Shoulder arthroplasty has become the treatment of choice for many patients
with a glenohumeral injury or disease. The frequency of shoulder
arthroplasties has increased substantially over the past decade from
approximately 10,000 in 1990 to 20,000 in
20001.
An inverse relationship between hospital and surgeon volume and nonoptimal
clinical outcomes after surgery has been demonstrated for total hip
arthroplasty and total knee
arthroplasty2-10.
Shoulder arthroplasty is a technically demanding procedure that is not
performed routinely, as most surgeons perform only one or two arthroplasties
per year11. The
combination of surgical difficulty and minimal surgeon experience could
influence patient outcomes. To the best of our knowledge, no study has
investigated the relationship between volume and outcomes for shoulder
arthroplasty.
The objective of this study was to examine the relationship between surgeon
and hospital volume and outcomes in shoulder arthroplasty with use of the
Nationwide Inpatient Sample databases. We hypothesized that surgeons and
hospitals with higher caseloads for total and partial shoulder arthroplasty
have better outcomes as measured by a decreased mortality rate, shorter length
of hospital stay, reduced postoperative complications, and routine disposition
of patients on discharge.
Design
We performed a secondary analysis of the Nationwide Inpatient Sample (NIS)
databases for the years 1988 through 2000.
Database Description
The NIS database for the years 1988 through 2000 was used for this
study12. The NIS is
part of the Healthcare Cost and Utilization Project (HCUP), sponsored by the
Agency for Healthcare Research and Quality (AHRQ). Only patients who were
admitted to the hospital are included. The NIS is the largest database for
all-payer inpatient care that is publicly available in the United States, and
it contains approximately five to eight million records of inpatient stays per
year from about 1000 hospitals, which represent a 20% stratified sample of
community hospitals in the United
States13. To ensure
maximal representation of hospitals in the United States, the following
sampling strata based on five important hospital characteristics were used for
the creation of the NIS database: geographic region (Northeast, North Central,
West, and South), ownership (public, private not-for-profit, and private
investor-owned), location (urban and rural), teaching status (teaching
hospital and nonteaching hospital), and size (small, medium, and large) in
terms of the number of beds. Information on hospital ownership was obtained
from the American Hospital Association (AHA) Annual Survey of Hospitals and
includes categories for government nonfederal (public), private not-for-profit
(voluntary), and private investor-owned
(proprietary)14.
NIS datasets provide the following information: hospital identifiers
(AHRQ-sponsored and AHA identifiers), synthetic surgeon identifiers, unique
patient visit identifier, patient demographic data, and procedure and
diagnostic codes classified according to the International Classification
of Diseases, Ninth Edition, Clinical Modification
(ICD-9-CM)15.
The HCUP assigned validation and quality assessment of these datasets to an
independent
contractor16. To
perform the validation, the contractor reviewed the univariate statistics for
all numeric data elements and the frequency distributions for all categorical
and some continuous data elements, checked the range against standard norms,
and performed edit checks to identify inconsistencies between related data
elements. The NIS database has also been extensively validated against the
National Hospital Discharge Survey and was confirmed to perform very well for
many
estimates17.
The combined datasets (1988 through 2000) contain information on 12,876
patients who had a total shoulder arthroplasty and 17,999 patients who had a
hemiarthroplasty.
Sample Selection
Data were extracted separately for total shoulder arthroplasties and
hemiarthroplasties. The records with an ICD-9-CM procedure code for total
shoulder replacement (81.80) and for partial shoulder replacement (81.81) were
initially included in the analysis (see Appendix). Each record in the datasets
represents a single patient visit and has a unique identification number. As
our dataset does not contain a unique patient identifier, patients who were
readmitted could not be tracked. Patients with a procedure code for revision
shoulder arthroplasty (81.83 — Other repairs of shoulder including
revision shoulder arthroplasty) were not included in the study.
Patients who had a primary or secondary diagnosis of infection, malignant
tumor, or pathological fracture in the bones of the shoulder region were
excluded from the analysis. Cases of patients with evidence that the present
surgery was performed as a result of complications of previous shoulder
arthroplasty were also excluded (see Appendix). Stratifications based on the
diagnosis of osteoarthritis and fracture of the humerus, scapula, or glenoid
were attempted for both total shoulder arthroplasty and hemiarthroplasty.
There were 12,594 records for total shoulder arthroplasty and 17,452 records
for hemiarthroplasty included in the final analyses.
Outcome Measures
The outcomes of interest included inhospital mortality rate, length of stay
in the hospital, disposition of the patient on discharge, and inhospital
postoperative complications. The mortality rate was based on whether the
patient died during hospitalization or was discharged alive. Length of stay
was calculated in days by subtracting the admission date from the date of
discharge.
The disposition of the patient on discharge was coded into routine and
nonroutine disposition. Nonroutine disposition included transfer to a
short-term hospital, skilled nursing facility, intermediate care facility,
another type of facility, or home health care. Routine disposition reflected
patients who were discharged home. The variable for postoperative
complications was created on the basis of the information from fourteen
secondary diagnoses included in the datasets. Patients with a secondary
diagnosis of postoperative wound infection, other infections, a nonhealing
surgical wound, disruption of the operative wound, pulmonary embolism,
thrombophlebitis, and other unspecified complications were considered to have
a postoperative complication (see Appendix).
Main Effects
The primary predictor variables included surgeon and hospital volume. The
databases contained a synthetic primary surgeon identifier for each surgeon,
which was consistent over thirteen years. The synthetic surgeon identifier was
a fixed-key (one-to-one) encryption of the primary surgeon number and served
as a unique identifier for each surgeon. Surgeon volume was calculated by
counting the number of total shoulder arthroplasties of hemiarthroplasties
performed for each of the respective subpopulations during a given year with
use of this unique identifier. Surgeon volume was then divided into three
categories (fewer than two procedures, two to four procedures, or five
procedures or more per year) for both total shoulder arthroplasty and
hemiarthroplasty. Synthetic primary surgeon identifiers were missing for 41.3%
of total shoulder arthroplasties and 45.5% of hemiarthroplasties. In order to
test the impact of missing surgeon identifiers on our results, a sensitivity
analysis was conducted. Imputation by best subset regression of missing values
for surgeon volume was used to conduct the sensitivity analysis. We calculated
the value for missing surgeon volume on the basis of other characteristics
such as hospital volume, hospital location, teaching status of hospital,
hospital identifier, hospital size in terms of the number of beds, and year of
the operation. The combined characteristics of these known variables were used
to find the most likely estimate of surgeon volume. The calculation was based
on patterns of other characteristics observed in patients for whom the
information regarding surgeon volume was available. The method then matched
these characteristics to patients without surgeon volume and imputed the most
likely value. This method is called imputation, which represents an
established and frequently used statistical
tool18-23.
Additionally, to test the potential impact of missing surgeon identifiers on
our findings, patients with surgeon identifiers were compared with patients
without surgeon identifiers in terms of outcomes and demographic variables. We
found that the two patient subsets were very similar.
Similarly, each hospital had a unique hospital identifier, and this was
used to determine the three categories (fewer than five procedures, five to
nine procedures, and ten procedures or more per year) for hospital volume of
total shoulder arthroplasties and hemiarthroplasties. None of the hospital
identifiers were missing.
We chose surgeon and hospital volume categories to obtain approximately
similar percentages of procedures in each category and also to have clinically
meaningful cut-offs. The terms high volume and low volume
for hospitals and surgeons in this article reflect only surgical volume of
either total shoulder arthroplasty and hemiarthroplasty and not total surgical
volume. In this article, the term high volume is used for surgeons or
hospitals in the highest volume category; medium volume, for the
surgeons or hospitals in the middle volume category; and low volume,
for surgeons and hospitals in the lowest volume category.
Covariates
Covariates that are available from NIS include age, sex, race, household
income, and comorbidity (according to the Charlson index as modified by Deyo
et al.) of the
patient24,25.
The Charlson index measures comorbidity by assigning scores of 1, 2, 3, or 6
to each of the comorbid conditions present in a patient. These scores are then
added to provide a single index score, which measures the overall comorbidity
of the patient. Income is estimated by the median household income in the
patient's zip code.
Statistical Analysis
Each of the analyses mentioned below was performed for both total shoulder
arthroplasty and hemiarthroplasty. Univariate analyses were performed with use
of means and proportions in percentage. Bivariate analyses were performed to
measure the association between surgeon and hospital caseload and the
remaining covariates. This analysis yielded the proportions in percentage of
each surgeon-volume category across hospital-volume categories. Hospital
volume was also tabulated across hospital size, total hospital charges, and
hospital teaching-status categories. Hospital size categories (small, medium,
and large) were based on the number of hospital beds and are specific to the
hospital's location and teaching status. The hospital teaching status was
obtained from the AHA Annual Survey of Hospitals. A hospital is considered to
be a teaching hospital if it has an American Medical Association-approved
residency program of any type, is a member of the Council of Teaching
Hospitals, or has a ratio of full-time-equivalent interns and residents to
beds of 0.25 or
higher12.
Multivariate logistic regression models were used to examine the
risk-adjusted association between the surgeon and hospital volume and the
outcomes. The surgeon-volume models were controlled for hospital volume (as a
continuous variable), but surgeon volume was not used as a confounder for
models with hospital volume as the main effect to avoid exclusion of records
with missing surgeon volume. Each model was adjusted for age, sex, race,
household income, and comorbidity of the patient.
Length of stay in the hospital was examined with use of multivariate linear
regression models. Length of stay, which was used as a continuous variable,
had a skewed distribution and therefore was modeled with use of a logarithmic
transformation. Estimated mean length of stay was obtained by the
exponentiation of regression coefficients.
Adjusted odds ratios with 95% confidence intervals were used to express the
strength of association between the surgeon and hospital volume and the
outcomes. Generalized estimating equations were used to control for clustering
of patients within hospitals. Adjusted estimates were calculated for length of
stay with use of linear regression. The White
test26 was
performed to determine heteroscedasticity in the linear regression models. The
estimated parameters were also corrected with use of a smearing factor to
adjust for heteroscedasticity (as the White test was significant, p <
0.001) and logarithmic
transformation27,28.
Because of the very low mortality rates in our dataset, which are in
agreement with those in other
datasets29 and
previous studies30,
the outcome rate of mortality for total shoulder arthroplasty was not
sufficient to do a regression analysis for surgeon volume. Hence, surgeon
volume was recategorized (fewer than two procedures, two to fewer than four
procedures, and four procedures or more) to calculate adjusted odds ratios. A
sensitivity analysis with surgeon-volume categories of fewer than two
procedures, two to fewer than three procedures, and three procedures or more
was also conducted to add robustness and validity to the analysis.
Incremental odds ratios were used to determine whether every increase in
hospital or surgeon volume (category) is associated with an increased risk of
the outcome. This approach is more stringent and accurate than the Mantel
extension trend
statistic31, and it
requires that all of the incremental odds ratio estimates be greater than
(less than) 1.0 in order to confirm a dose-response relation.
Stratification based on the diagnosis of osteoarthritis or fracture of the
humerus, glenoid, or scapula was attempted. Because of the extremely small
percentage of patients who died, multivariate logistic regression was
conducted only on postoperative complications and nonroutine disposition of
the patient on discharge as outcome variables.
Statistical analyses were conducted with use of Inter-cooled Stata for
Windows (version 7.0; Stata, College Station, Texas) and SAS for Windows
(version 8.02; SAS Institute, Cary, North Carolina).
Patients included in our analysis were predominantly white (66.9% for total
shoulder arthroplasty and 65.2% for hemiarthroplasty) and female (61.2% for
total shoulder arthroplasty and 70.1% for hemiarthroplasty), and they had a
mean age of approximately sixty-eight years for both procedures. The mean
Charlson index was 1.5 ± 3.2 for total shoulder arthroplasty and 1.6
± 3.6 for hemiarthroplasty, and the mean number of diagnoses on
discharge was 4.0 ± 2.4 for total shoulder arthroplasty and 4.6
± 2.7 for hemiarthroplasty (Table
I).
On the average, patients undergoing hemiarthroplasty had a longer stay in
the hospital (4.9 ± 5.3 days) than those who had a total shoulder
arthroplasty (3.9 ± 3.8 days). Postoperative complications were
uncommon, affecting 1.2% of the patients who had a total shoulder arthroplasty
and 1.3% of the patients who had a hemiarthroplasty. Nonroutine disposition
was recorded for 27.5% of the patients who had a total arthroplasty and 34.6%
of patients who had a hemiarthroplasty
(Table II).
Surgeons in the low-volume category performed 33.5% of the total shoulder
arthroplasties and 42.0% of the hemiarthroplasties; those in the medium-volume
category, 37.6% and 44.0%, respectively; and those in the high-volume
category, 28.9% and 14.1%. The percentage of procedures performed consistently
decreased for low-volume surgeons across low to high-volume hospitals for
total shoulder arthroplasty (21.5% to 4.5%) and hemiarthroplasty (24.5% to
6.0%). In contrast, for surgeons with a higher caseload, the proportion of
procedures performed consistently increased across low to high-volume
hospitals for total shoulder arthroplasty (0.3% to 20.8%) and hemiarthroplasty
(0.4% to 9.4%) (see Appendix). In an attempt to better understand the
distribution of surgeon volume across hospital volume, a bar graph was drawn
that displays the distribution of surgeon volume as a proportion of individual
hospital-volume categories (Fig.
1).
Bivariate analysis of hospital volume and hospital characteristics was
performed. These results are displayed in the Appendix.
The multivariate logistic regression modeling demonstrated that patients
undergoing total shoulder arthroplasty were 4.4 times (95% confidence
interval, 0.6 to 31.2) more likely to die during the hospital stay if the
operation was done by a surgeon who performed fewer than two procedures per
year and 4.2 times (95% confidence interval, 0.6 to 29.6) more likely to die
if the operation was performed by a surgeon who did two to fewer than four
procedures per year than were patients managed by surgeons who performed at
least four procedures every year (Table
III). The trends analysis also yielded incremental odds ratios of
4.2 and 1.04 for decreasing surgeon-volume categories compared with surgeons
with volume of at least four procedures (see Appendix). The sensitivity
analysis performed with use of hospital-volume categories of fewer than two
procedures, two to fewer than three procedures, and three procedures or more
per year also yielded similar results. The results obtained for hospital
volume were similar, as patients managed at low-volume and medium-volume
hospitals were 2.1 times (95% confidence interval, 0.7 to 6.6) and 1.5 times
(95% confidence interval, 0.4 to 5.5), respectively, more likely to die
compared with those managed at high-volume hospitals
(Table IV). Incremental odds
ratios of 1.5 and 1.4 showed a positive trend effect (see Appendix).
Patients undergoing hemiarthroplasty performed by surgeons who did fewer
than two procedures and by those who did between two and four procedures were
2.2 times (95% confidence interval, 1.1 to 4.4) and 1.5 times (95% confidence
interval, 0.7 to 3.2), respectively, more likely to have postoperative
complications than were those managed by higher-volume surgeons. Trends
analysis also showed a dose-response relationship with incremental odds ratios
of 1.5 for both categories of decreasing surgeon volume. Significant results
were obtained for the relationship between hospital volume for total shoulder
arthroplasty and postoperative complications. Odds ratios of 2.5 (95%
confidence interval, 1.5 to 4.2) for low-volume hospitals and 2.1 (95%
confidence interval, 1.2 to 3.6) for medium-volume hospitals compared with
high-volume hospitals with a positive trends analysis (odds ratio, 2.1 and 1.2
for decreasing hospital volume) were obtained.
The risk-adjusted odds ratios of nonroutine discharge after
hemiarthroplasty were 1.3 (95% confidence interval, 1.1 to 1.5) and 1.3 (95%
confidence interval, 1.1 to 1.6) for patients managed with surgeons who
performed fewer than two procedures per year and for those managed by surgeons
who performed two to less than five procedures per year, respectively,
compared with those managed by surgeons with a case-load of at least five
procedures per year. These results were significant (p = 0.01)
(Table III). Similar results
were obtained for hospitals that had a low or medium volume of total shoulder
arthroplasties and hemiarthroplasties compared with those that had a high
volume (Table IV).
Surgeons who performed five total shoulder arthroplasties or more per year
discharged their patients an average of seventeen hours earlier than did
surgeons who performed fewer than two procedures per year (p < 0.001).
Similarly, patients who had the operation in hospitals with a volume of ten
procedures or more per year were discharged an average of twelve hours earlier
than those who had the operation in hospitals with caseloads of between five
and nine procedures per year and 1.1 days earlier than those who had the
operation in hospitals with fewer than five procedures per year (p <
0.001). The mean length of stay for patients managed by surgeons who performed
fewer than two hemiarthroplasties per year (5.4 ± 1.3 days) was
significantly higher than that for those managed by surgeons who performed
five or more procedures (4.1 ± 1.1 days) (p < 0.001). On the
average, patients managed in hospitals where ten hemiarthroplasties or more
were performed every year were discharged twelve hours before patients managed
in hospitals with a volume between five and nine procedures per year and 1.1
days before patients managed in hospitals with a caseload of fewer than five
procedures per year; the difference was significant (p < 0.001)
(Table V).
The sensitivity analysis performed with use of imputation by best-subset
regression for missing surgeon volume revealed a variation of up to 13.6% in
the odds ratios for all outcomes with a positive trends analysis, except
mortality rate for patients managed with total shoulder arthroplasty, which
still had odds ratios of >1.0 for both medium and low-volume surgeons and a
positive trends analysis. Hence, we concluded that our analysis is robust and
valid.
Stratification was performed on the data for patients with a diagnosis of
osteoarthritis and those with fractures of bones in the shoulder region.
Similar to the observations made on analysis of the results of total shoulder
arthroplasty and hemiarthroplasty, associations between better outcomes and
higher surgeon and hospital volume were observed for some of the variables
(see Appendix).
To the best of our knowledge, this study is the first attempt to
investigate the relationship between volume and outcomes for shoulder
arthroplasty. We used thirteen years of data from a 20% stratified probability
sample of community hospitals in United States to examine whether surgeon and
hospital volume were related to patient outcomes such as mortality,
postoperative complications, disposition of patient on discharge, and length
of stay. The multivariate logistic regression modeling demonstrated that the
likelihood of mortality during hospitalization associated with both total
shoulder arthroplasty and hemiarthroplasty increases as the volume of such
procedures performed by the surgeons decreases and that the likelihood of
postoperative complications associated with hemiarthroplasty is higher in
patients managed by low-volume surgeons. An incremental pattern in the
possibility of postoperative complications and nonroutine disposition of the
patient on discharge after total shoulder arthroplasty was also observed in
association with declining hospital volume. The mean length of stay was
significantly lower for patients of high-volume surgeons and those treated in
high-volume hospitals compared with those managed by low-volume surgeons and
those treated in low and medium-volume hospitals.
In the risk-adjusted analysis, the likelihood of inhospital mortality after
total shoulder arthroplasty was found to be low for high-volume surgeons and
for high-volume hospitals with a confirmatory trends analysis; however; these
findings need to be interpreted with caution. Wide confidence intervals as a
result of the small percentage of patients who died necessitate further
investigation of this outcome. Similar observations were made in the study by
Kreder et al., who attempted to determine the relationship between surgeon and
hospital volume and mortality for total hip arthroplasty during initial
elective
hospitalization5.
Infection after shoulder arthroplasty has been well documented in previous
studies32-34,
and, although pulmonary embolism is not frequently associated with shoulder
arthroplasty, it has been
reported35,36.
Katz et al. 7
examined the association between surgeon and hospital volume of total hip
arthroplasties and pulmonary embolus and found no significant association;
however, in agreement with our findings, they concluded that lower rates of
deep infection following total hip arthroplasty were associated with greater
hospital volumes. With use of data from the Washington State Department of
Health, Kreder et al. found that patients managed by low-volume surgeons had
more infections after total hip arthroplasty than did patients managed by
high-volume
surgeons5.
The effect of length of stay on patient outcome is an important area of
research37-41.
Our results are in conformity with those of Kreder et al., who found an
association between patients managed with a total hip arthroplasty by
high-volume surgeons and a shorter length of hospital
stay6. Lavernia and
Guzman, in a study of arthroplasty procedures, noted a similar association
between a prolonged length of stay and low-volume
surgeons2. Length of
stay is an important determinant of
cost42,43.
Hence, this finding is very relevant to health care today in light of the
pressure placed on hospitals to reduce costs. Additionally, surgeons and
hospitals with higher caseloads are more likely to discharge their patients
routinely. These results imply that high-volume surgeons and hospitals not
only discharge patients earlier but also discharge them with a reduced chance
of postoperative complications that would necessitate transfer of the patient
to another facility.
It is evident from the results that patients treated by high-volume
surgeons or in high-volume hospitals have better outcomes than do patients
treated by low-volume surgeons or in low-volume hospitals. Halm et al., in a
recently published systematic review of studies on the relationship between
volume and outcome, noted that 71% of the studies on hospital volume and 69%
of the studies on physician or surgeon volume reported associations between
higher volume and better health
outcomes44. Other
studies of the volume and outcome relationship with respect to musculoskeletal
disorders2,3,6,7,
coronary artery
bypass8,45,
coronary
angioplasty46,47,
cardiac
transplantation48,49,
cancer
surgery50,51,
gastrointestinal
surgery52,53,
liver
transplantation54,55,
and postoperative wound
infection56 have
described similar findings.
Despite the improvements seen in the findings in the present study compared
with those in previous volume-outcome studies, our analysis has limitations.
First, primary synthetic surgeon identifiers were missing for 41.3% of the
total shoulder arthroplasties and for 45.5% of the hemiarthroplasties.
However, sensitivity analyses and a comparison of the cases of patients with
and without surgeon identifiers indicated that this missing information did
not impact our findings. Several investigations have compared the results
obtained from actual data and from imputation, and they concluded that
imputation is an appropriate method for dealing with missing
data18-23.
Second, the NIS does not provide information with regard to the severity
grading for patients requiring shoulder arthroplasty, which prevents
additional risk-adjusting of outcomes. Third, complications (e.g.,
postoperative wound infections and dislocations) occurring after hospital
discharge cannot be ascertained, even if patients were readmitted to the
hospital. Fourth, clinical outcome indicators on function, strength, range of
motion, and patient satisfaction are not available. Last, there is no evidence
that the coding of diagnoses in the NIS has been validated against clinical
data. However, it is unlikely that miscoding would occur systematically in a
certain volume group of hospitals or surgeons, and thus bias can be assumed to
be minimal.
Our study showed that better outcomes can be achieved for shoulder
arthroplasty when patients are referred to high-volume surgeons and hospitals.
This additional evidence may help in the formulation of health policies to
encourage better outcomes.
Inclusion and exclusion criteria categorized by ICD-9-CM codes, the
algorithm used for case inclusion or exclusion, and tables showing the
bivariate analysis of hospital volume and hospital characteristics, the trend
analysis of surgeon and hospital volume, and stratification of data by
diagnosis (osteoarthritis or fracture) are available with the electronic
versions of this article, on our web site at
(go to the article citation and click on "Supplementary Material")
and on our quarterly CD-ROM (call our subscription department, at
781-449-9780, to order the CD-ROM).