Pediatric Cardiac Surgery: The Effect of Hospital and Surgeon Volume on In-hospital Mortality Edward L. Hannan, PhD* Michael Racz, MS* Rae-Ellen Kavey, MD��� Jan M. Quaegebeur, MD, PhD�� and Roberta Williams, MDi ABSTRACT. Objective. To examine the relationship between annual provider (hospital and surgeon) volume of pediatric cardiac surgery and in-hospital mortality. Design. Population-based retrospective cohort study using a clinical database. Setting. The 16 acute care hospitals in New York with certificate of need approval to perform pediatric cardiac surgery. Patients. All children undergoing congenital heart surgery in New York from 1992 to 1995. Main Outcome Measures. Risk-adjusted mortality rates for various hospital and surgeon volume ranges. Adjustments were made for severity of illness using logistic regression. Results. A total of 7169 cases were analyzed. After controlling for severity of preprocedural illness using clinical risk factors, hospitals with annual pediatric car- diac surgery volumes of fewer than 100 had significantly higher mortality rates (8.26%) than hospitals with vol- umes of 100 or more (5.95%), and surgeons with annual volumes of fewer than 75 had significantly higher mor- tality rates (8.77%) than surgeons with annual volumes of 75 or more (5.90%). Conclusions. Both hospital volume and surgeon vol- ume are significantly associated with in-hospital mortal- ity, and these differences persist for both high-complex- ity and low-complexity pediatric cardiac procedures. Pediatrics 1998 101:963���969 pediatric cardiac surgery, mortality, hospital volume, surgeon volume. ABBREVIATIONS. CSRS, New York Cardiac Surgery Reporting System CAC, Cardiac Advisory Committee. Nbetween umerous studies in the past 2 decades have documented significant inverse relationships adverse outcomes for certain types of patients and the amount of experience providers have in treating those patients.1���8 Generally, patients have been identified according to the type of proce- dure they underwent or their medical condition (principal diagnosis), with surgical examples being more frequent. Also, provider volume has been mea- sured both on the hospital and the physician/sur- geon level, with some studies investigating both vol- ume measures. For the most part, in-hospital or short-term mortality has been used as the measure of adverse outcome, although complications of treat- ment and hospital length of stay have also been used. The more sophisticated studies have attempted to investigate volume-outcome differences after having adjusted for patient severity of illness using various demographic and diagnostic indicators of severity. The reader is referred to the book by Luft et al9 for a thorough description and review of the methods used, assumptions made, procedures and medical conditions investigated, and current areas of re- search in this large body of work. Only one study in the literature has investigated the relationship between adverse outcomes and the volume of pediatric cardiac surgery. Probably one of the reasons for this is that pediatric cardiac surgery involves a myriad of different procedures, involving different surgical challenges and entailing a wide range of risks to patients. Also, the annual volumes for each of these individual procedures tend to be quite low, even when accumulated across an entire state or large geographical region. The purpose of this study is to examine the rela- tionship between in-hospital mortality and provider (hospital and surgeon) volume for pediatric cardiac surgery in New York State between 1992 and 1995. This study extends the earlier study conducted by Jenkins et al1 in a few respects. First, both hospital volume and surgeon volume are available as mea- sures of provider volume, whereas Jenkins et al had access only to hospital volume. Second, clinical data from New York���s Cardiac Surgery Reporting System (CSRS) are available for conducting risk-adjustments rather than having to rely on administrative data collected for other purposes. Third, in addition to investigating aggregate differences in risk-adjusted mortality for various provider volume groups, this study examines risk-adjusted mortality rate differ- ences between provider volume groups for different levels of procedure complexity. DATA AND METHODS Data The database used in the study is the part of New York���s CSRS dedicated to pediatric cardiac surgery. The CSRS was initiated in 1989 by the New York State Department of Health and its Cardiac Advisory Committee (CAC) for the purpose of improving quality From the *Department of Health Policy, Management, and Behavior, State University of New York, University at Albany, Albany, New York the ���Department of Pediatric Cardiology, SUNY Health Science Center, Syra- cuse, New York the ��Department of Surgery, Columbia-Presbyterian Hos- pital, New York, New York and the iDepartment of Pediatrics, University of North Carolina, Chapel Hill, North Carolina. Received for publication May 30, 1997 accepted Oct 16, 1997. Reprint requests to (E.L.H.) University at Albany, State University of New York, Department of Health Policy, Management, and Behavior, One Uni- versity Place, Rensselaer, NY 12144���3456. PEDIATRICS (ISSN 0031 4005). Copyright �� 1998 by the American Acad- emy of Pediatrics. PEDIATRICS Vol. 101 No. 6 June 1998 963

of care in New York. The CAC is a group of cardiac surgeons, cardiologists, health services researchers, and consumers charged with advising the Department of Health on issues related to the quality of and access to cardiac care, and the prevention of coro- nary heart disease. The CSRS contains information on patient demographics, types of procedures performed, risk factors/diag- noses, complications of surgery, discharge status, and surgeon and hospital identifiers. The information is collected concurrently under the direction of each hospital���s director of cardiac surgery. Completeness of registry data is assured by matching it with New York���s administrative database, the Statewide Planning and Re- source Cooperative System. Data accuracy is assured by conduct- ing medical record reviews in each hospital. The quality of coronary artery bypass graft surgery has been improved by identifying significant risk factors related to coronary artery bypass graft surgery, risk-adjusting adverse outcomes using these factors, computing risk-adjusted mortality rates for surgeons and hospitals, and providing this information to hospitals, surgeons, and the public.10,11 This has not been done to date for other proce- dures in the database (eg, valve surgery, pediatric cardiac surgery) primarily because these procedure groups are not as homogenous and have much lower volumes. Although pediatric cardiac surgery data have been available since 1989, the set of risk factors in the database was expanded substantially in 1991. The data in this study comprises all pediatric cardiac surgery performed in New York State between 1992 and 1995 that is reportable in the state���s pediatric cardiac surgery registry mentioned above, a total of 7169 procedures. These pro- cedures are performed in the 16 hospitals in New York that have certificate of need approval. The pediatric cardiac surgery report form that is part of the CSRS contains information on patient demographics (age, date of birth, sex, ethnicity, race) admission, procedure, and discharge dates primary diagnosis and procedure codes mode of cardio- pulmonary bypass weights at birth and at time of operation number of previous open heart and closed heart operations pre- vious catheter interventions 11 other risk factors and a variety of complications after surgery. The procedures used in the study were identified and defined by the CAC so as to accurately characterize the treatments ob- tained and the risks incurred by patients. They are identified using special codes created by the CAC that are similar to, but do not map on a one-to-one basis with, ICD-9-CM codes. Complexity Categories In CSRS, procedures are classified according to recommendations from the CAC. As mentioned above, one of the challenges in exam- ining volume-outcome relationships in pediatric cardiac surgery is that there is a wide variety of procedures performed, with no one procedure having a sufficient volume in most databases to sustain complicated multivariate analyses. Consequently, the best analysis option is the one used by Jenkins et al,1 which is to combine proce- dures into groups that are as homogenous as possible with respect to patient severity of illness and to use the groups as severity measures (risk factors) in the risk-adjustment process. The method used in this study to define complexity categories consisted of the following steps: (1) order the procedures accord- ing to in-hospital mortality rate from low to high, (2) look for natural breaks in the contiguous mortality rates so that between three and five groups could be identified (it was decided that this was a good range to use so that enough groups would be available for describing severity of illness differences while maintaining a sufficiently large number of cases in each group), (3) have the pediatric cardiac surgeon (J.Q.) and the two pediatric cardiologists (R.E.K. and R.W.) revise the categories so that similar procedures would be contained in the same category even if they had some- what dissimilar mortality rates, and to reflect procedural complex- ity not necessarily captured by mortality rates in the database (perhaps because of low volumes), and (4) repeat steps 1 and 2. Four procedure complexity categories were identified. Methods The first step in the analysis consisted of calculating the fre- quencies and mortality rates associated with each of the demo- graphic and diagnostic risk factors contained in New York���s clin- ical pediatric cardiac surgery database. Later, these analyses were expanded to include frequencies and mortality rates for each of two hospital volume ranges (,100 pediatric cardiac procedures annually, 100 or more procedures annually). Significance of the various categorical variables (eg, sex, race, binary risk factors such as congestive heart failure) was determined using x2 tests. Next, a stepwise logistic regression model was constructed to determine which of the potential risk factors were significant predic- tors of in-hospital mortality, and how to predict mortality on the basis of those risk factors. The dependent variable was binary, with a 1 denoting in-hospital mortality and a 0 denoting a live discharge. The candidate independent variables were the demographic and diagnostic risk factors described above, and the four complexity categories. Independent variables were retained in the model if they were significant in the stepwise analysis (P , .05). Complexity categories were tested in the model by using the group with the lowest complexity as a reference group, and the other three groups as binary risk factors in the model. Age was tested as a categorical variable by splitting it into ranges that were defined after examining their bivariate relationship with mortal- ity. The ranges ,7 days, 7 to 29 days, 30 to 89 days, 90 to 179 days, 180 to 359 days, and 360 days or more were used. Also, after a set of significant independent variables was identified, two-way in- teractions among these variables were tested in a new stepwise model. As a test of the adequacy of the complexity categories, observed and expected (using the model) mortality rates were compared for each of the complexity categories. After the final statistical model was identified, hospital and surgeon volume measures associated with each case were com- puted as the number of pediatric cardiac procedures performed in that calendar year in the hospital the procedure was performed and by the surgeon performing the procedure, respectively. Then, risk-adjusted mortality rates were calculated for different hospital and surgeon volume groups in an attempt to determine which split that created high-volume and low-volume groups yielded the largest differential in risk-adjusted mortality rates between the two groups while maintaining reasonably large volumes in the two groups. Splits at 100 procedures annually in hospitals and 75 procedures annually for surgeons were obtained. As a confirma- tion that the two volume measures were significant predictors of mortality, they were also tested by adding them to the logistic regression model described above, and testing their significance in the model. Another test consisted of using the average 4-year hospital and surgeon volumes in lieu of annual volumes to deter- mine if the conclusions changed substantially (they did not). In computing risk-adjusted mortality rates, the first step con- sisted of computing the expected mortality rate for the volume groups by summing the predicted probabilities of death for each patient in the group using the logistic regression model, and then dividing by the number of patients. For each volume group, this rate was divided into the observed mortality rate (number of deaths/number of patients), and then multiplied by the overall mortality rate for all pediatric surgery patients to obtain the risk- adjusted mortality rate for the group. This rate represents the best estimate of what each hospital volume or surgeon volume group���s mortality rate would have been if it had had an average severity of illness that was the same as that of the state as a whole. The rate for each group was then tested to determine if it was (statistically) significantly higher or lower than the statewide rate by calculating confidence intervals for risk-adjusted rates.12 The next set of analyses was aimed at determining if there were significant differences in risk-adjusted mortality rates for hospital volume groups and for surgeon volume groups by complexity category. Because the analysis used complexity categories rather than the entire data set (and this reduced the sample sizes), the relationship between each of the two provider volume measures and risk-adjusted mortality was examined separately rather than observing intersections of the volume measures. The purpose of these analyses was to determine if the relationship between mor- tality rate and provider volume was limited to the more complex procedures, or if it persisted across all complexity categories. In the next set of analyses, four hospital volume/surgeon vol- ume groups were obtained by splitting patients on the basis of their hospital and surgeon volumes (hospital volume ,100/sur- geon volume ,75, hospital volume ,100/surgeon volume .75, and so forth). Then, observed and risk-adjusted mortality rates were calculated for each intersection of hospital volume group and surgeon volume group to determine whether there were significant differences in risk-adjusted rates among the resulting 964 PEDIATRIC CARDIAC SURGERY MORTALITY