Our study cohort was drawn from the linked SEER-Medicare database. SEER is a collection of population-based cancer registries that collect tumor and clinical information for approximately half of incident cancers in the USA. SEER was designed to capture a diverse population with respect to race, ethnicity, region, and geography. SEER has been linked to Medicare claims and enrollment data through a collaboration between the National Cancer Institute (NCI) and the Centers for Medicare & Medicaid Services (CMS). Over 95% of adults age 65 years and older with incident cancer living in an SEER area have been successfully linked to Medicare claims and enrollment data [21]. Our study used Medicare claims and enrollment data from January 1, 2003 through December 31, 2019 for incident breast cancer cases diagnosed between January 1, 2004 and December 31, 2017.

We identified a cohort of women newly diagnosed with primary stage I–III breast cancer between 2004 and 2017, who were at least 66 years of age at diagnosis. Incident breast cancer cases were identified using ICD-O-3 codes (C500-C509) and month and year of diagnosis variables in the SEER registry data. Since only month and year of diagnosis are provided in SEER, we assigned the first of the month of diagnosis as the diagnosis date. Women were required to be continuously enrolled in Medicare fee-for-service (Parts A and B) with no enrollment in a health maintenance organization (i.e., Medicare Advantage/Part C) for at least 365 days prior to the diagnosis date.

The Faurot frailty index is a Medicare claims-based algorithm that calculates a predicted probability of frailty using demographic, diagnosis code, or procedure code-based frailty indicators that are assessed during a user-specified frailty ascertainment window [14]. The Faurot frailty index was originally developed and validated as a predictor of dependency in the activities of daily living (ADL) using data linkage between Medicare claims and enrollment data and the Medicare Current Beneficiary Survey (MCBS). Claims-based frailty predictors were determined based on clinical expertise and the final choice of predictors and parameterization was determined based on backwards selection and bootstrapping. The Faurot frailty index was externally validated as a predictor of the frailty phenotype using the Atherosclerosis Risk in Communities (ARIC) cohort with Medicare linkage [22] and in the National Health and Aging Trends Study (NHATS) [19]. It was recently updated for use in the International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM) system and was re-validated as a predictor of mortality, hospitalizations, and SNF admissions [23]. A macro to estimate the Faurot frailty index is available online: https://sph.unc.edu/epid/harry-guess-research-community/.

We calculated the predicted probability of frailty on the date of the breast cancer diagnosis using 3-, 6-, 8-, and 12-month frailty ascertainment windows (Figure 1). We also used an “all-available lookback” approach, where the entire history of Medicare fee-for-service continuous enrollment available for each woman prior to the breast cancer diagnosis date was used to estimate the predicted probability of frailty. The earliest availability for claims was January 1, 2003. The predicted probability of frailty was categorized as low (< 0.05), low-medium (0.05– < 0.10), medium (0.10– < 0.20), medium-high (0.20– < 0.40), and high (≥ 0.40) based on cut-points used in prior research [14, 23].

We followed women from the date of the breast cancer diagnosis for up to 1 year to assess incident outcomes relevant to a frail population: all-cause mortality, inpatient admission, and skilled nursing facility (SNF) admission. A one-year period was selected based on the strong association between frailty and short-term outcomes. For the analyses of inpatient and SNF admissions, follow-up time was administratively censored at disenrollment from Medicare fee-for-service and death was considered a competing event. For the analysis of one-year mortality, follow-up time was not censored at disenrollment, since Medicare captures vital status information regardless of enrollment status.

Demographic, clinical, and cancer characteristics were assessed to describe the study population and to conduct subgroup analyses. Demographic factors included age at diagnosis, race (Asian or Pacific Islander, Black, White, Other), ethnicity (Hispanic, non-Hispanic), and census region (Northeast, West, Midwest, South). Comorbidities were assessed during the year prior to the cancer diagnosis using the Gagne combined comorbidity score [24, 25]. Cancer and metastatic cancer were given weights of 0 when calculating the combined comorbidity score. Cancer characteristics included the American Joint Committee on Cancer (AJCC) stage at diagnosis, tumor size, lymph node involvement, tumor grade, and tumor subtype. Tumor subtype was only assessed in patients diagnosed in 2010 or later, since HER2 status was only reliably captured in the SEER-Medicare database beginning in 2010.

We identified 128 857 women with stage I–III breast cancer who had at least 365 days of continuous enrollment prior to breast cancer diagnosis. The median age at diagnosis was 75 (interquartile range: 70, 81) years (Table 1). 87% were White, 8% were Black, 4% were Asian or Pacific Islander, and < 1% were another race. 5% were Hispanic. 45% of women had stage I, 42% had stage II, and 13% had stage III cancer at diagnosis.

Median predicted probability of frailty was similar across ascertainment windows (range 0.03–0.04; Table 1). Mean predicted probability of frailty increased with longer ascertainment windows (3 months: 0.06, all-available lookback: 0.13). The proportion of women classified as having low frailty decreased with increasing duration of lookback, while the proportion classified as having high frailty increased (Table 2).

Five percent of women died within 1 year of diagnosis. Higher predicted probability of frailty was strongly associated with increased one-year mortality risk using each frailty ascertainment window (Figure 2). The difference in one-year cumulative incidence of mortality for individuals with high vs. low frailty was largest using the shortest frailty ascertainment window (28% vs. 2%, difference = 26%, 95% CI 24%–28%) and smallest using the all-available lookback approach (16% vs. 2%, difference = 15%, 95% CI 14%–15%). Discrimination of one-year mortality risk was similar across all frailty ascertainment windows (Figure 3), ranging from 0.758 when using all-available lookback to 0.770 when using a 12-month frailty ascertainment window.

Results from the subgroup analyses are presented in Table 3. The cumulative incidence of one-year mortality differed across subgroups, with older women, Black women, those with higher stage cancer, and those with HR−/HER2+ and triple negative subtypes having a higher cumulative incidence of mortality compared to their counterparts. Across all frailty ascertainment windows, discrimination of one-year mortality risk was similar in subgroups defined by age, race, region, cancer stage, and cancer subtype. Discrimination was good in Hispanic and non-Hispanic subgroups (C-statistics > 0.75), but the C-statistic was lower in non-Hispanic women compared to Hispanic women using all lookback periods.

Higher predicted probability of frailty was associated with higher incidence of inpatient and SNF admissions using all frailty ascertainment windows (Table 4). For example, when using an 8-month frailty ascertainment window, 35% of women with a low predicted probability of frailty had a hospitalization within 1 year of diagnosis, compared to 66% of women with a high predicted probability of frailty. The cumulative incidence differences suggested that associations between the frailty strata and one-year inpatient and SNF admissions were similar using 3-, 6-, 8-, and 12-month frailty ascertainment windows, as measured by cumulative incidence differences. The cumulative incidence differences were lower when using an all-available lookback approach.

Several limitations have implications for the interpretation of these findings. First, the exact date of diagnosis is not available in SEER, so we assigned the first of the month of diagnosis as the index date. This method ensured that the frailty ascertainment window would not overlap with the period following diagnosis when cancer treatments and care may impact frailty and billing practices. However, this may lead to differential misclassification of frailty between those diagnosed early vs. late in the month. Second, we restricted our sample to women with fee-for-service Medicare coverage and our results may not be generalizable to women covered by Medicare Advantage plans, which are becoming increasingly prevalent [34]. Linkage between SEER and Medicare Advantage encounter data recently became available for research purposes [35]. Third, our study included older women with stage I–III breast cancer and our findings may not be generalizable to men with breast cancer, metastatic breast cancer, or other cancer types. In addition, our analysis of inpatient admissions included any hospitalization and may have included hospitalizations for breast cancer treatment (e.g., surgery). Finally, we only evaluated the performance of the Faurot frailty index, and our findings may not be generalizable to other frailty indices derived from administrative data (e.g., the Kim deficit accumulation claims-based frailty index or frailty indices derived from electronic health records) [36–38]. The optimal duration of lookback will vary across indices due to differences in the data structure, the frequency of health encounters and assessments, care settings, and differences in the conditions included in the frailty models.

Frailty is a condition where older adults become less able to cope with illnesses, injuries, and other health challenges. In older women with breast cancer, frailty is linked to a higher risk of dying. In our study, we measured frailty in 128 857 older women with breast cancer using a tool called the Faurot frailty index. This tool measures frailty based on information from healthcare encounters recorded in Medicare data, such as hospital stays or doctors’ appointments. We tested how the length of time used to gather this information affected the tool’s ability to predict who might pass away within a year of their breast cancer diagnosis. Shorter timeframes focus more on recent health issues, whereas longer timeframes provide a broader view of a person’s medical history. We found that women with higher frailty scores, regardless of the timeframe used, were more likely to die within a year of breast cancer diagnosis. Additionally, the Faurot frailty index worked well across different groups of women, including those of various ages, races, and cancer types. Our study shows that the Faurot frailty index is a useful tool for understanding frailty in older adults with breast cancer.