Internal Review Board approval was obtained from the Human Subjects Division of the University of Washington (IRB 37473)(15
). Data for this analysis came from a linkage between the Surveillance Epidemiology, End Results (SEER) database provided by the National Cancer Institute (NCI) and Medicare healthcare claims records provided by the Center for Medical Services (CMS)(16
). SEER registries identify 97% of all incident cancer cases among persons residing in SEER regions (14% of the US population in 1995 and 26% in 2005) (17
), and 93% of persons in these registries over 65 had Medicare data successfully matched to SEER records in the linkage process(16
This study identified all women over the age of 65 in the SEER-Medicare database diagnosed with ovarian cancer from January 1, 1995 to December 31, 2005. Women were included if they had American Joint Cancer Committee (AJCC) stage III or IV ovarian cancer (n=13,998). Women were excluded if they had a diagnosis based on autopsy or death certificate only, non-invasive pathology, disease that was not pathologically confirmed, non-epithelial malignancies, or they had a second primary malignancy diagnosed any time in the six months before or after the date of the ovarian cancer diagnosis (1488 excluded). Women had to be continuously covered by Medicare parts A+B and not be enrolled in an HMO from the 12 months prior to diagnosis and at least 9 months after diagnosis (4264 excluded). This study was further limited to the 5475 women from the above cohort who had evidence in Medicare records of a debulking surgery for their ovarian cancer. The first episode of cytoreductive surgery for ovarian cancer in the year after the diagnosis date was identified as defined below from the Medicare records which were available for claims through December 31, 2007.
SEER data was used to identify and categorize age (5 year groups), race (white, black or other) and marital status (married or unmarried). SEER registries were grouped according to geographic region (Northeast, Midwest, South or West). Population density of area of residence was categorized as defined in the SEER files. Median household income from zip code of residence was used as a proxy for socioeconomic status and was derived from 2000 census data included in the SEER files (categorized into quartiles low (1
) to high(4
)). Tumor stage, grade and histology were determined from SEER. Tumor grade was missing for over 20% of the subjects and so was not utilized as a variable in the analysis. Comorbidity score was determined using claims for the 12 months prior to ovarian cancer diagnosis to calculate the Deyo adaptation (18
) of the Charleson comorbidity index (19
). One point is assigned for evidence of each of the following: dementia, congestive heart failure, coronary artery disease (heart attack, angina or revascularization), diabetes, hypertension, peripheral vascular disease, pulmonary disease, renal disease or stroke. Two points are assigned for previous malignancy and three points assigned for hepatic disease.
Hospital volume was determined as the number of cases from January 1, 1995-December 31, 2006 for hospitals that were located within SEER registry areas as indicated in the hospital files in the SEER-Medicare database. Hospital volume was categorized by the number of ovarian cancer cases over the study period as described by Schrag et al. as low (1
), intermediate (13
) and high (>28)(21
). Only women whose surgeries occurred at hospitals located in SEER areas were included in the multivariable analysis that accounted for treatment variables to ensure that all Medicare cases for that hospital were identified. While this volume does not include non-Medicare cases, it has been shown to correlate with overall hospital volume and allow ranking of hospitals into volume categories(22
). Ovarian cancer surgeon was determined by the use of a unique provider identification number when available on provider claims (when available) associated with ovarian cancer surgeries and surgeon volume was categorized into low (1
), intermediate (5
) and high (>25). Provider specialty was determined from both Medicare files and AMA files(23
). Surgeon specialty was categorized as gynecological oncologist, gynecologist, surgeon or other/unknown.
Admission type was identified in 97.5% of the patients from the inpatient hospital billing records for the surgical episode. Patients were categorized as having an emergent admission if they were admitted through the emergency department or if the admitting physician indicated on the billing claim that the admission was an emergency.
Surgical treatment for ovarian cancer was identified in the MEDPAR files using ICD-9 procedure codes and in the physician claims using CPT codes indicating surgical resection of the primary tumor as previously described(24
). Complexity of the primary surgery may influence 30-day mortality. Identification of upper abdominal procedures at the time of the primary surgery was performed by searching for International Classification of Disease-9 codes (ICD-9) in the inpatient billing records. Patients were classified as having an upper abdominal procedure if they were noted to have a liver (50.22, 50.3), diaphragm (34.81), spleen (41.2, 41.3, 41.5) or pancreatic (52.5, 52.6) resection. Large bowel resections were identified from ICD-9 codes (45.52, 45.7 45.8, 45.92, 45.95, 45.93, 45.94, 48.4, 48.5, 48.6) of the inpatient records. Chemotherapy was identified as previously defined if either the inpatient record, outpatient file or physician claims indicated that chemotherapy was given(15
). Chemotherapy was classified as neoadjuvant if administered prior to the date of the primary surgical episode. Previous studies have determined a high level of agreement between Medicare data and chart review in the identification of surgery and chemotherapy among cancer patients (25
The primary outcome in this study was 30 day mortality after surgery, defined as death from any cause in the 30 days after the primary surgical episode. Death was identified from the Medicare records which is verified with the social security administration and captures all patient deaths regardless of location (in hospital, home, hospice etc).
The chi-squared test was used to compare the frequency distributions of categorical variables. All analyses were stratified based on the admission type and all models empirically included year of diagnosis as a confounding variable to account for possible temporal changes. Because the outcome of interest was not rare, a Poisson regression was used to model incident rate ratios which were interpreted as a relative risk for the outcome of interest. Outcomes at a particular institution may be related to unmeasured factors from the individual institution and thus as described previously a generalized estimating equation was used to account for clustering by hospital(27
). Models were fit using generalized estimating equations with a Poisson family, a log link and the hospital identifier as a clustering variable. Variables of interest were classified as either patient (age, race, median household income, marital status, geographic region, size of area of residence, stage, histology and co-morbidity score) or treatment (hospital and surgeon volume, surgeon specialty, upper abdominal procedures, large bowel resection, neoadjuvant chemotherapy) related. The first model fit included all patient related variables significantly associated with 30-day mortality on univariable analysis. The second model included all variables in the first model and the treatment related variables found to be significant on univariable analysis. Model fit was assessed by the use of generalized Pearson residuals. All p values are 2 sided and a p<0.05 was considered significant. No statistical corrections were made for multiple comparisons. STATA SE version 11.0 (College Station, TX) was used for all calculations.
Lastly a decision tree was then constructed for women with routine admissions. Because we aimed to identify women who may benefit from alternative treatment strategies such as neoadjuvant chemotherapy, women who already received neoadjuvant chemotherapy were excluded from this tree (n=605). Similarly only women admitted routinely were included in this tree, as it was hypothesized that women admitted emergently were much more likely to have bowel obstructions and other acute symptoms that would dictate the need for immediate surgery, regardless of the surgical mortality. Only variables that were significant in the multivariable model were included and we limited variables in the analysis to the patient related variables as these were easily ascertained pre-operatively and objective. Age was dichotomized at age 75 (65-75 vs. 75+), stage was classified as III or IV and co-morbidity score was categorized as 0-1 vs. 2+. We were then able to categorize women into risk groups based on the observed 30 day mortality in these subgroups.