We found that 15.6% of BC patients were registered with at least one CVD diagnosis in hospital records up to five years before BC diagnosis, and a considerably higher percentage of patients had a history of CVD drug prescriptions (i.e. drugs related to CVD disease management). Overall, BC patients and cancer-free controls were similar with regards to CVD comorbidity at index date. The most common types of CVD diagnoses in both BC patients and controls were hypertension, ischemic heart disease/acute coronary syndrome, and cerebrovascular disease.
After BC diagnosis, the incidence of CVD was significantly higher in BC patients than controls for all CVD diagnoses combined up to approximately 6 years after the index date (BC diagnosis) taking account of the presence of competing risk of death. After 10 years, 28% of both BC patients and controls (without any CVD diagnosis in NPR five years before the index date) had at least one CVD diagnosis in NPR. Only the cumulative incidence of heart failure, thrombophlebitis/thrombosis and pulmonary heart disease including pulmonary embolism remained higher in BC patients than controls during the entire 10-year follow-up period. Furthermore, the risk of heart failure and thrombophlebitis/thrombosis was significantly higher in BC patients who had received chemotherapy compared to BC patients who had not as well as cancer-free controls. Heart failure has been associated with chemotherapy in previous studies , and it is well established that cancer patients have increased risk of thrombotic complications including e.g. deep vein thrombosis, pulmonary embolism and arterial thrombosis .
Contrary to our results, Abel-Qadir et al.  found that the prevalence of different types of CVD was significantly higher in BC patients compared to cancer-free controls. A possible explanation for this discrepancy could be that Abel-Qadir et al. used different algorithms to determine the presence of CVD preceding the index date. Other possible explanations are differences in the definition of the study population and matching criteria, as Abel-Qadir et al. included women diagnosed with early stage BC only, and controls were matched to BC patients according to gender and age, but not education as in our study.
Abel-Qadir et al.  found a significantly higher incidence of CVD hospitalizations due to heart failure, cerebrovascular disease and arrhythmias in BC patients compared to cancer-free controls whereas we did not find this for cerebrovascular disease or arrhythmias. In accordance with our results, Strongman et al.  found a statistically significantly higher incidence of heart failure and venous thromboembolism in BC patients compared to cancer-free controls, but not significant differences in the incidence of arrhythmia, stroke or peripheral vascular disease. Furthermore, Strongman et al. found that the incidence of coronary artery disease was significantly lower in BC patients compared to controls.
Existing studies have shown an elevated risk of CVD in BC patients after anthracycline-based chemotherapy, radiation therapy and antibody therapy [11, 26–28]. Our study confirms an elevated risk of CVD after chemotherapy. Unfortunately, data available for our study did not allow distinction between different types of chemotherapy. We also saw a tendency towards a higher risk of CVD after radiation and antibody therapy, but these results were in general not statistically significant.
Hormonal agents like tamoxifen approved for BC treatment more than 30 years ago can have both beneficial and detrimental effects on the cardiovascular system . Studies have shown that tamoxifen has a protective effect on lipid metabolism [29–31], but at the same time increases the risk of venous thrombosis and thromboembolism . In our study, we did not find a higher risk of CVD after hormonal therapy, but we see a higher incidence of thrombotic complications in BC patients compared to cancer-free controls.
The present study has several strengths. Firstly, it is a study at the population level based on a large real-world dataset with no selection bias or other potential issues related to more segmented and selective populations. Secondly, the study population was identified from CAR, which has high completeness and validity due to use of notifications from different data sources and quality control . Thirdly, we used prescription data from NPRD to identify CVD patients as a supplement to data from NPR, as the latter registry does not include patients who require primary care only. NPRD is the national registry of prescription drugs dispensed from community pharmacies and is considered both complete and valid from 1995 and onwards . Fourthly, the study included a cancer-free control group, and the use of exact matching ensured that BC patients and controls were comparable with regards to age, region of residence, and education as well as gender (they were all women). We did not include CCI as a matching criteria to avoid overmatching, and we preferred exact matching to propensity score matching because exact matching ensures that groups are identical with regard to the matching criteria used. Finally, we allowed for the presence of competing risk of death when estimating the cumulative incidence function [22, 33]. Since BC patients had a higher risk of dying during the 10-year follow-up period than controls, ignoring the presence of competing risk could result in substantial bias.
The study also has limitations. The study relies on diagnosis and procedure codes in NPR, which may contain errors or be incomplete, e.g. many doctors code the primary diagnosis only, even if the patient has other relevant diseases . Furthermore, NPR does contain information on patients who require primary care only as mentioned above, and it is not straightforward to identify CVD patients based on NPRD data as prescription drugs used for CVD disease management may have other indications. Finally, there is a risk of confounding due to the observational nature of the study. The risk of confounding related to observable baseline characteristics was minimised by exact matching and a multivariable regression design, but BC patients and controls may differ on non-observable lifestyle factors, that influence the risk of CVD. However, we expect lifestyle factors to be similar in BC patients and controls because education is a proxy for lifestyle and was included as a matching criteria. Similar comorbidity at index date also indicate a comparable lifestyle between BC patients and controls.