Comprehensive evaluation of the incidence of late effects in 5-year survivors of breast cancer.

With improving survival of breast cancer patients, patient and provider attention has shifted to the relative occurrence of diseases, conditions, symptoms and complications long after the breast cancer diagnosis [ 1 , 2 ]. These adverse events, sometimes called “late effects,” affect the quality of survivorship and may be avoidable or manageable if they are treated prophylactically or detected and successfully treated before they fully develop [ 2 , 1 ]. Research on the late effects of breast cancer has focused on pre-specified hypotheses pertaining to particular organ systems—such as the cardiovascular system [ 3 ], reproductive system [ 4 ], and bone health [ 5 ]—or pertaining to particular functional consequences, such as loss of shoulder function [ 6 , 7 ] or quality of life [ 8 ]. Alternatively, research has been organized according to the adverse effects potentially related to radiation therapy [ 9 ], chemotherapy [ 10 ], or adjuvant endocrine therapies [ 11 ]. Rare, unusual, or late effects not routinely hypothesized to be associated with breast cancer or its therapies are often investigated only as case reports or case series [ 12 - 14 ]. The Long-term Survivorship in Older Women with Early Stage Breast Cancer study has constructed a cohort of older breast cancer patients treated in integrated health care systems [ 15 ], and a matched cohort of women free of breast cancer [ 16 ]. We have investigated whether the breast cancer survivors are at risk for heart disease [ 17 ], fractures [ 18 ], cancer [ 16 ], and the diseases included in the Charlson index (Jordan et al. , submitted). Given our access to administrative databases containing healthcare information for all members of both cohorts, we recognized the potential to investigate the association between breast cancer survivorship and a comprehensive set of late events. We therefore implemented a hypothesis scanning investigation to identify potential new associations for late effects of breast cancer by determining whether the distribution of these new associations was centered on the null overall and within disease categories, and whether any individual associations merit further investigation. The resulting data achieved these aims, and provide a pilot data resource for other investigations of late events among breast cancer survivors.

We identified a breast cancer survivor cohort which consisted of 1361 Medicare-eligible women 65 years or older, diagnosed with TNM stage I, IIA, or IIB breast cancer between 1 January 1990 and 31 December 1994, who survived five years after diagnosis, and who were cared for in one of six integrated health care systems: Group Health Cooperative, Seattle, Washington; Kaiser Permanente, Southern California, Pasadena, California; Lovelace Health System, New Mexico; Henry Ford Hospital and Health System, Detroit, MI; HealthPartners, Minnesota; and Reliant Medical Group (formerly Fallon Clinic), Massachusetts [ 15 ]. These sites, which are members of the HMO Cancer Research Network (CRN), were chosen to achieve diversity in geography, system size, and patient populations while optimizing study feasibility. At the time the study began, the CRN consisted of the research programs and enrollee populations of 14 integrated health systems with over 11 million enrollees. The overall goal of the CRN is to improve the effectiveness of preventive, curative, and supportive interventions for both major cancers and rare tumors [ 19 ]. We matched members of a breast cancer free comparison cohort to the members of the breast cancer survivor cohort [ 16 ]. Comparison members were selected from women enrolled in the health plan at least one year before the matched breast cancer patient's diagnosis date, to satisfy the same enrollment criterion as applied to breast cancer patients. Comparison cohort members were further matched one to one on the breast cancer cases' age at diagnosis date. The index date for breast cancer patients was the date of their breast cancer diagnosis. The index date for members of the comparison cohort was the date of breast cancer diagnosis for their matched member of the breast cancer cohort. The protocol for this study was reviewed and approved by the Institutional Review Boards for each participating institution. The study was conducted in compliance with U.S. regulations governing the protection and privacy of human subjects. We collected demographic and comorbidity data for members of both cohorts. Data were collected from electronic sources when available and were supplemented and/or collected from the women's medical records. Standardized medical record reviews were conducted at each site by trained medical record abstractors, and the data were entered directly into a computer-based, menu-driven data collection system [ 20 ]. We gathered information on each woman's date of birth, race, and ethnicity for members of the breast cancer cohort from cancer registry databases for sites with cancer registries, from the women's medical records at the sites without cancer registries, and for comparison subjects from medical records. We classified race and ethnicity at all sites using the National Cancer Institute Surveillance, Epidemiology and End Results (SEER) coding instructions for consistency. We identified women of Hispanic ethnicity and grouped the women into the following SEER race categories: white, African-American, American Indian or Alaskan Native, Asian, and native Hawaiian or Pacific Islander. For members of the breast cancer and comparison cohorts, we collected information on prevalent comorbid conditions one year before and five years after the index date, and incident comorbid conditions over ten years of follow-up beginning on the first day of the sixth year after the index date. We used the prevalent comorbidity data to calculate the Charlson Comorbidity Index (CCI) [ 21 ], previously validated in a breast cancer cohort. Data on incident diseases, conditions, and events were collected by harvesting all ICD9 codes that appeared during the ten years of follow-up from the electronic administrative data maintained by the study health plans. These ICD codes were recorded electronically for each patient encounter, including inpatient, outpatient, pharmacy, laboratory, and out-of-plan claims data. We obtained date of death by matching identifying information with the records of the National Death Index. We computed the frequency and proportion of the breast cancer and comparison cohorts within categories of the demographic variables. We restricted the analysis to ICD codes that had no V prefix (indicating an encounter with a disease or injury, but not a current illness) or E prefix (indicating the source of an external injury) and that were present in at least two members of each cohort. For all incident ICD codes included in the analyses, we computed the follow-up time as the time from the first day of the sixth year after index date to the first recording of that ICD code or end of follow up due to death, disenrollment, or completion of ten years of follow-up (year 6 through 15 after index date). Treating each ICD code as an independent event, we computed the hazard ratio (HR) comparing the incidence rate in the breast cancer cohort with the incidence rate in the comparison cohort. Comorbidities known to have an association with breast cancer were excluded from this analysis. One of us (TLL) examined the ICD codes and their hazard ratios and prepared a list of those that were candidates for exclusion because their occurrence was clearly expected to be elevated given their close relation to breast cancer and its sequelae. Another of us (RAS) examined this list of candidates for exclusion, blinded to the hazard ratio, and determined a final list of ICD codes for exclusion. This list, and the initial hazard ratio with its 95% confidence interval (CI), included: (a) ICD9 174 malignant neoplasm of female breast (HR=25, 95% CI=20, 33), (b) ICD9 196 secondary and unspecified malignant neoplasm of lymph nodes (HR= 2.4, 95% CI=1.6, 3.8), (c) ICD9 197 secondary malignant neoplasm of respiratory and digestive systems (HR=2.9, 95% CI=2.1, 4.0), (d) ICD9 198 secondary malignant neoplasm of other specified sites (HR=6.6, 95% CI=4.7, 9.2), (e) ICD9 217 benign neoplasm of breast (HR=1.3, 95% CI=0.80, 2.0), (f) ICD9 233 carcinoma in situ of breast and genitourinary system (HR=4.3, 95% CI=2.7, 6.8), (f) ICD9 611 other disorders of breast (HR=1.9, 95% CI=1.6, 2.3). After excluding the above associations that we expected a priori , we applied semi-Bayes shrinkage methods to the distribution of remaining hazard ratios [ 22 , 23 ]. Semi-Bayes shrinkage narrows the distribution of observed (conventional) hazard ratios and improves their precision by applying a shrinkage estimator derived from regression methods. Associations well above or below the null, but imprecisely measured, are drawn towards the center of the distribution to reduce the potential to take notice of what are likely to be overestimates of the true association. To implement the semi-Bayes shrinkage method, we assumed that 95% of true hazard ratios would fall between 0.25 and 4. In the appendix , we present the frequency of each incident ICD9 code recorded in the breast cancer and comparison cohorts, along with the hazard ratio before and after semi-Bayes shrinkage. We plot histograms of the frequency of occurrence of hazard ratios of varying strength, and cumulative probability plots to show the distribution of the strength of associations overall and within disease categories. Statistical analyses were conducted using SAS version 9.2 (Cary, NC) and Excel® 2010 (Redmond, WA).

Table 1 shows the demographic characteristics of the breast cancer cohort and matched comparison cohort. The two cohorts were well-balanced with regard to the distribution of age categories, race and ethnicity, and prevalent comorbidities. There were 564 ICD9 codes that met our inclusion criteria for this analysis. In the ten years of follow-up that began at the start of six years after the index date, the conventional hazard ratios (HR) ranged from 0.21 (95% C=0.05, 0.96) for other congenital anomalies (ICD9 751) to 5.8 (95% CI=2.0 to 17) for malignant neoplasm of connective and other soft tissue (ICD9 171). After semi-Bayes adjustment, the adjusted hazard ratios (sbHR) ranged from 0.51 (95% CI=0.18, 1.47) for other congenital anomalies of digestive system (ICD9 751) to 3.2 (95% CI=1.89, 5.48) for noninfectious disorders of lymphatic channels (ICD9 457). Figures 1 and 2 show the distribution of the 564 conventional and semi-Bayes adjusted hazard ratios. The median of the conventional hazard ratios equaled 1.06 (95% CI=0.72, 1.6) and the interquartile range was from a hazard ratio of 0.92 (95% CI=0.33, 2.5) to 1.3 (95% CI=1.1, 1.5). The 2.5 th % conventional hazard ratio equaled 0.53 (95% CI 0.18, 1.5) and the 97.5 th % conventional hazard ratio equaled 1.3 (95% CI 1.1, 1.5). As expected, all of the hazard ratios shifted towards the null after semi-Bayes adjustment. The median of the semi-Bayes adjusted hazard ratios equaled 1.06 (95% CI 0.46, 2.45), and the interquartile range was from an adjusted hazard of 0.94 (95% CI 0.72, 1.23) to 1.24 (95% CI 0.87, 1.77). The 2.5 th % semi-Bayes adjusted hazard ratio equaled 0.69 (95% CI 0.37, 1.26) and the 97.5 th % semi-Bayes adjusted hazard ratio equaled 1.89 (95% CI 0.99, 3.59). We compared the distribution of conventional hazard ratios within four disease subgroups, encompassing all included ICD codes, to the overall distribution ( Figure 3 ). The distributions in the infectious diseases subcategory (ICD9 0-139, 680-686, 711) and in the injury and poisoning subcategory (ICD9 800-999) overlapped nearly perfectly with the overall distribution. The distribution in the circulatory system subcategory (ICD9 390-459) was shifted slightly in the direction of negative associations, whereas the distribution in the neoplasms subcategory (ICD9 140-239) was shifted slightly in the direction of positive associations. In the circulatory diseases subcategory, we noted that members of the breast cancer cohort were less likely to have codes for aortic aneurysms (ICD9 441) over the course of follow-up (n=32) than members of the matched comparison cohort (n=63), yielding a conventional hazard ratio of 0.53 (95% CI=0.34, 0.81) and semi-Bayes adjusted hazard ratio of 0.56 (95% CI=0.35, 0.89). In the positive direction, members of the five-year breast cancer survivor cohort were twice as likely (HR=2.1, 95% CI 1.1, 4.2) to have ICD9 code 338 for pain, not elsewhere classified (n=26) than members of the matched comparison cohort (n=13), with semi-Bayes adjusted hazard ratio of 1.89 (95% CI 0.99, 3.59).

The median survival duration after breast cancer diagnosis has continuously improved over the past few decades [ 24 ], as a result of screening-related migration of the stage at diagnosis towards earlier stages and as a result of improving primary and adjuvant therapies [ 25 ]. With this improving survival, patient and provider attention has shifted toward late effects of breast cancer among long-term survivors [ 1 , 2 ]. Using the cohort of older breast cancer survivors, and their matched comparison cohort, we have previously reported on the relative incidence of heart disease [ 17 ], fractures [ 18 ], cancer [ 16 ], and the comorbid conditions included in the Charlson index (Jordan et al. , submitted). Given the administrative setting from which cohort members were identified, we recognized that we could evaluate the complete spectrum of ICD-coded diseases, conditions, and symptoms with marginal effort and cost. We expected this hypothesis scanning exercise to yield new ideas for hypothesis-directed studies of late effects of breast cancer and to provide a data resource of associations that patients and providers could use to evaluate whether incident health effects may have been a late effect related to an earlier diagnosis of breast cancer or its treatment. Our results generated findings worthy of further consideration. For example, members of the breast cancer cohort were less likely to have ICD codes for aortic aneurysms than members of the matched comparison cohort. Breast cancer and its therapies have been associated with both higher [ 3 , 26 ] and lower [ 27 ] risks of heart disease, but we know of no previous report regarding the association between long-term survivorship and the occurrence of aortic aneurysm. As a second example, members of the five-year breast cancer cohort were twice as likely (HR=2.1, 95% CI 1.1, 4.2) to have ICD9 code 338 for pain, not elsewhere classified (n=26) than members of the matched comparison cohort (n=13). Although pain has been identified as a concern in long-term breast cancer survivors [ 28 ], and pain is more common in breast cancer survivors than in comparable women at five-years post-diagnosis [ 29 ], we know of no study that has compared the occurrence of pain in long-term survivors with a matched comparison group. These are examples of new directions for further research suggested by our scanning of 564 hypotheses. Many others, with both negative and positive associations, may draw interest from researchers, physicians, and patients interested in late effects associated with breast cancer and its treatments. Those who use the results to investigate a specific late event should focus on the conventional hazard ratio, whereas those who scan the results to identify new avenues for research should focus on the semi-Bayes hazard ratios, since these are less likely to lead down a false-positive path in the hypothesis-generating setting [ 22 , 23 ]. While these two examples and other results may merit further investigation, many of the associations were centered near the null, with a median of the distribution of 564 conventional adjusted hazard ratios equal to 1.06 and with interquartile range from 0.92 to 1.3. Overall, these results suggest that the health of long-term breast cancer survivors may be very similar to comparable women without a history of breast cancer. If there truly is no difference, then a distribution of hazard ratios above and below the null would be expected, which is approximately what we observed. Alternatively, these results may suggest a slight excess in recording of ICD codes for diseases and conditions among five-year survivors of breast cancer than among their matched comparison cohort. Part of this excess may be attributable to codes for neoplasms, the distribution of which was somewhat shifted toward positive associations compared with the overall distribution. This shift persisted despite the fact that we excluded ICD codes for which we expected and observed elevated hazard ratios likely related to the original breast cancer diagnosis. We note that our earlier hypothesis-driven comparison of the risk of second primary neoplasms in the breast cancer survivors, compared with the matched cohort, showed no overall excess risk [ 16 ]. In the present investigation, some elevated hazard ratios in the neoplasms category may have resulted from miscoding of breast cancer recurrences or from coding of history of breast cancer, particularly in the ambulatory setting. For example, the conventional hazard ratio for malignant neoplasm of uncertain nature (ICD9 239) equaled 2.1 (95% CI 1.7, 2.5). The slight shift toward positive associations for all disease categories cannot be entirely explained by the neoplasms categories, however, since the distributions of conventional hazard ratios for the infectious diseases category (median of 31 hazard ratios equaled 1.05) and for the injury and poisoning category (median of 109 hazard ratios equaled 1.06) were very similar to the overall distribution, and therefore similarly shifted toward positive associations. In contrast, the distribution of conventional hazard ratios for the circulatory system ICD codes was slightly shifted toward negative associations compared with the overall distribution (median of 52 hazard ratios equaled 0.97). This hypothesis scanning investigation has both strengths and limitations. The comparison cohort was matched to breast cancer patients on health plan, enrollment year, and age. Prevalent comorbidities were well balanced in the two cohorts, and there was little loss-to-follow-up aside from that due to deaths. As a result, non-null hazard ratios are unlikely to result entirely from confounding or differential loss to follow-up. We recognized the potential for measurement of multiple associations to generate imprecisely measured, strong estimates of relative effect, and used semi-Bayes methods to address this potential for overestimation. The administrative data used to generate outcomes are not, however, ideally suited for identifying incident diseases and conditions. Some codes likely reflect rule-out diagnoses as opposed to incident conditions, some probably pertain to screening tests rather than to diagnoses, and some likely reflect coding errors. We note, for example, that three members of the all-female breast cancer cohort and five members of the all-female comparison cohort received an ICD code for hyperplasia of the prostate. In conclusion, this study used administrative data to generate new hypotheses for the late effects among older breast cancer survivors. As the size, complexity, and completeness of administrative health databases has grown, opportunities for high-quality hypothesis-directed research have expanded as well [ 30 ]. The marginal effort to harvest additional information, beyond that required for the original aims, is often low. Prudent analysis and interpretation of these hypothesis-scanning datasets can yield new directions for research, such as the potential to investigate previously poorly studied late effects among breast cancer survivors (e.g., aortic aneurysms or pain). However, the centering of the overall distribution of hazard ratios near the null suggests that, beyond the expected sequelae of the initial breast-cancer, there is little difference in the overall health of long-term breast cancer survivors and comparable women without a history of breast cancer.