• 2022-05
  • 2022-04
  • 2021-03
  • 2020-08
  • 2020-07
  • 2018-07
  • METHODS br Surveillance Epidemiology and End Results Program Medicare database


    Surveillance, Epidemiology, and End Results Program-Medicare database
    The National Cancer Institute’s Surveillance, Epidemi-ology, and End Results (SEER) tumor registry linked to the Medicare database was used for this DETA NONOate study. The SEER-Medicare data link 2 national databases to provide detailed information about Medicare beneficiaries with cancer. To link SEER with Medicare data, the registries participating in the SEER program send individual identi-fiers for all persons in their files. These identifiers are matched with identifiers contained in Medicare’s master enrollment file. For each of the linkages, 95% of persons aged 65 years or older in the SEER files are matched to the Medicare enrollment file.5 Quality control is an impor-tant component of the SEER program. The current stan-dard for accuracy of data is an error rate <5%. Data for the years 2004 to 2012 was requested for this study. As a population-based study with no patient identifiers involved, our investigation was exempt from IRB approval. 
    The Patient Entitlement and Diagnosis Summary Files contain the data elements collected by the SEER registries, as well as information pertaining to Medicare eligibility and enrollment. Cancer-related variables in Patient Enti-tlement and Diagnosis Summary Files include demo-graphic characteristics, previous cancer diagnoses, date of cancer diagnosis, stage at diagnosis, and date of death if applicable. The National Claims History, Outpatient Claims, and Medicare Provider Analysis and Review files were used to find diagnosis codes for each patient up until the date of cancer diagnosis given in the Patient Entitle-ment and Diagnosis Summary Files.
    Case selection
    Case selection was performed by identifying patients with a biopsy-proven diagnosis of invasive ductal adenocarci-noma (ICD-9 histology codes 8500, 8521, 8230, 8522, 8523, and behavior code 3) who underwent operation for the primary tumor between 2004 and 2012. Our cohort was restricted to patients with stage I disease (as defined by the American Joint Committee on Cancer’s Cancer Staging Manual, 7th edition) who were enrolled in Medicare during the year of their diagnosis.
    The exclusion criteria ruled out ductal carcinoma in situ, stage other than I or unstaged cancer, cases identified by au-topsy only, and patients not enrolled in both Medicare parts A and B or without enrolment in an HMO. Patients aged younger than 65 years enrolled in Medicare because of end-stage renal disease or chronic disability also were excluded from the study.
    The cause of death variable code was used to identify our 3 primary outcomes of interest: death due to breast cancer (DOD), death due to other cancers (DOC), and death due to non-cancer causes (NCD). The cause of death var-iable in SEER is derived from death certificates, and although concerns have been expressed about the accuracy and use of this variable, it has been shown to be a valid estimate of cancer-specific survival.6,7 For all survival ana-lyses, time zero was the date of diagnosis of breast cancer. Last follow-up is the date of the last time the patient was seen in a clinical setting for which Medicare was charged, that is, the final Medicare claim.
    Neural network analysis
    Using ICD-9-CM diagnosis and procedure codes, relevant comorbid conditions in all medical claims (inpatient, outpatient, and physician claims) for the index breast can-cer claim were identified for each patient. To produce clin-ically meaningful analyses, the comorbidities were grouped together into distinct categories associated with mortality
    32 Wasif et al Competing Risk of Death in Breast Cancer J Am Coll Surg
    outcomes of interest and based on earlier work in the field.8 After grouping the diagnoses codes, a neural network model was used to explore the predictive ability of these diagnosis groups on the mortality modes, with each mor-tality mode being an output node classification. This approach was used to recognize underlying relationships in the data without making a priori assumptions about comorbidities and the association with mortality. Because there were 3 different outcomes of interest and multiple co-morbidity groupings, the associations between the 2 groups could best be explored by using this technique rather than traditional multivariable regression analysis. The input var-iables for the neural network analyses were the diagnosis codes and the output variables; the 3 mortality outcomes of interestdDOD, DOC, and NCD (Fig. 1).
    Statistical analyses
    Output from the neural network analysis was used to identify comorbidity groupings associated with our mortality out-comes of interest, which were then used in the subsequent analyses detailed here. The association of age, race, comor-bidity groupings, and tumor variables with mortality risk was studied using Fine and Gray multivariable competing risk regression to predict the probability of DOD, taking into account the competing risks of DOC and NCD. Competing risk analysis is a special form of survival analysis in which a competing risk (dead of other cancer, non-cancer death) is an event whose occurrence precludes the occurrence of the primary event of interest (DOD).9 Variables included in both models were age, race, psychiatric disorders, injuries, neurologic disorders, infectious disorders, neoplasms, car-diovascular disorders, and estrogen receptor (ER) status. The assumptions of the Fine-Gray model were tested using