Data sources

The Surveillance, Epidemiology, and End Results (SEER) database (https://seer.cancer.gov) collects cancer incidence data from 17 regional registries, representing an estimated one-quarter of the US population, and is considered the gold standard for data quality among cancer registries in the US as well as globally. We gained access to the SEER database via a standard request and used SEER*Stat 9.0.42.2 software (https://seer.cancer.gov/seerstat) to access the SEER 17 registries research data (2023 submission), in October 2025 [17].

The SEER database regularly collects and updates patient data including demographic characteristics, clinico-pathological characteristics, and patient follow-up information. Researchers need to submit a data access permission application form and download the required data through SEER*Stat software. The SEER database is publicly accessible and comprises of de-identified secondary data. Therefore, patients’ informed consent was not required, and the study was also exempt from review by an ethics committee.

Concurrently, to facilitate comparative assessment with the BC patient population in the SEER database in this study, data were sourced from the Centers for Disease Control and Prevention’s Wide-ranging Online Data for Epidemiologic Research (CDC WONDER) for a standard cohort that represented the US general population [18]. For the standardized mortality ratio (SMR) analysis, the CDC WONDER reference mortality rates were obtained for the 2009–2019 period. This reference period was selected to avoid incorporating the anomalous elevation in cardiovascular mortality observed in the US general population during 2020 associated with the COVID-19 pandemic. The CDC is a government agency of the US Department of Health and Human Services. CDC WONDER publishes the Morbidity and Mortality Weekly Report (MMWR) of the general population of the US and various public health data, including scientific research data, survey supervision data, health statistics and laboratory information based on the death certificates of American residents. Therefore, patients’ informed consent was not required, and use of data was exempt from review by an ethics committee.

Inclusion criteria

This retrospective cohort study employed the following inclusion criteria: (1) female patients diagnosed between 2010 and 2020; (2) the 7th edition of the American Joint Committee on Cancer (AJCC) staging clearly defined as a single primary BC stage I–III at diagnosis; (3) age between 18 and 85 years old; (4) complete follow-up information; (5) receipt of BC-related surgery.

Exclusion criteria

Patients were excluded from the study based on the following criteria: (1) identification of BC diagnosis from death certificate or autopsy; and (2) incomplete information or unknown records. ‘No landmark’ exclusion was applied in the primary analysis; all eligible patients meeting the inclusion criteria were retained and followed from the date of diagnosis.

Covariates definitions

Analyses were conducted for pre-specified subgroups based on clinically meaningful characteristics, including demographic and clinical characteristics.

Demographic characteristics

Patient demographic variables included age at diagnosis (categorized as 18–49, 50–59, 60–69, and 70–85 years old), race (categorized as White, Black, other races), and marital status (categorized as unmarried and married patients).

Clinical characteristics

Clinical variables collected from the SEER data included AJCC stage (I–III), estrogen receptor (ER), progesterone receptor (PR), and tumor human epidermal growth factor receptor 2 (HER2) status (categorized as negative and positive), pathological grade (categorized as Ⅰ, Ⅱ, Ⅲ, and Ⅳ), and treatment forms.

Treatment variables included surgery, chemotherapy, and radiotherapy. Breast-conserving therapy [BCT; less than total mastectomy (ME): segmental ME, lumpectomy, tylectomy, wedge resection, nipple resection, excisional biopsy, or partial ME not otherwise specified (NOS)] or ME (radical or total) were required for patients to be included in this study. Data on adjuvant chemotherapy and adjuvant hormonal therapy were not available in the SEER. Therefore, the chemotherapy and radiation variables were simply classified as yes or no/unknown.

Outcome measures

In this study, the patients were divided into four groups according to their outcomes at the end of follow-up as follows: (1) censored observations (also called “alive group”): alive till the last follow-up or lost to follow-up; (2) BCSM: the outcome of BCSM was defined if their cause of death was classified by International Classification of Diseases, tenth Revision (ICD-10) codes as death due primary malignant BC [19]; (3) CSM: the outcome of cardiovascular-specific death was similarly defined using ICD-10 codes for death due to diseases of the heart, hypertension without heart disease, cerebrovascular disease, atherosclerosis, aortic aneurysm and dissection and other diseases of arteries, arterioles and capillaries [20]; (4) other cause mortality (OCM): other cancer-related deaths or non-cancer-related deaths.

Statistical analysis

The distributions of all baseline characteristics were summarized by calculating the median and range for continuous variables and frequencies for categorical variables, respectively, and the chi-square test were applied to compare the distribution of these categorical demographic and clinical characteristics across the four outcome groups.

The relative risk of CSM was compared to general US female population obtained from CDC WONDER and presented as the SMR, and the 95% confidence intervals (CI) for all SMR were obtained using an approximation based on an assumption of a Poisson regression model [21]. The cumulative incidence curves were plotted for cause-specific deaths overall and stratified by age to show the cumulative incidence of different outcomes in all included patients. Follow-up time for the primary no-landmark analysis was calculated from the date of diagnosis to death or last follow-up. The primary analysis included all eligible patients (n = 255,731), including those who died within the first 3 or 6 months after diagnosis. For the 3-month and 6-month landmark sensitivity analyses, patients who died before the corresponding landmark were excluded, and follow-up was redefined from the respective landmark time (3 or 6 months after diagnosis) to death or last follow-up. Thus, the landmark analyses evaluated subsequent mortality among patients who remained alive at the specified landmark and were intended as sensitivity analyses to assess the robustness of the primary findings to exclusion of early deaths. The 3-month (n = 255,206) and 6-month (n = 253,945) landmark analyses are shown in Tables S1, S2 and S3.

The approximate post-diagnosis time point at which cumulative CSM overtook cumulative BCSM within the 70–85 age stratum was identified by visual and numerical comparison of the non-parametric cumulative incidence function estimates for the two outcomes, taken as the earliest follow-up month from which the CSM cumulative incidence estimate remained above the BCSM estimate through the end of follow-up. The uncertainty of this crossover time was quantified by nonparametric bootstrap resampling with 1,000 replicates.

Finally, Fine-Gray’s competing-risk regression analysis was used to identify the risk factors that significantly correlated with CVD. When treating CVD as the event of interest, death from other causes (including BC) were considered as competing events, and the sub-distribution hazard ratio (SHR) with corresponding 95% CI was acquired. Rather than univariate screening, a prespecified, clinically informed multivariable model was fitted, forcing in all pre-specified covariates, including age, race, marital status, T stage, N stage, ER, PR, HER2, histological grade, surgery type, radiation, and chemotherapy, regardless of univariate significance. This full-model approach is deliberately more inclusive than univariable screening at P < 0.10 or P < 0.15. Since every covariate was forced in irrespective of its univariable P value, no variable meeting such a threshold could have been excluded [22]. As presented in Table S4, collinearity was assessed by variance-inflation factors (all < 2.3), and the proportional-subdistribution-hazards assumption was examined by covariate-by-time interaction. Given the large sample, these interaction tests were highly powered, with marginal departures from proportionality observed. In effect, the SHRs are interpreted as time-averaged effects over the follow-up period. All statistical analyses mentioned above were performed using R software (version 4.5.2). A two-sided P < 0.05 was considered statistically significant.

Patient characteristics

Between 2010 and 2020, 255,731 (82.56%) female BC patients out of 309,750 diagnosed cases qualified for inclusion in our study. There were 216,895 (84.81%) patients who were alive at last follow-up, 17,210 (6.73%) who died of BCSM, 6,250 (2.44%) who died of CSM, and 15,376 (6.01%) who died of OCM. The median follow-up period was 50 months [interquartile range (IQR): 29–72] for patients in the CSM group and 49 months (IQR: 29–70) for those in the OCM group, respectively. The median follow-up period was shorter for BCSM group patients [40 months (IQR: 24–59)].

As Table 1 shows, a majority of patients were White (n = 5,127, 82.03%), unmarried (n = 3,808, 60.93%), diagnosed with BC at age 70–85 (n = 4,401, 70.42%), ER positive (n = 5,371, 85.94%), PR positive (n = 4,654, 74.46%), HER2 negative (n = 5,589, 89.42%), grade Ⅱ (n = 2,894, 46.30%), stage Ⅰ (n = 3,309, 52.94%), T1 status (n = 3,828, 61.25%), N0 status (n = 4,637, 74.19%) in the CSM group. In terms of treatment, most patients underwent BCT (n = 3,516, 56.26%), but did not receive radiotherapy (n = 3,767, 60.27%) or chemotherapy (n = 5,031, 80.50%). Patient characteristics in the OCM group were similar to those in the CSM group.

Some obvious differences can be found in the BCSM group, patients’ age at diagnosis was largely concentrated between 18–59 years old (n = 8,341, 48.46%), a majority of patients were married (n = 8,952, 52.02%), grade Ⅲ/Ⅳ (n = 10,336, 60.06%), stage Ⅱ (n = 7,365, 42.79%), T2 status (n = 7,976, 46.35%), N1 status (n = 5,561, 32.31%), underwent ME (n = 11,128, 64.66%), and received chemotherapy therapy (n = 11,319, 65.77%).

The CSM risk in BC patients and general population

To ensure optimum comparability between the CDC and SEER databases, the analysis was limited to the US-based population. Consistency in baseline characteristics was ensured via the standardization of the selection criteria across datasets, including age, gender, year of diagnosis, and CVD-related causes of death. From 2009 to 2019, the cardiovascular-related mortality rate of the general female population aged 18–85 in the US was 146.8 (/100,000). The reference period was restricted to 2009–2019 to preclude potential distortion of baseline cardiovascular mortality rates attributable to the COVID-19 pandemic, which is well-documented to have caused anomalous elevation of cardiovascular-related deaths in the general US population in 2020 [20]. The cardiovascular-related mortality rates per 100,000 of the general American female population aged 18–49, 50–59, 60–69, and 70–85 were 14.9 (/100,000), 92.4 (/100,000), 231.2 (/100,000), and 1,035.3 (/100,000), respectively.

At the same time, using the data collected from the SEER database in this study, the corresponding patient person-years for the entire group of patients and each age subgroup can be calculated as: 1,576,158.16, 338,596.83, 402,364.08, 458,182.17, and 377,015.08, respectively. The expected numbers of cardiovascular deaths were calculated using indirect standardization: age-group-specific CDC WONDER mortality rates (per 100,000 person-years) were first converted to rates per person-year, then multiplied by the corresponding SEER cohort person-years at risk within that same age stratum; the resulting expected deaths were summed across the four age strata to obtain the total expected deaths. Standardization was performed by age group only; calendar year, attained age during follow-up, and cumulative follow-up time were not separately incorporated into the expected-death calculation.

The overall age-standardized SMR for CSM was 1.16 (95% CI: 1.13–1.19), as displayed in Figure 1 and Table 2. Although the overall excess was modest, the risk of CSM was higher in BC patients than in the general female population across all age groups, with the relative excess greatest in the youngest stratum. The risk of CSM was higher in BC patients across all age groups than it was in the general female population (< 50 years, SMR = 2.894, P < 0.001; 50–59 years, SMR = 1.138, P < 0.001; 60–69 years, SMR = 1.208, P < 0.001; 70–85 years, SMR = 1.128, P < 0.001). The fact that CVD mortality was significantly higher among patients under the age of 50 (SMR = 2.894, 95% CI: 2.461–3.404) likely reflects the low rate of CVD death in the general female population at these ages.

Display full size

Figure 1. The observed event numbers and SMR for overall cohort and age subgroups. CI: confidence intervals; SMR: standardized mortality ratio.

The changing risk of death over time

The hazard curves demonstrated that the risk of BCSM and CSM changed over time, as illustrated in Figure 2. Given that the risk of BCSM grew significantly and peaked within the third year following surgery, it is likely that prognostic factors that contribute to BCSM will have a different effect. Similar to BCSM, the hazard of CSM has not yet peaked and has increased far more slowly. The higher risk of CSM among BC survivors does not become apparent until several years following diagnosis, according to these findings as well.

Display full size

Figure 2. Hazard curves of cancer-specific death and heart-specific death over time. BCSM: breast cancer-specific mortality; CSM: cardiovascular-specific mortality.

Cumulative incidence of all causes of death

The cumulative mortalities for all causes of death among BC patients included in this study are illustrated in Figure 3 and Table 3. By far the highest cumulative mortality of death was caused by BC, followed by OCM. CSM was lower than cumulative mortality due to OCM at all three time points. The cumulative incidences of BCSM, CSM and OCM were 3.09% (95% CI: 3.02–3.16%), 0.84% (95% CI: 0.80–0.87%) and 2.10% (95% CI: 2.04–2.16%), respectively, at the thirty-six months after diagnosis. At 60 months after diagnosis, the cumulative incidences were 5.35% (95% CI: 5.26–5.44%), 1.61% (95% CI: 1.57–1.67%), and 4.06% (95% CI: 3.98–4.14%), respectively. At 84 months after diagnosis, the cumulative incidences were 7.16% (95% CI: 7.05–7.27%), 2.53% (95% CI: 2.46–2.60%), and 6.28% (95% CI: 6.17–6.38%), respectively.

Display full size

Figure 3. The cumulative incidence curves for BCSM, CSM, and OCM by age at diagnosis. BCSM: breast cancer-specific mortality; CSM: cardiovascular-specific mortality; OCM: other cause mortality.

Stratified by age groups, women aged younger had the highest cumulative BCSM compared with CSM and OCM. Nevertheless, it was observed that cumulative CSM steadily increased with age at diagnosis and length of follow-up. These age-stratified patterns are demonstrated in Figure 4. In the oldest age group (70–85 years), cumulative CSM had not yet overtaken cumulative BCSM at 7 years (84 months; 6.98% vs. 7.74%). The two curves crossed later, at approximately 8 years (97 months) after diagnosis. While the 95% CI for cumulative BCSM and CSM in this age cohort are clearly separated at 84 months, they converge and overlap as the crossover is approached, as shown in Table 3. The uncertainty of this crossover time was quantified by nonparametric bootstrap resampling with 1,000 replicates, which placed the crossover at approximately 97 months (95% CI: 91–101 months). This crossover time should therefore be interpreted as an approximate, exploratory landmark rather than a precisely bounded threshold.

Display full size

Figure 4. The cumulative incidence curves for BCSM, CSM and OCM stratified by age at diagnosis (A, < 50 years; B, 50–59 years; C, 60–69 years; D, 70–85 years). BCSM: breast cancer-specific mortality; CSM: cardiovascular-specific mortality; OCM: other cause mortality.

Univariate and multivariable competing-risk regression analysis

Since SEER lacks data on established cardiovascular risk factors such as hypertension, diabetes, hyperlipidemia, smoking, and body mass index, the following associations should be interpreted as factors independently associated with CSM after adjustment for the measured covariates, rather than as fully independent or causal predictors.

As shown in Table 4, in the univariate analysis, age at BC diagnosis, race, marital status, T stage, N stage, ER status, HER2 status, histological grade, radiation, and chemotherapy were substantially linked with CSM (P < 0.05). Through further multivariable analysis, age, race, marital status, T stage, N stage, radiation and chemotherapy remained associated with CSM after adjustment for the measured covariates in the pre-specified model. Progesterone-receptor status was not associated with CSM after adjustment (SHR 0.97, 95% CI 0.90–1.04), whereas surgery type showed a modest association (ME vs. BCT, SHR 0.90, 95% CI 0.85–0.96). Variance-inflation factors indicated no material collinearity (all < 2.3) (Table 4).

In detail, CSM increased most strongly with age, with SHR (95% CI) as high as 2.37 (1.96–2.86), 6.01 (5.06–7.14), and 19.37 (16.37–22.91) for patients aged 50–59, 60–69, and 70–85 relative to patients aged 18–49, respectively. Black patients faced a higher risk [Black versus White: SHR = 1.37 (1.27–1.48)], but other racial groups had a lower risk of CSM [other races versus White: SHR = 0.73 (0.65–0.81)]. Married women have a lower risk of cardiovascular death compared to unmarried women [unmarried patients versus married patients: SHR = 1.59 (1.51–1.68)].

CSM also increased with larger tumor size and more positive lymph nodes [T4 stage: SHR = 1.60 (1.35–1.89), N3 stage: SHR = 1.34 (1.15–1.57)]. Patients receiving chemotherapy and radiation demonstrated a low risk of CSM relative to patients who did not receive chemotherapy or radiation [SHR = 0.59 (0.55–0.64), and 0.64 (0.61–0.68), respectively].

Strengths and limitations

Notably, our study is associated with several strengths. The outcome of the study offers significant statistical power and generalizability, as we employed an extensive registry (SEER) that optimally represents the BC population in the US, on a national scale. Also, leveraging the general population through CDC WONDER for comparative assessment enabled robust standardization of CSM risk stratification using SMR analysis. This study further employed Fine-Gray competing risk regression, enabling the addressing of competing BC-specific risk and other-cause related deaths, which is often missed during studies involving the analysis of BC survivals, despite its significant role in ensuring unbiased assessment of CVD-specific risk among this population. We mitigated confounding from unmeasured baseline CVD comorbidities by limiting the cohort to patients who underwent surgery as an internal control. The primary analysis applied ‘no landmark’ restriction, and hence patients who died within the first months after diagnosis were retained, thus avoiding immortal-time or healthy-survivor selection. To confirm that this choice did not materially influence the findings, we repeated the analysis using 3-month and 6-month landmarks as presented in Tables S1–S3. The SMR, cumulative incidence, and CSM-BCSM crossover remained materially unchanged. Numbers at risk by age group and follow-up interval are provided beneath Figure 4. Follow-up was nonetheless shorter for patients diagnosed closer to 2020, such that estimates in the final months, including the crossover, rest on a progressively diminishing risk set and should therefore be interpreted as exploratory.

Despite the strengths noted, several limitations were identified in the study. Granular data relating to individual chemotherapeutic agents, cumulative drug doses, radiation field and dosing details, or follow-up treatments. However, these data were lacking in the SEER registry. Another noted limitation is that the SEER database lacked data on vital risk factors of CVD, such as diabetes, hypertension, hyperlipidemia, smoking, body mass index, and baseline cardiac function. The study’s findings on absolute and comparative risk estimates may be impacted by these unmeasured confounders. Additionally, the study may be limited by generalizability outside the US population, especially in cohorts characterized by varying healthcare systems, racial compositions, and BC treatment protocols. External validation in differently characterized populations is therefore recommended. Even though the incidence of competing outcomes at individual level may be accounted for by competing risk models, the study may still be associated with possible residual confounding from unmeasured patient variables. Also, careful interpretation is needed for the causal relationship between higher prevalence CVD related deaths and exposure to specific treatment modalities, particularly considering the lack of differentiation of CVD mortality directly consequent to BC treatment exposure from CVD mortality underlying or incidental CVD comorbidities. In addition, cause of death in SEER is based on death certificates coded to ICD-10. It is noteworthy that misclassification of the underlying cause of death is a known limitation of registry-based mortality studies and may have led to under- or over-estimation of the SMR and cumulative incidence estimates of CSM. Furthermore, the SMR analysis compared SEER cohort person-years accrued largely between 2010 and 2020 against CDC WONDER reference mortality rates restricted to 2009–2019. While this restriction was applied to avoid the anomalous 2020 COVID-19-related mortality spike, it introduces a partial temporal mismatch with the diagnosis window of the SEER cohort that could not be resolved through sensitivity analysis within the scope of this study. It is therefore a necessary recommendation that both should be reported in future analyses to further support the identified crossover point.

In conclusion, this large population-based analysis showed that early-stage BC women in the US, particularly those aged 70–85, experienced significantly elevated CSM relative to the age-matched general population, with an approximate, exploratory crossover between cardiovascular-specific and breast-cancer-specific mortality at about 8 years (97 months). Older age, Black race, unmarried status, and advanced T and N stage were independently associated with increased CSM after adjustment for measured covariates. The apparent association of radiotherapy and chemotherapy with lower CSM (SHR = 0.64 and 0.59) is most plausibly explained by healthy-patient selection and residual confounding rather than a cardioprotective effect, given the absence of granular treatment and cardiovascular risk-factor data in SEER. These findings suggest that age-stratified cardiovascular risk assessment may warrant increasing attention during longer-term survivorship, particularly from approximately 5 years after diagnosis and among patients with demographic or tumor characteristics associated with higher CSM in this study. However, the approximately 5-year period should be regarded as an exploratory interval rather than a prescriptive clinical threshold, and the observed associations should not be interpreted as establishing specific referral criteria. Prospective studies and external validation are needed to determine whether these findings can inform clinically actionable timing and risk-stratification strategies.