We performed a retrospective, population-based cohort study using the Korean National Health Insurance Service (K-NHIS) database. Korea has a mandatory social insurance system with insurance premiums that are determined by income level and not by health status. The K-NHIS is a single insurer that covers approximately 97% of the population, while the remaining 3% of beneficiaries are covered by the Medical Aid Program. Data on the use of medical facilities and records of prescriptions with International Statistical Classification of Diseases and Related Health Problems, 10th Revision (ICD-10), diagnosis codes are gathered by the NHIS. The K-NHIS claims database also includes information on demographics, medical treatment, procedures, prescription drugs, diagnostic codes, and hospital use. Vital status and cause of death were obtained from death certificates collected by Statistics Korea at the Ministry of Strategy and Finance of South Korea [17]. Use of the K-NHIS database was approved by the NHIS review committee.

The main exposure was the incidence of cancer in participants 15–39 years old. To define incident cancer, we used a special registration code V193 in addition to the relevant ICD-10 diagnosis code. The K-NHIS has established a special copayment-reduction program to enhance health coverage and reduce the financial burden of patients with cancer. Once cancer patients are registered in the system, they pay only 5% of the total medical bill incurred for cancer-related medical care. Since enrollment in this copayment-reduction program is indicated by a special copayment-reduction code for cancer (V193) and requires a medical certificate from a physician, the cancer diagnoses included in this study are considered to be sufficiently reliable, and this method has been used in previous studies [18].

For this study, we considered all Korean men and women aged 15–39 years enrolled between 2006 and 2019 in the K-NHIS database between 2005 and 2020. Data access was restricted by a data-share policy; we selected all patients with cancer defined by the presence of ICD-10 code C or a special copayment reduction code for cancer (V193) between 2006 and 2019 (n = 681,752) and fourfold the number of age- and sex-matched samples of men and women who did not develop cancer during the study period (n = 2,670,558). To select newly diagnosed cancer as the exposure and incident cases of CVD as the outcome, we excluded 109,269 participants with CVD (n = 66,317) or any cancer (n = 45,732) before January 1, 2006.

Among the eligible participants (n = 3,243,041), we mimicked sequential emulation of the target trial (detailed methods are presented in the Statistical Analysis section) [19, 20]. The number of cloned participants aged 15–39 years enrolled in the K-NHIS database between 2006 and 2019 was 62,985,785. During the process of enrolling the new cohort, we excluded participants who had any cancer (n = 1,570,072) or a history of CVD at each baseline point (n = 1,859,670). Participants who met the eligibility criteria in the previous cohort were excluded from the next cohort if they were > 40 years old, had cancer or a history of CVD, or died prior to the start date. These processes were repeated every 6 months until June 1, 2019 (n = 59,561,447/unique n = 3,108,601). Since our focus was on AYA gastric cancer, patients who underwent gastric cancer surgery, including endoscopic operation of an upper gastrointestinal tumor and endoscopic submucosal dissection, as well as total and subtotal gastrectomy, within the 2 months prior to diagnosis and within 1 year thereafter with ICD-10 diagnosis code (C16) were selected for the gastric cancer group (n = 6562). Then, we also selected a threefold larger control group using propensity score matching (n = 19,678) (Fig. 1).

Flow chart of study population selection

The Institutional Review Board of the Samsung Medical Center approved the study and waived the requirement for informed consent because K-NHIS data were de-identified (SMC 2022-03-028).

The primary endpoint is a composite outcome of any cardiac outcomes, including myocardial infarction (ICD-10: I21–I22), stroke (ICD-10: I60–I64), heart failure (ICD-10: I50), cerebrovascular disease (ICD-10: I63–I69), atrial fibrillation (ICD-10: I48), arrhythmia (ICD-10: I47–I49), cardiomyopathy (ICD-10: I42–I43, I23.5), valvular heart disease (ICD-10: I01–I08, I34–I37), venous thromboembolism (VTE): deep venous thromboembolism (DVT) (ICD-10: I80.1–I80.3), and pulmonary embolism (PE) (ICD-10: I26). The cardiovascular outcomes were identified by diagnostic records, according to the ICD-10 codes from either outpatient visits or hospitalization. The definitions of outcomes are summarized in Supplemental Table 1. In regard to myocardial infarction diagnosis, 93% accuracy was achieved in the validation study [21].

For the covariates, we included age, sex, comorbidities, income, and residential area at baseline. The presence of diabetes (ICD-10, E100-E149), hypertension (ICD-10, I10-I15), or dyslipidemia (ICD-10, E780-E785) was defined as having had at least one clinic visit or hospitalization with the corresponding ICD-10 code within the previous year. Data on income were obtained from the insurance eligibility database. Income level was categorized by percentile (≤ 30th, > 30th– ≤ 70th, and > 70th percentiles). Residential area was classified as metropolitan or rural. Metropolitan areas were defined as Seoul, six metropolitan cities, and 15 cities with > 5,00,000 residents that have been officially designated as municipal cities (http://www.mois.go.kr). We conducted a sensitivity analysis to examine whether body mass index (BMI), alcohol drinking, smoking, and physical activity have an additional impact on CVD outcome. For this analysis, we restricted the participants to those who underwent a health screening exam 4 years prior to baseline.

The participants were then classified into two groups based on whether they developed gastric cancer or not. In this process, we generated a propensity score using logistic regression with incident cancer as the outcome variable and age, sex, income, residential area, and the presence of comorbidities (diabetes mellitus, hypertension, and hyperlipidemia) at cohort entry as covariates. Then, a 1:3 matching ratio was applied using the propensity score through greedy matching methods (caliper < 0.1). To compare the distribution of variables used for matching, a standardized mean difference between the gastric cancer and control groups was estimated. The variables used for matching were updated based on the first date for each subsequent cohort.

The primary endpoint was development of CVD. Each endpoint was analyzed separately, and we included only participants who had not experienced the endpoint of interest prior to the endpoint analysis. We followed the participants from the baseline of each subsequent cohort until the development of CVD, death, or December 2020, whichever occurred first. After completing all processes, we pooled the data from all trials into a single model and included the day at baseline of each cohort.

The cumulative incidence of each outcome was estimated with the Kaplan–Meier method, and log-rank tests were applied to evaluate differences between the groups. We calculated hazard ratios (HR) with 95% confidence intervals (CIs) for clinical outcome incidence using a Cox regression model. The time scale was the calendar year. We examined the proportional hazards assumption using plots of the log (-log) survival function and Schoenfeld residuals.

In sensitivity analysis, to account for competing risks due to mortality, we fitted a proportional subdistribution hazards (subHR) regression model [22] with death as the competing event. Since the year 2020 was complicated by the coronavirus disease 2019 pandemic, we performed a separate sensitivity analysis to exclude the year 2020. To understand the differential impacts of treatment modalities on the outcomes of interest, additional subgroup analyses were performed in the surgery-only group and the chemotherapy group. These groups were compared to their matched controls from the general population without gastric cancer at baseline. In addition, individual control group tests were conducted for each subgroup. We also performed multivariable Cox regression to find risk factors influencing the incidence of CVD among AYA gastric cancer patients after adjusting for covariates of BMI, smoking status, alcohol consumption and regular physical activity. As the risk of VTE significantly increased in AYA gastric cancer patients, we also presented risk factors affecting the incidence of VTE.

All P values were two-sided, and P < 0.05 was considered statistically significant. Analyses were performed with SAS® Visual Analytics (SAS Institute Inc., Cary, NC, USA) and R 4.1.2 (R Foundation for Statistical Computing, Vienna, Austria).