International Journal for Quality in Health Care 2004; Volume 16, Supplement 1: pp. i27–i35
10.1093/intqhc/mzh012
Distinguishing hospital complications
of care from pre-existing conditions
JAMES M. NAESSENS AND TODD R. HUSCHKA
Divisions of Health Care Policy & Research and Biostatistics, Mayo Clinic, Rochester, MN, USA
Abstract
Objective. To compare cases identiWed through the Complications Screening Program (CSP) as complications with cases using
the same ICD-9 secondary diagnosis codes, where the identifying diagnosis is also indicated as not present at admission.
Design. Observational study comparing two sources of potential hospital complications: published computer algorithms
applied to coded diagnosis data versus a secondary diagnosis indicator, which distinguishes pre-existing from hospitaldeveloped conditions.
Setting. All patients discharged from Mayo Clinic Rochester hospitals during 1998 and 1999. The Mayo Clinic is a large integrated delivery system in southeastern Minnesota, USA, providing services ranging from local, primary care to tertiary care for
referral patients. Approximately 35% of Mayo patients travel >200 km for medical care.
Study participants. Hospital patients (total = 84 436). The numbers of cases with complications ranged from 0 to 2444 per
algorithm.
Main outcome measures. Percent of algorithm complication cases indicated as developing in the hospital, and percent of
acquired conditions of that type detected by the computer algorithms. Incremental hospital charges, length of stay (LOS) and
mortality associated with acquired complications.
Results. The percent of cases identiWed through the computer algorithm that were also coded as acquired varied from 8.8% to
100%. The ability of the computer algorithms to detect acquired conditions of that type also varied greatly, from 2% to 99%.
Incremental charges and LOS were signiWcant for patients with acquired complications except for hip fracture/falls. Many
acquired complications also increased hospital mortality.
Conclusions. Complication rates based strictly on standard discharge abstracts have limited use for inter-hospital comparisons
due to large variability in coding across hospitals and the insensitivity of existing computer algorithms to exclude conditions
present on admission from true complications. However, complications do carry high costs, including extended stays and
increased hospital mortality. Enhancing secondary diagnoses with a simple indicator identifying which diagnoses were present
on admission greatly increases the accurate identiWcation of complications for internal quality and patient safety improvements.
Keywords: hospital complications, positive predictive value, pre-existing conditions, quality indicators, sensitivity
Renewed attention has been focused on adverse events in the
hospital through the publication of the two Institute of Medicine
(IOM) reports on patient safety [1,2]. Methods of efWciently
identifying hospital complications using existing administrative
databases would enable the epidemiological study of risk factors
leading to their development and give providers a way of identifying targets for improvement efforts. These methods might
also conceivably provide tools to assess and compare institutions
or individual providers. Having a relative cost estimate for complications would enable institutions and agencies to prioritize their
efforts at reducing these untoward events better.
Administrative hospital data (billing claims for discharge
abstracts) provide an inexpensive way of identifying and making
visible medical errors. Many vendors for the Joint Commission on
Accreditation of Healthcare Organizations ( JCAHO) ORYX
initiative offer hospitals several complication rates as potential
measures to be routinely reported (http://www.jcaho.org/
oryx.html; http://www.medstart.com/products/outcomes.
html). Many current hospital report cards from both state
agencies and the Internet are using complications as part of
their assessment of the quality of hospital care (http://
www.healthgrades.com) [3]. Many of these assessments rely on
applying computer algorithms to administrative data collected
in the billing process [4,5]. Most of these methodologies try to
exclude complications due to disease progression, but in that
process also exclude cases where care could be improved.
Address reprint requests to J. M. Naessens, Division of Health Care Policy and Research, Mayo clinic, 200 First Street SW,
Rochester, MN 55905, USA. E-mail: [email protected]
International Journal for Quality in Health Care vol. 16 Supplement 1
© International Society for Quality in Health Care and Oxford University Press 2004; all rights reserved
i27
J. M. Naessens and T. R. Huschka
For example, an algorithm to identify postoperative myocardial infarction excludes cardiac surgery patients from those
considered at risk. Furthermore, most claims data are not able
to distinguish between those complications that developed
during the hospitalization from those that were present
before the patient was admitted.
Lawthers et al. [6] performed an assessment of validity of
the Complications Screening Program (CSP) [4] through a
reabstraction study to ascertain the accuracy and completeness
of coding complications, and the ability to separate conditions
present on admission from these occurring in the hospital.
They recommended the widespread adoption of a modiWer to
secondary diagnosis codes which indicates whether that
condition was present on admission.
Beginning in July 1990, hospital abstractors at Mayo Clinic
Rochester began to collect a modiWer on every secondary diagnosis, which indicated whether that condition was present on
admission, developed during the hospitalization, or whether it
was unclear or uncertain with respect to timing. A previous
study [7] showed that the reliability of the coders to make this
determination was good and that the additional cost of
collecting the information was small. Interobserver agreement
was at least 80% for the six conditions assessed, with Kappa
statistics in the 0.58–0.67 range for renal failure, pneumonia
and decubitus ulcer, and in the 0.82–0.89 range for acute myocardial infarction (AMI), pulmonary embolism and stroke.
This study undertook to compare cases identiWed through
the CSP as complications with cases using the same ICD-9
secondary diagnosis codes, where the identifying diagnosis is
also indicated as not present at admission. This comparison
enables us to both identify those ‘complications’ that were
present on admission (false positives) and the cases with those
conditions that were excluded from the CSP algorithm (false
negatives). Incremental cost estimates of resources used,
length of stay (LOS) and mortality were determined for the
complications indicated as acquired conditions. As the previous study [6] has shown variation in the validity and extent of
pre-existing problems identiWed by the individual complication
algorithms, we present our results stratiWed by whether the
patients were those with a surgical, major invasive non-surgical
procedure (e.g. cardiac catherization or gastrointestinal endoscopies) or whether they were medical patients.
Methods
Setting
Data for this study were obtained from discharges at the
Mayo Clinic hospitals in Rochester Minnesota (Mayo) during
1998 and 1999. These Mayo hospitals are organized into three
separately licensed entities (Rochester Methodist, St Marys
and the Mayo Clinic Psychiatry and Psychology Treatment
Center) under one governing board, totaling 1951 beds. Combined, the three facilities cover a range of hospital services,
from those for typical community patients to tertiary referrals
for transplant, gamma knife and cardiac surgery. Provided services include most medical and surgical specialties for both
i28
adults and pediatrics, obstetrics, neonatalogy, psychiatry and
rehabilitation. Approximately 35% of Mayo patients travel
>200 km for medical care.
Complications
There are several sets of computer algorithms that attempt to
identify patients with hospital complications by screening for
selected secondary diagnosis codes identiWed through ICD-9CM codes in administrative data among a restricted population
at risk. For this study we chose to focus on the CSP developed
by Iezzoni [4]. Table 1 provides the ICD-9-CM diagnoses
codes used for identifying patients with postoperative stroke.
The subset of the hospital’s population considered at risk for
this example includes all major and minor surgery patients,
excluding patients admitted with neurological problems [major
diagnostic category (MDC) = 1] or those with preoperative
stays of >1 day. These patients are excluded to reduce the
rate of identifying patients where the ‘complications’ may be
pre-existing or due to underlying disease. Patients with a secondary diagnosis of pre-existing stroke may reasonably receive
neurologic surgery. Additionally, those with an acquired
stroke after neurologic surgery are less likely to be of iatrogenic
cause than other surgeries. The secondary diagnosis code,
which was identiWed by the algorithm, is considered the trigger
condition. Not every patient is at risk for each of the 28
complications of care algorithms. There are four risk groups:
major surgery, minor surgery, invasive non-surgical procedures
and medical patients. Each algorithm is applied to appropriate
risk groups with selected types of patients being excluded.
Three of the 28 algorithms were not assessed due to the complex
logic used. One algorithm (No. 27: Technical DifWculty with
Med Care) was excluded because it strictly relies on E-codes,
which are not consistently coded at our institution. Results
are reported for each of the major risk groups.
Table 1 Sample complication: postoperative stroke
Complication of care #1:
Postoperative stroke
Present/absent deWnition:
Secondary diagnosis: 433.x1*,
434.x1**, 997.02***
At risk deWnition1
Any Procedure
Excludes
Major or minor surgery
Preoperative >1 day
MDC = 1 (neurology
body system)
MDC, Major Diagnostic Category.
1
This “at-risk” deWnition includes any case with Any procedure of a
major or minor surgery, but Excludes surgeries with preoperative
stays longer than one day and any surgery in major diagnostic category one, neurology body system.
*
Occlusion and stenosis of precerebral arteries.
**
Occlusion of cerebral arteries with cerebral infarction.
***
Iatrogenic cerebrovascular infarction or hemorrhage postoperative
stroke.
Acquired versus pre-existing conditions
Acquired conditions
A modiWer was collected on each secondary diagnosis
code, which indicated whether the condition was present on
admission (P), acquired or developed during the hospitalization (A), or whether the timing was unclear (U). All cases for
each complication algorithm where the trigger condition was
identiWed as acquired were identiWed, irrespective of risk group
or exclusions.
acquired complication cases for each algorithm were based on
two sample t-tests or rank sum tests, dependent on the skewness of the data. Comparisons of cases of acquired complications between those in the risk pools and those normally
excluded were also performed. Cost estimates were calculated
on any complication with at least 30 cases in the 2 years.
Results
Assessment of complication accuracy
Descriptive
Hospital discharge abstracts including the acquired condition
Xag were used from patients discharged from hospitals at our
institution from 1998 to 1999. In compliance with the laws of
the State of Minnesota and our Institutional Review Board,
we used data only from patients who had authorized use of
their medical records for research [8]. Patients included the
full range, from local, primary care patients to tertiary care
referrals. These abstracts were processed through the set of
computer algorithms to identify cases that likely had a hospital complication. The acquired condition indicator for the
diagnosis that triggered the complication was then used to
determine if the condition was present at admission. Any case
at risk for the complication whose trigger condition was
present at admission was considered a false positive. Secondary diagnoses not present at admission which triggered complication codes were also identiWed among patients ‘not at
risk’ to assess the rates of false negatives for each algorithm.
Sensitivity, the percent of all acquired conditions of that type
detected by the algorithm, and positive predictive value (PPV),
the percent of all computer algorithm identiWed cases determined to have developed in the hospital, were calculated for
each complication (Table 2).
Of the 84 438 hospital discharges from 1998 to 1999, 7.8%
had one or more likely complication identiWed. The percent
of cases with complications was highest among major surgical
patients (13.5%). Minor surgical patients (9.7%) and diagnostic
procedure cases (8.9%) had relatively similar complication
rates, while the remaining cases, typically medical patients,
had relatively few complications (2.6%). Of the 24 studied
complication types, cases were found with secondary diagnosis
that met the criteria for 22 different complications. Eighteen
complications had at least 20 cases in our dataset.
Meanwhile, the percent of discharges with any acquired
condition was higher (13.9%), with highest levels among
major surgery (23.3%), similar rates among minor surgery
(12.2%) and diagnostic procedures (11.1%), and lower rates
among medical patients (7.2%). It should be noted that not
all acquired conditions are of equal severity, nor do they
necessarily reXect independent events. Table 3 provides the
most frequent acquired conditions among surgical patients
and among medical patients in 2 years at our institution.
Twenty-four different complication algorithms were deWned
in our assessment. In the 2 year experience we encountered
cases in 22 of these complications, while 19 complications
Estimation of impact of complications
The cost of complications in terms of additional charges, hospital days and mortalities were calculated. Expected charges,
LOS and hospital mortalities were based on statistical models
determined by the University HealthSystems Consortium
using standard hospital discharge data [9]. We estimated the
impact of the complications by using the difference between
observed and expected values normalized for the overall institutional difference for non-complicated cases. Ninety-Wve
percent conWdence intervals were calculated, assuming a
Gaussian distribution. Comparisons between pre-existing and
Table 2 Assessment of a complication
Complication
present
...........................................................................................................
Acquired condition present
Yes
No
Positive predictive value = a/(a+c)%
Sensitivity = a/(a+b)%
Assessment of complications
Yes
a
c
No
b
d
Table 3 The most frequent acquired secondary diagnosis codes
recorded during 1997–1998
ICD-9-CM
diagnosis code
Description
..........................................................................................................
Surgical cases
285.1
788.20
998.12
427.31
560.1
Medical cases
599.0
285.1
276.7
276.1
276.8
Acute posthemorrhagic anemia
Retention of urine, NOS
Hematoma complication of procedure
Atrial Wbrillation
Paralytic ileus
Urinary tract infection, NOS
Acute posthemorrhagic anemia
Hyperpotassemia
Hyposmolality
Hypopotassemia
NOS, not otherwise speciWed.
i29
J. M. Naessens and T. R. Huschka
had ≥10 cases. For one complication all of the identiWed cases
had the trigger condition indicated as developing within the
hospital, however it was based on the presence of a procedure
rather than a diagnosis. Table 4 provides the number of cases
with each computer algorithm complication, number of cases
with the acquired conditions that would have met the numerator
for each algorithm, and sensitivity and positive predictive
value for each of the complications encountered. For example,
101 qualifying surgical cases were identiWed with a secondary
diagnosis of stroke. However, 35 of these cases had indications
that the stroke diagnosis was present at admission, resulting in
a positive predictive value of 65%. There were 143 total cases of
stroke not identiWed at admission. Of these cases, 21 patients
were medical cases, 10 patients had non-surgical procedures,
while 46 patients had surgery but were excluded from the
postoperative stroke complication algorithm due to a neurological principal diagnosis (13) or a preoperative stay of >1 day
(33). The sensitivity of the algorithm to detect acquired strokes
was 46% overall and 59% when restricted to surgical cases.
Overall, Wve complications had sensitivities of ≥50%, with
four complications exceeding 80% sensitivity. Six complications
identiWed <20% of the speciWc acquired problems. PPV
exceeded 50% for seven complication algorithms, with one
algorithm based on procedure codes at 100% PPV. When
restricting analysis to only surgical patients, 10 complication
algorithms had sensitivities of ≥50%, with Wve complications
exceeding 80% sensitivity. Five algorithms identiWed <20%
of the speciWc acquired problems. PPV exceeded 50% for
nine complication algorithms when assessing only surgical
patients. Less than half of the acquired cases not detected by
the algorithm were surgical cases for aspiration pneumonia
(38.9%), gastrointestinal hemorrhage (37.1%), mechanical
complications (47.5%), miscellaneous complications (48.2%),
AMI (49.7%) and venous thrombosis and pulmonary embolism
(36.2%). Best screening performance was seen among the following complications: postoperative hemorrhage, iatrogenic
complications, and wound infections. The poorest screening
performances in terms of in-hospital-developed conditions were
among decubitus ulcers, postoperative complications relating
to urinary tract anatomy, postoperative cardiac abnormalities
other than AMI, shock or cardiorespiratory arrest and septicemia.
Only Wve of the complication algorithms we assessed are
applicable to medical cases and 10 apply to non-surgical
procedures. Post-procedural hemorrhage or hematoma did
identify all of the acquired cases among non-surgical cases
and 56.5% of algorithm-identiWed cases were not present on
admission. Similarly, both wound infections and iatrogenic
complications identiWed all of the acquired cases among nonsurgical cases; however, only 20% and 43%, respectively, of
identiWed cases were not pre-existing. The other complication
algorithms among non-surgical complications identiWed preexisting conditions >80% of the time and detected <50% of
the acquired conditions of that type.
Cost impact of complications
The 19 complication groups correctly identifying ≥30 acquired
cases in the time-frame were evaluated. Table 5 provides the
i30
sample size and estimated incremental cost in hospital
charges, LOS and hospital mortality. Overall, all acquired
complications were seen to add signiWcant costs and to add
days to LOS, except hip fracture/falls. The only complications
estimated with signiWcantly additional mortalities were stroke,
aspiration pneumonia, pulmonary compromise, gastrointestinal
hemorrhage, septicemia, surgical site reopening, cardiorespiratory arrest/shock, AMI, other cardiac abnormalities, hemorrhage/hematoma, and iatrogenic complications. Patients
with acquired postoperative stroke had mean charges exceeding
their institution-normalized expected charges by $16 315
[95% conWdence interval (CI) $9903–23 537]. Patients correctly
identiWed by the algorithm had mean incremental charges of
$14 854 and patients with acquired strokes missed by the
algorithm had mean incremental charges of $17 567, while
algorithm-detected patients with preexisting strokes had
incremental charges of $7293. Similarly for these patients, the
acquired conditions increased the average LOS by 6.1 days
(5.3, 6.8 and 3.7 days for identiWed-acquired, missed-acquired
and identiWed preexisting strokes, respectively), and increased
hospital mortality by 9.7 deaths per 100 admissions (95% CI
8.1–12.1). Postoperative septicemia on average increased the
hospital bill by $54 976 (95% CI $41 647–68 309), added 16.3
days to the stay (95% CI 13.2–19.6) and increased the hospital
mortality by 15.3 per 100 admissions (95% CI 13.7–16.8).
Aspiration pneumonia on average increased the hospital bill by
$30 122 (95% CI $21 143–39 011), added 10.9 days to the stay
(95% CI 7.2–14.0) and increased hospital mortality by 11.9
per 100 admissions (95% CI 9.8–14.1). Although surgical complications dominated the list of most expensive complications, medical complications also have major impacts. We
additionally examined the costs of urinary tract infections
(UTI) not coded as present at admission and compared those
among surgical cases (postoperative UTI complication) with
those among medical cases. The costs among surgical cases
were higher, but the costs added to medical cases were not
insubstantial, with hospital charges increased by 27%, average
LOS increased by 20% and hospital mortality increased by
35% over expected values.
In many, though not all cases, acquired complications had
signiWcantly higher costs and/or average LOS compared with
the algorithm-identiWed cases with preexisting conditions.
For example, cases with venous thrombosis or pulmonary
emboli identiWed as acquired had an incremental average LOS
of 11.8 days compared with the incremental increase of 4.8
days for preexisting cases of this type. There were no
instances where preexisting complications were signiWcantly
more costly or longer than acquired complications.
Discussion
Computer algorithms for identifying possible hospital complications vary in their ability to exclude cases where the ‘complication’ is actually a pre-existing condition. The experience at
our institution, which includes the range of patients from
primary to tertiary care, found that among the 18 complications
with at least 20 cases, 10 had ≥50% of cases with the trigger
Cases with complication
as deWned by algorithm, N
Percent of algorithm complications
identiWed as developing in hospital %
Cases with acquired
condition coded, N
Percent of acquired conditions
detected by algorithm %
1. Postoperative stroke
101
2. Aspiration pneumonia
76
3. Postoperative pulmonary compromise
311
4. Postoperative GI hemorrhage or
56
ulceration following non-GI
surgery
5. Postoperative complications relating to
34
urinary tract anatomy
6. Cellulitis or decubitus ulcer
13
7. Septicemia
25
8. Post/intra-operative shock due to
0
anesthesia
9. Reopening of surgical site
60
10. Mechanical complications due to
498
device, implant, or graft, except
organ transplant
11. Miscellaneous complications
605
12. Shock or cardiorespiratory arrest in
13
hospital
13. Postoperative complications relating to
32
the central or peripheral nervous system
14. Postoperative acute myocardial infarction 125
15. Postoperative cardiac abnormalities except 7
acute myocardial infarction
143
152
653
165
8
77
238
0
539
314
432
174
19
215
152
65.3%
57.9%
59.8%
44.6%
8.8%
42.6%
48.0%
–
100.0%
26.7%
52.9%
46.2%
21.9%
52.8%
42.9%
30.7%
2.0%
36.8%
74.1%
3.4%
11.1%
42.4%
7.8%
5.0%
–
37.5%
46.2%
28.9%
28.5%
15.2%
continued
...........................................................................................................................................................................................................................................................................................................
Complication1
Table 4 The accuracy of computer algorithm to capture complications using secondary diagnosis modiWer for hospital acquired conditions
Acquired versus pre-existing conditions
i31
i32
24.5%
NA
0.0%
46.9%
NA
0.0%
20.3%
47.6%
66.2%
40.9%
59.5%
NA
NA
NA
4
194
NA
0
98
241
572
1602
66
2444
NA
NA
30
1455
NA
NA
274
1064
118
0
2
315
NA
NA
52
GI, gastrointestinal; N, number; NA, not available.
1
Complications 17, 20 and 28 were excluded due to their complexity. Complication 27 was excluded due to inconsistent collection of E-codes.
16. Postoperative infections except
pneumonia and wound
17. Procedural related perforations or
lacerations
18. Postoperative coma or stupor
19. Postoperative pneumonia
20. Postoperative physiologic and
metabolic derangements
21. Complications relating to anesthetic
agents or other central nervous system
depressants
22. Venous thrombosis and pulmonary
embolism
23. Wound infection
24. Post-procedural hemorrhage
or hematoma
25. Inhospital hip fracture or fall
26. Iatrogenic complications
27. Technical difWculty with medical care
28. Complications due to medication events
Table 4 continued
90.0%
99.9%
NA
NA
99.3%
99.7%
41.5%
0.0%
0.0%
28.9%
NA
NA
46.2%
J. M. Naessens and T. R. Huschka
Acquired versus pre-existing conditions
Table 5 Incremental cost of complications identiWed through algorithm diagnosis codes with acquired condition modiWers
applied to all patients
Complication1
N
Incremental hospital
charges US$ mean
(95% CI)
Incremental LOS
days mean
(95% CI)
Incremental mortality
rate per 100 cases
mean (95% CI)
........................................................................................................................................................................................................................
1. Postoperative stroke
143 16,315 (9,093 to 23,537) 6.1 (3.9, 8.4)
2. Aspiration pneumonia
152 30,122 (21,143 to 39,011) 10.9 (7.2, 14.0)
3. Postoperative pulmonary compromise
653 38,015 (31,967 to 44,062) 10.1 (8.5, 11.7)
4. Postoperative GI hemorrhage or
165 21,411 (13,946 to 28,876) 6.5 (4.2, 8.8)
ulceration following non-GI surgery
6. Cellulitis or decubitus ulcer
77 21,329 (8,453 to 34,204) 13.3 (8.1, 18.6)
7. Septicemia
238 54,976 (41,647 to 68,309) 16.3 (13.2, 19.6)
9. Reopening of surgical site
539 17,712 (13,755 to 21,668) 5.0 (3.9, 6.0)
10. Mechanical complications due to
314 32,615 (24,106 to 41,125) 11.3 (9.1, 13.4)
device, implant or graft graft, except
organ transplant
11. Miscellaneous complications
432 6,264 (2,165 to 10,363)
3.2 (2.3, 4.2)
12. Shock or cardiorespiratory arrest in hospital 174 23,523 (14,831 to 32,215) 4.8 (2.3, 7.4)
14. Postoperative acute myocardial infarction
215 15,112 (9,685 to 20,539) 4.2 (2.6, 5.8)
15. Postoperative cardiac abnormalities
152 9,441 (2,194 to 16,687)
2.6 (0.6, 3.8)
except acute myocardial infarction
16. Postoperative infections except
52 44,480 (21,807 to 67,154) 17.6 (10.5, 24.8)
pneumonia and wound
19. Postoperative pneumonia
315 30,028 (22,044 to 38,013) 10.5 (8.5, 12.5)
22. Venous thrombosis and pulmonary
118 31,512 (15,121 to 47,903) 11.8 (7.6, 16.0)
embolism
23. Wound infection
274 27,955 (18,414 to 37,495) 10.8 (8.4, 13.2)
24. Post-procedural hemorrhage or hematoma 1064 10,149 (7,274 to 13,066) 3.1 (2.4, 3.8)
25. Inhospital hip fracture or fall
30 1,162 (−4,176 to 6,499)
1.2 (−1.1, 3.6)
26. Iatrogenic complications
1455 4,857 (2,792 to 6,925)
3.0 (2.4, 3.6)
9.7 (8.1, 12.1)
11.9 (9.8, 14.1)
11.8 (10.7, 12.8)
5.1 (3.3, 6.9)
−3.7 (−6.7, −0.6)
15.3 (13.7, 16.8)
3.1 (2.6, 3.7)
−1.8 (−3.0, −0.6)
−0.2 (−1.4, 0.4)
42.6 (40.4, 44.9)
9.2 (7.4, 11.1)
12.6 (10.9, 14.4)
−3.4 (−6.9, 0.0)
−2.6 (−4.1, −1.1)
0.8 (−0.9, 2.5)
−4.1 (−5.3, −2.9)
0.6 (0.2, 1.0)
−0.5 (−1.2, 0.1)
0.9 (0.5, 1.3)
Note: All mean values are based on the “excess” costs of acquired complication: (Observed − Expected) for each patient normalized for
institutional differences.
CI, conWdence interval; LOS, length of stay; GI, gastrointestinal.
1
Complications 17, 20, and 28 were excluded due to their complexity. Complications 5, 8, 13, 18, 21, and 27 had fewer than 30 cases.
condition identiWed as present at admission. Only one algorithm
had all cases identiWed as acquired and this algorithm was
based on the presence of a procedure. Some of the complications
including large numbers of pre-existing cases were venous
thrombosis and pulmonary emboli, urinary tract complications,
cardiac abnormalities other than AMI, and mechanical complications of devices and implants. This is similar to the
results reported by Lawthers et al. [6] from the Validation
Study of the CSP. In their study of Medicare patients from
Connecticut and California, they found 9% of postoperative
myocardial infarction cases identiWed by their complication
algorithm on billing data as present upon admission after
medical record review. They also found 87% of secondary
diagnoses of hip fracture or fall to be present on admission.
Their recommendations included a call for the adoption of a
modiWer digit to secondary diagnosis codes to indicate
whether that condition was present upon admission, as we
have reported here. Best and colleagues [10], in a study that
compared the ICD-9-CM diagnosis codes from routine
discharge abstracts for patients from US Veterans Affairs
(VA) hospitals with speciWc complication data obtained from
medical records for the National VA Surgical Risk Study,
found relatively low sensitivity and positive predictive values
for detecting the complications from the routine data and also
stated it was difWcult to separate preoperative characteristics
from the postoperative adverse events using the ICD-9-CM
codes. Without this clariWcation, coupled with the concern
about discrepancies in data quality across hospitals [11–14],
many complication algorithms could clearly lead to erroneous
conclusions in comparing institutions. In a study on patients
with elective lumbar diskectomies, Romano et al. [15] concluded
that more than half of the difference in risk-adjusted complication rates between hospitals was attributed to reporting
variations, with those hospitals labeled with more complications
than expected being more thorough in reporting. Furthermore, in a related article by Weingart et al. [16], a potential
quality problem existed in <30% of surgical patients and 16%
of medical patients Xagged by their computer algorithms as
i33
J. M. Naessens and T. R. Huschka
potential complications. Similarly, we did see better performance
of the algorithms among surgical performance than among
medical patients.
Many of the complication algorithms are deWned with risk
groups that attempt to minimize patients who developed their
conditions before admission to the hospital. As for acquired
myocardial infarctions, more cases were excluded from the risk
pool than were identiWed as postoperative, even when restricted
to surgical cases. Many of the acquired myocardial infarctions
were associated with cardiac surgery (31), many had ≥2 day preoperative stays (43), while many were among non-surgical procedures (30) or medical patients (45). In these groups, the
likelihood that the myocardial infarction is disease-related rather
than health-care-related is much higher. Acquired condition
codes provide better tools for quality improvement efforts within
an institution than computerized complication algorithms.
Although complication rates based strictly on standard
discharge abstracts have limited use for inter-hospital comparisons; complications carry very high costs, including
extended stays and increased hospital mortality. Fifteen different complications increased hospital charges by >$10 000,
11 speciWc complications added at least 5 days to the stay, and
seven complications increased mortality by at least Wve per
100 admissions. These data add to those of other studies [17],
which showed higher hospital charges, LOS and mortality
among cases identiWed using the CSP. Dimick et al. [18] found
complications to help explain the mortality differences for
abnormal aortic surgery between high- and low-volume hospitals. Enhancing secondary diagnoses with a simple indicator
identifying which diagnoses were present on admission
greatly improves the accurate identiWcation of problem cases
with higher mortality and resource costs for internal quality
and patient safety improvements.
There are several limitations to our study. A weakness is
the reliance on administrative data. The only potential diagnoses
considered as complications had to be recorded on the
abstract. During the study time-frame, our coders recorded
up to 15 diagnoses. McCarthy et al. [19] have reported on their
experience of Wnding clinical evidence to support the presence
of selected complications. We have not assessed validity
through reexamination of clinical evidence for complications
in the medical record, but have relied on the coders accurately
reXecting the physician’s recording in the chart. Conditions
may not have been consistently recorded, but we believe that
the reliability of the acquired condition Xag has been maintained.
Another limitation of our study is our method of estimating
the cost attributed to an acquired condition. Rather than making
a best effort of deciding which services were actually provided
for the complication, we attributed the average difference
between observed and expected values as the incremental
estimate in cost due to the presence of the problem. Furthermore, our methodology counts a patient with multiple
acquired conditions in multiple complications. Even if the
true costs of complications are lower than we estimate, the
relative value of these problems is likely to remain valid.
Finally, the rate of patients with pre-existing conditions,
particularly among surgical patients, is likely to be much
higher at a tertiary care referral center than among the community
i34
hospital population. Similarly, the cost of complications in
terms of charges, hospital days and mortality may all be
greater among the more severely ill patients at academic centers
than for those same complications among healthier patients.
Nevertheless, any inter-institutional comparisons need to
consider these possible differences also.
Further work needs to be performed to understand better
the utility of complication algorithms and acquired conditions.
A multi-institutional study with participants from California
or New York, where the pre-existing condition Xag has been
in use for a number of years, would help determine the generalizability of these Wndings. It may also be fruitful to examine
our data more closely for subgroups of patients. Many
patients with acquired complications had very high severity
scales at admission. It may be possible to discover risk factors
related to the development of iatrogenic events; Iezzoni et al.
[20] found a higher rate of complication after hospital discharge
among patients with comorbidities.
Conclusions
Conditions that are not present on admission clearly add
substantial costs to both surgical and non-surgical hospitalized
patients. Computer algorithms, which use secondary diagnoses
codes from administrative data do identify some patients who
have had complications, however, they tend to perform much
better among surgical than medical patients. Among patients
at a tertiary care referral center, some of the complication
algorithms also have a tendency to identify patients with the
‘complication’ being present on admission, while excluding
many patients who develop these problems from their populations considered at risk. The use of a Xag on each secondary
diagnosis to indicate whether the condition was present at
hospital admission is an inexpensive way to enrich the usefulness of administrative data to identify which patients may have
had iatrogenic complications.
Acknowledgements
The authors would like to thank Erika WohlWel for manuscript preparation and organizational assistance. This work
was presented in part at the International Society for Quality
in Health Care Fifth International Summit on Indicators in
Paris, France, in November 2002.
References
1. Kohn LT, Corrigan JM, Donaldson MS (eds), Committee on
Quality of Health Care in America, Institute of Medicine. To Err
Is Human: Building a Safer Health System. Washington, DC:
National Academy Press, 1999.
2. Committee on Quality Health Care in America, Institute of
Medicine. Crossing the Quality Chasm: a New Health System
for the 21st Century. Washington, DC: National Academy
Press, 2001
Acquired versus pre-existing conditions
3. Solucient. Solucient’s Top 100 Hospitals: Orthopedic Benchmarks for
Success—2000. Evanston, IL: Solucient (available online at http://
www.100tophospitals.com/studies/ortho00/default.htm).
4. Iezzoni LI, Daley J, Heeren T et al. Identifying complications of
care using administrative data. Med Care 1994; 32: 700–715.
5. DesHarnais SI, Forthman MT, Homa-Lowry JM, Wooster LD.
Risk-adjusted quality outcomes measures: Indexes for benchmarking rates of mortality, complications, and readmissions.
Qual Manage Health Care 1997; 5: 80–87.
6. Lawthers AG, McCarthy EP, Davis RB, Peterson LE, Palmer H,
Iezzoni LI. IdentiWcation of in-hospital complications from
claims data: is it valid? Med Care 2000; 38: 785–795.
7. Naessens JM, Brennan MD, Boberg CJ, Amadio PC, Karver PJ,
Podratz RO. Acquired conditions: An improvement to hospital
discharge abstracts. Qual Assur Health Care 1991; 3: 257–262.
8. Jacobsen SJ, Xia Z, Campion ME et al. Potential effect of
authorization bias on medical record research. Mayo Clinic Proc
1999; 74: 330–338.
9. University HealthSystems Consortium. Risk Adjustment of the Clinical Data Base. Oak Brook, IL: University HealthSystems Consortium (available online at http://www.uhc.edu).
10. Best WR, Khuri SF, Phelan M et al. Identifying patient preoperative
risk factors and postoperative adverse events in administrative
databases: results from the Department of Veteran Affairs
National Surgical Quality Improvement Program. J Am Coll Surg
2002; 194: 257–266.
11. Iezzoni LI. Assessing quality using administrative databases.
Ann Intern Med 1997; 127: 666–674.
12. Myers ER, Stage JF. Risk adjustment for complications of
hysterectomy: limitations of routinely collected administrative
data. Am J Obstet Gynecol 1999; 181: 567–575.
13. Geraci JM, Ashton CM, Kuykendall DH, Johnson ML, Wu L.
International ClassiWcation of Diseases, 9th Revision. Clinical
modiWcation codes in discharge abstracts are poor measures of
complication occurrence in medical inpatients. Med Care 1997;
35: 589–602.
14. Romano PS, Mark DH. Bias in coding of hospital discharge data
and its implications for quality assessment. Med Care 1994; 32:
81–90.
15. Romano PS, Chan BK, Schembri ME, Rainwater JA. Can
administrative data be used to compare postoperative complication
rates across hospitals? Med Care 2002; 40: 856–867.
16. Weingart SN, Iezzoni LI, Davis RB et al. Use of administrative
data to Wnd substandard care validation of complication screening
program. Med Care 2000; 38: 796–806.
17. Kalish RL, Daley J, Duncan CC, Davis RB, Coffman GA,
Iezzoni LI. Costs of potential complications of care for major
surgery patients. Am J Med Qual 1995; 10: 48–54.
18. Dimick JB, Pronovost PJ, Cowan JA, Ailawadi G, Upchurch
GR,Jr. The volume-outcome effect on abdominal aortic surgery:
Differences in case-mix or complications? Arch Surg 2002; 137:
828–832.
19. McCarthy EP, Iezzoni LI, Davis RB et al. Does clinical evidence
support ICD-9-CM diagnosis coding of complications? Med
Care 2000; 38: 868–876.
20. Iezzoni LI, Mackiernan YD, Cahalane MJ, Phillips RS, Davis
RB, Miller K. Screening inpatient quality using post-discharge
events. Med Care 1999; 37: 384–398.
Accepted for publication 27 October 2003
i35