma and COPD: is the general practitioner
to blame? Monaldi Arch Chest Dis. 2002;57:65-68.
14. Stratelis G, Jakobsson P, Molstad S, Zetterstrom O. Early detection of
COPD in primary care: screening by invitation of smokers aged 40 to 55
years. Br J Gen Pract. 2004;54:201-06.
15. Buffels J, Degryse J, Heyrman J, Decramer M. Office spirometry significantly
improves early detection of COPD in general practice: the DIDASCO study.
Chest. 2004;125:1394-99.
16. Mapel DW, Hurley JS, Frost FJ, Peterson HV, Picchi MA, Coultas DB.
Health care utilization in chronic obstructive pulmonary disease: a casecontrolled study in a health maintenance organization. Arch Intern Med.
2000;160:2653-58.
17. Mapel DW, Picchi MA, Hurley JS, Frost FJ, Peterson HV, Coultas DB.
Utilization in chronic obstructive pulmonary disease: patient characteristics
and diagnostic evaluation. Chest. 2000;117(suppl):346S-353S.
18. Mapel DW, McMillan GP, Frost FJ, et al. Predicting the costs of managing
patients with chronic obstructive pulmonary disease. Respir Med. 2005;99:
1325-33.
19. Anthonisen NR, Skeans MA, Wise RA, Manfreda J, Kanner RE, Connett
JE, and the Lung Health Study Research Group. The effects of a smoking cessation
intervention on 14.5-year mortality: a randomized clinical trial. Ann Intern
Med. 2005;142:233-39.
Vol. 12, No. 6
July/August 2006
JMCP
Journal of Managed Care Pharmacy 465