2.1 Source of data

Data were collected from the Clinical Data Analysis and Reporting System (CDARS), a population-based administrative database that prospectively records basic demographics, diagnoses, laboratory tests and medical procedures of all patients attending public hospitals and clinics in Hong Kong, which cover the entire territory of Hong Kong and serve an estimated 90% of the population.⁶ Diagnoses are encoded by the International Classification of Diseases, Ninth revision (ICD-9) codes regardless of the time of data input, as ICD-10 codes have not been implemented in CDARS to date. Mortality data were collected from the linked Hong Kong Death Registry, a governmental registry of all Hong Kong citizens' death records. Both databases have been used extensively in research, and have been shown to have good coding accuracy and data completeness.^7-13

2.2 Patient population

Adult (aged 18 years old or above) patients with PCa who received ADT between 1 January 1993 to 31 March 2021 were identified. ADT included medical castration and bilateral orchiectomy (BO). There were no exclusion criteria. This cohort has been published earlier.¹⁴

The following baseline characteristics were recorded: age, type of androgen deprivation therapy, comorbidities (hypertension, diabetes mellitus, dyslipidemia, chronic kidney disease, chronic liver disease, stroke, myocardial infarction (MI), heart failure (HF) and arrhythmias (atrial fibrillation, ventricular tachycardia or ventricular fibrillation)) and use of medications or prior procedures (radiotherapy, chemotherapy, radical prostatectomy, androgen receptor signaling inhibitor [ARSI], dihydropyridine calcium channel blockers, metformin, sulphonylurea, dipeptidyl peptidase-4 inhibitors, glucagon-like peptide-1 receptor agonist, insulin, antiplatelet, anticoagulant and corticosteroid).

2.3 Follow-up and outcomes

All patients were followed up from the date of ADT initiation until 31 September 2021, death or the primary endpoint, whichever occurred earlier. The primary endpoint was a composite of cardiovascular hospitalization and cardiovascular mortality. Cardiovascular hospitalizations were identified by ICD-9 codes listed in Table S1, while cardiovascular mortality was identified by ICD-9 or ICD-10 codes listed in Table S2.

2.4 Major cardiovascular comorbidities and their relative impact

We focused on prior history of HF, MI, stroke and arrhythmia as the major cardiovascular comorbidities of interest. These comorbidities were chosen among broader cardiometabolic conditions, such as hypertension and diabetes mellitus, as (a) we had shown that these accounted for >75% of cardiovascular hospitalizations in patients with PCa receiving ADT,¹⁴ (b) they are clinically more severe conditions that are likely to lead to substantially worse morbidity and (c) the inclusion of broader cardiometabolic comorbidities would have created too many combinations of comorbidities with limited sample sizes, preventing any meaningful analysis. Patients were categorized into the following groups by the presence of these major cardiovascular comorbidities for comparisons: none of HF/MI/stroke/arrhythmia, HF only, MI only, stroke only, arrhythmia only and ≥2 of HF/MI/stroke/arrhythmia. To explore the effects of stroke subtypes, we carried out a post hoc analysis in which patients with stroke only were split into those having had only ischemic stroke, only hemorrhagic stroke or both ischemic and hemorrhagic stroke.

As these comorbidities commonly overlap, we further explored the relative impact of each major cardiovascular comorbidity in those with overlapping comorbidities (ie, with ≥2 major cardiovascular comorbidities). This was done by comparing patients with both a condition of focus and any of the “nonfocus” conditions against patients with ≥2 of the “nonfocus” conditions. For instance, to explore the relative impact of HF, we compared patients with HF and any of MI/stroke/arrhythmia against those with ≥2 of MI/stroke/arrhythmia.

2.5 Statistical analyses

Continuous variables were expressed as medians with interquartile ranges (IQRs). There was no missing value due to the nature of database and variables used. All variables used for multivariable adjustments were determined based on clinical expertise. The number of major cardiovascular comorbidities was compared between groups using multivariable Poisson regression adjusting for all recorded baseline variables, with the ratio of cardiovascular comorbidity counts and corresponding 95% confidence intervals (CIs) as summary statistics. Noncardiovascular mortality constituted a competing event for the primary outcome. As Kaplan-Meier curves over-estimate cumulative incidences in the presence of competing events,^{14, 15} the cause-specific cumulative incidence of the primary endpoint was estimated and visualized using the Aalen-Johansen estimator.^{14, 15} The 3-, 5- and 10-year cause-specific cumulative incidences were estimated for each group (none of HF/MI/stroke/arrhythmia, HF only, MI only, stroke only, arrhythmia only and at least two of HF/MI/stroke/arrhythmia). The cumulative incidence of the primary endpoint was quantitatively compared between groups using multivariable Fine-Gray competing risk regression, with subhazard ratios (SHRs) and 95% CIs as summary statistics and adjusting for all collected baseline variables. All reported SHRs and corresponding CIs are adjusted estimates.

As the exploratory analysis for the relative impact of each major cardiovascular comorbidity focused on patients with ≥2 of these comorbidities, the corresponding Fine-Gray regressions were additionally adjusted for the number of major cardiovascular comorbidities present to account for potential imbalances in the number of comorbidities between groups. Furthermore, in view of the results from the exploratory analysis, a post hoc analysis was conducted to explore the prognostic effects of the number of major cardiovascular comorbidities on the primary endpoint among patients with ≥2 of these comorbidities. Because very few patients had all four of these comorbidities, this post hoc analysis compared those with two of these comorbidities against those with ≥3 of these comorbidities.

Fine-Gray regression was separately performed in two prespecified and one post hoc subgroup analyses. First, to clarify any impact that heterogeneity in the type of ADT may have on the observed effects, subgroup analysis was performed for each type of ADT (medical castration, bilateral orchidectomy and both; all patients received bilateral orchidectomy after medication castration, mostly because, until recently, the local reimbursement system did not subsidize medical castration and bilateral orchidectomy was often performed as a cost-saving long-term alternative to medical castration). On the other hand, as cancer staging and histology were not available from the database used, we used ever-prescription of ARSI or chemotherapy as a surrogate for metastatic disease, as these were only indicated in metastatic PCa.^{16, 17} The second subgroup analysis thus stratified patients by whether they were ever prescribed ARSI or chemotherapy, as a surrogate of whether they had metastatic disease. The third, post hoc subgroup analysis explored the effects of metabolic dysfunction or hypertension on our findings, with stratification for the presence of hypertension, diabetes mellitus or dyslipidemia. Due to the small sample sizes for each risk category, interactions between subgroups were not tested. Additionally, due to the relatively small number of patients with ≥2 major cardiovascular comorbidities, these subgroup analyses were not performed for the exploratory analysis of the relative impact of each comorbidity in those with overlapping comorbidities.

Two-sided P < .05 were considered statistically significant. All analyses were performed on Stata version 16.1 (StataCorp LLC, College Station, Texas).

3.1 Impact of cardiovascular comorbidities

Over a median follow-up of 3.3 [1.5-6.7] years, 3225 (23.8%) met the primary endpoint and 6662 (49.2%) died of noncardiac causes without meeting the primary endpoint; 9113 (67.6%) died in total. Of the 3225 patients who met the primary endpoint, 1 had cardiovascular mortality and 3224 had cardiovascular hospitalization (among whom 670 [4.9% of all patients; 20.8% of those who had cardiovascular hospitalization] went on to have cardiovascular mortality). Cumulative incidence of the primary endpoint is shown in Figure 1, and each group's estimated 3-, 5- and 10-year cumulative incidences of the primary endpoint are summarized in Table S4. Compared to patients who had none of prior HF/MI/stroke/arrhythmia, patients with only prior HF (SHR 1.67 [95% confidence interval (CI): 1.37-2.02], P < .001), prior arrhythmia (SHR 1.63 [1.35-1.98], P < .001) or prior MI (SHR 1.43 [1.14-1.79], P = .002) had significantly higher incidence of the primary endpoint, but not those who only had prior stroke (SHR 1.06 [0.92-1.23], P = .391). Those who had ≥2 of HF/MI/stroke/arrhythmia had the highest incidence of the primary endpoint (SHR 1.94 [1.62-2.33], P < .001). In post hoc analysis, patients with stroke only did not have significantly different incidence of the primary endpoint regardless of the type of stroke (ischemic stroke: SHR 1.07 [0.92-1.24], P = .408; hemorrhagic stroke: SHR 1.10 [0.77-1.56], P = .598; both ischemic and hemorrhagic stroke: SHR 0.93 [0.53-1.63], P = .808), compared to those without any of major cardiovascular comorbidities.

Among the 1881 patients with only one of the major cardiovascular comorbidities, HF (SHR 1.51 [1.21-1.89], P < .001), arrhythmia (SHR 1.39 [1.11-1.73], P = .004) and MI (SHR 1.31 [1.03-1.66], P = .026) were all associated with significantly higher incidences of the primary endpoint compared to those with only stroke.

3.2 Subgroup analysis

Subgroup analysis by the type of ADT (Table 3) showed mostly consistent trends in patients who received only medical castration (N = 6944) or only bilateral orchidectomy (N = 5116), with those having ≥2 of HF/MI/stroke/arrhythmia having the highest incidence of the primary endpoint, and with only HF, MI and arrhythmia being associated with significantly higher incidence of the primary endpoint when compared to those without any of HF/MI/stroke/arrhythmia. However, no significant association was observed in the 1234 patients who received both medical castration and bilateral orchidectomy.

On the other hand, while MI and arrhythmia remained to be associated with significantly higher incidences of the primary endpoint regardless of whether patients had metastatic disease (ie, prescribed ARSI or chemotherapy; Table S5), such association for HF was significant in those who did not have metastatic disease (ie, never prescribed ARSI or chemotherapy), but only approached significance in those with metastatic disease (P = .059). Having ≥2 of HF/MI/stroke/arrhythmia remained to be associated with the highest incidence of primary endpoint regardless of whether patients had metastatic disease.

In the post hoc subgroup analysis by the presence of hypertension, diabetes mellitus or dyslipidemia, findings in both subgroups were largely consistent with the main analysis (Table S6), although the association for patients with only MI among those with hypertension, diabetes mellitus or dyslipidemia only approached statistical significance (P = .055).

3.3 Relative impact of each comorbidity in patients with overlapping comorbidities

The relative impact of each specified major cardiovascular comorbidity were explored among the 494 patients with ≥2 of HF/MI/stroke/arrhythmia, with cumulative incidences of the primary endpoint stratified by different combinations of cardiovascular comorbidities (Figure 2). HF, MI, stroke and arrhythmia did not have significantly different impact on the incidence of the primary endpoint (Table 4). Post hoc analysis suggested that compared to those with two of HF/MI/stroke/arrhythmia, those who had ≥3 of these comorbidities had significantly higher incidence of the primary endpoint (SHR 1.42 [1.02-1.97], P = .037).

4.1 Strengths and limitations

Using data from a prospectively recorded population-based database, our study is representative of the clinical practice in Hong Kong, a modern Asian metropolitan, and our findings are likely generalizable to many other Asian cohorts. The population-based nature also allowed a sizeable cohort, reinforcing the validity of our findings. Furthermore, we carefully considered and modeled for the important issue of competing risks in our study. Competing risks are critically important in cardio-oncology studies, as many patients die of cancer or noncardiovascular causes before experiencing cardiovascular events. By recognizing this issue and modeling for it using appropriate statistical models, we likely avoided significant biases in estimates that would have occurred if other methods not accounting for competing risks were used.

Nonetheless, our study is not without limitations. First, owing to the nature of the database used, cancer staging and histology were not available. Nonetheless, ADT is unlikely to be used for nonadvanced disease.^{16, 17} Additionally, we attempted to mitigate this by using prescription of ARSI or chemotherapy as a surrogate for metastatic disease^{14, 16, 17, 22} and performing a subgroup analysis accordingly. The results appeared qualitatively similar for most risk groups with all CIs showing substantial overlaps, supporting that the associations were likely to be independent of whether the PCa is metastatic or not.

Second, owing to the observational nature of our study, unmeasured and residual confounders may be present. For instance, due to the nature of the database used, some important cardiovascular risk factors, such as smoking, metabolic syndrome and obesity, could not be accounted for. We adjusted for the widest range of relevant risk factors available to us, which should sufficiently account for a significant proportion of confounders. Nonetheless, further studies with greater data granularity are required to verify our findings. Third, the database used (CDARS) did not allow individual adjudication of data. However, CDARS has been demonstrated to have good data completeness and accuracy,⁷ and all data were input by treating clinicians independent of the authors.