The data were obtained from the National Health Insurance Service (NHIS) Customized Database, a population-based national healthcare database covering the entire Korean population. The database included eligibility information, health screening and questionnaire data, healthcare utilization claims, and mortality records [14]. Data from 2011 to 2022 were used to assess smoking trajectories during the 6 years preceding the index date and to identify incident cardiovascular and cerebrovascular disease cases from 2017 to 2022. This 6-year preindex period was selected to capture smoking patterns over time and provide sufficient repeated measurements for trajectory modeling. Because NHIS health screening is offered annually or biennially depending on employment type, this period allowed at least three screening opportunities even for biennially screened participants. For cases, the index date was defined as the date of first hospitalization with a principal diagnosis of cardiovascular or cerebrovascular disease.

The study population included NHIS beneficiaries between the ages of 30 and 69 who completed national general health screening from January 2015 through December 2016. In Korea, screening intervals vary by employment type, with annual screening for nonoffice workers and biennial screening for office workers. Participants screened in 2015 and 2016 were pooled to reduce selection bias related to screening year. For individuals screened in both years, the most recent examination was retained, resulting in a baseline population of 18,241,353 individuals. The baseline exclusions included deaths, individuals outside the target age range, medical aid beneficiaries, and individuals with missing covariate data; a total of 471,716 individuals were excluded. Of the baseline population of 17,769,637 individuals, 618,109 potential cases who were first hospitalized with a principal diagnosis of cardiovascular and cerebrovascular diseases between January 1, 2017, and December 31, 2022, were identified. Only individuals with screening records during the 6 years before diagnosis and at least three nonmissing smoking measurements were included. This threshold was applied to provide sufficient longitudinal information for estimating smoking trajectories and classifying participants into trajectory groups [7, 15].

To ensure adequate preindex exposure assessment and exclude prevalent disease, we excluded individuals without a health screening record within 2 years before the index date, as well as those with a history of cancer (ICD-10 codes C00–C99) or cardiovascular and cerebrovascular diseases (ICD-10 codes I20–I25 and I60–I69) during the 6 years preceding the index date. History of disease was defined as at least one hospitalization or three outpatient claims with a principal diagnosis of the corresponding disease, to reduce potential over-ascertainment from miscoding or administrative error. After applying these preindex eligibility criteria, 244,679 cases remained eligible for matching.

The pool of potential controls comprised participants who had no history of cardiovascular and cerebrovascular diseases throughout follow-up. After applying the preindex eligibility criteria, the retained cases were matched to potential controls on sex, age, and insurance premium quantile, an indicator of income-related socioeconomic status. Up to six controls were initially selected for each retained case. Matched controls were then assigned the index date of the corresponding case, after which the same preindex eligibility criteria were applied to matched controls. After excluding incomplete matched sets, the final matched study population comprised 241,163 cases and 865,065 controls (Supplementary Material 1, Supplementary Figure S1).

Smoking behavior was assessed using repeated health screening questionnaires and classified as never smoker, former smoker, smoking 1–9 cigarettes per day, smoking 10–19 cigarettes per day, and smoking ≥20 cigarettes per day. This categorization was used to reflect gradations in smoking intensity and maintain comparability with previous studies [9]. Because questions on electronic cigarettes were introduced only in 2018 and were not consistently available over time, only combustible cigarette smoking was included in the trajectory analysis.

The main outcome was cardiovascular and cerebrovascular disease, identified using ICD-10 codes I20–I25 and I60–I69, consistent with definitions used in previous epidemiologic studies [4, 9, 12]. The subtype-specific outcomes included ischemic heart disease (I20–I25), acute myocardial infarction (I21), cerebrovascular disease (I60–I69), intracranial hemorrhage (I60–I62), and cerebral infarction (I63).

Several covariates were included in the analysis: region, disability status, family history of heart disease or stroke, alcohol consumption, physical activity, body mass index (BMI), hypertension, diabetes, dyslipidemia, and the Charlson Comorbidity Index (CCI). Questionnaire- and examination-based covariates were obtained from the most recent health screening before the index date. BMI was classified following Asia-Pacific criteria [16]: underweight (<18.5 kg/m²), normal weight (18.5–22.9 kg/m²), overweight (23.0–24.9 kg/m²), and obesity (≥25.0 kg/m²). Hypertension, diabetes, and dyslipidemia were identified using screening results suggestive of disease (Supplementary Material 1, Supplementary Table S1), documented disease history, or use of medication for the relevant condition.

Exact matching was applied to improve the comparability between cases and controls and minimize confounding by factors associated with both smoking behavior and disease occurrence, in accordance with the established methodological principles for case-control studies [17–19]. The standardized mean differences were measured to assess balance on matching variables, and all were below 0.10. The cases and controls were matched on sex, age, and insurance premium quantile, with up to six controls randomly selected for each case within the same matching stratum. Controls were drawn from participants with no record of cardiovascular or cerebrovascular disease during follow-up, consistent with a cumulative sampling design [19]. Because smoking trajectories were defined using repeated preindex assessments and disease status was determined after follow-up, the study design helped preserve temporal ordering and reduce the likelihood of reverse causation.

Smoking trajectories during the 6 years preceding the index date were identified using group-based trajectory modeling, which is a semiparametric approach for identifying distinct longitudinal patterns within a population [7, 8]. In the trajectory analysis, time was treated as the longitudinal scale and smoking behavior as the repeated outcome variable. Because smoking behavior was measured as an ordered quantitative variable with numeric intervals, trajectories were estimated using the censored normal (CNORM) model for continuous outcomes, as described by Jones and Nagin [20]. To allow for changes in smoking behaviors over time, polynomials up to the cubic order were assessed, and the appropriate order for each trajectory group was determined during model fitting. The candidate solutions with two to five groups using the Akaike information criterion (AIC) and Bayesian information criterion (BIC) were compared. Solutions with six or more groups could not be estimated because of nonconvergence and were excluded from model selection.

The associations between smoking trajectory groups and cardiovascular and cerebrovascular diseases were estimated using conditional logistic regression to obtain odds ratios (ORs) and 95% confidence intervals (CIs) within matched strata [21, 22]. The matched sets were treated as strata in the regression models, and the matching variables (sex, age, and insurance premium quantile) were not additionally included as covariates. Model 1 included only smoking trajectory groups. Model 2 additionally adjusted for region, family history, and disability status. Model 3 was further adjusted for alcohol consumption, physical activity, and BMI. Model 4 was additionally adjusted for hypertension, diabetes, and dyslipidemia. In Model 5, the individual metabolic conditions in Model 4 were replaced with the CCI to account for the overall comorbidity burden.

The robustness of the primary findings was assessed via sensitivity analyses. First, conditional logistic regression analyses were repeated separately among those participants whose most recent health screening was conducted within 1 year before the index date and among those screened between 1 and 2 years before the index date. Second, to assess the validity of the trajectory-based approach, we compared the model fits between the trajectory-based smoking measures and single time-point smoking measures that were assessed 1 and 6 years before the index date using the AIC and BIC. All analyses were conducted using SAS version 9.4 (SAS Institute Inc., Cary, NC, USA), and trajectory modeling was performed using the PROC TRAJ procedure [23]. Statistical significance was set at a two-sided P value of <0.05.

Compared with the controls, the cases were more likely to reside in small-to-medium-sized cities or rural areas and to have a family history of heart disease or stroke and a disability status (Supplementary Material 1, Supplementary Table S2). The cases also showed less favorable lifestyle profiles, including heavier smoking, higher alcohol consumption, physical inactivity, and obesity, as well as a higher prevalence of metabolic conditions and a greater overall comorbidity burden, as reflected by the higher CCI scores.

The smoking trajectory groups were identified over the 6 years preceding the index date (Figure 1; Supplementary Material 1, Supplementary Table S3). Model fit was assessed for models with two to five groups using the AIC and BIC. The adequacy of group assignment was assessed using the average posterior probability of group membership, with values ≥0.70 considered acceptable [24, 25]. A five-group model was selected as the final model because it showed the best fit according to the AIC and BIC values and met the minimum group size criterion of 5% for all groups [25]. The final model identified five trajectory groups: never (50.6%), former (23.6%), high-decreasing (5.2%), moderate-stable (12.4%), and high-stable (8.2%) groups.

The never group was the largest and comprised higher proportions of women, individuals in their 60s, lower-income individuals, and rural residents (Table 1). The former group comprised higher proportions of men, older adults, and higher-income individuals and showed the highest proportion of family history of heart disease or stroke. The high-decreasing and moderate-stable groups were predominantly male and included relatively high proportions of individuals in their 40s and 50s. The high-decreasing group was additionally characterized by a higher prevalence of obesity, whereas the moderate-stable group showed more frequent alcohol consumption. The high-stable group was characterized by higher proportions of men, individuals in their 50s, and lower-income individuals and showed the highest prevalence of rural residence, disability, frequent alcohol consumption, physical inactivity, and metabolic conditions.

Table 2 shows the conditional logistic regression results for the associations between smoking trajectories and cardiovascular and cerebrovascular diseases. The direction and magnitude of the associations were largely similar across models. Model 5, adjusted for sociodemographic factors, health-related factors, and the CCI, was used as the primary fully adjusted model because it accounted for the overall comorbidity burden using the CCI.

In Model 5, compared with the never group, the odds of cardiovascular and cerebrovascular diseases were highest in the high-stable group (OR = 1.97; 95% CI 1.93–2.01), followed by the moderate-stable group (OR = 1.59; 95% CI 1.56–1.62), high-decreasing group (OR = 1.36; 95% CI 1.33–1.40), and former group (OR = 1.11; 95% CI 1.09–1.13).

Table 3 shows the subtype-specific associations from Model 5. Compared with the never group, the high-stable group showed increased odds across all disease subtypes. The strongest association was observed for acute myocardial infarction (OR = 3.51; 95% CI 3.34–3.69), followed by cerebral infarction (OR = 2.46; 95% CI 2.35–2.56), ischemic heart disease (OR = 2.04; 95% CI 1.99–2.09), overall cerebrovascular disease (OR = 1.90; 95% CI 1.84–1.96), and intracranial hemorrhage (OR = 1.49; 95% CI 1.37–1.61). All these associations were statistically significant (P < 0.001), with larger estimates for ischemic outcomes than for intracranial hemorrhage.

Supplementary Tables S4, S5 (Supplementary Material 1) show the results of the sensitivity analyses. In the analyses stratified by the timing of the most recent health screening before the index date, the pattern of associations across the smoking trajectory groups was similar to that observed in the main analysis. Moreover, the trajectory-based smoking measures showed a better model fit than the single time-point smoking measures at both time points, as indicated by the lower AIC and BIC values.

Several limitations should be noted. First, the study population was restricted to national health screening participants, which may introduce healthy screenee bias and limit the generalizability of the findings to nonparticipants. Second, smoking status was based on self-reported questionnaire data and may be subject to measurement error or reporting bias. Age at smoking initiation is an important aspect of smoking history, but this information was not collected in the health screening questionnaire and could not be assessed. Residual confounding by unmeasured factors, including education, employment status, deprivation, diet, sleep, and more detailed measures of alcohol consumption and physical activity, also cannot be excluded. Third, reverse causality remains possible because some participants may have reduced or stopped smoking in response to preclinical or existing health conditions before the index date. Although the consistency of associations across subgroups defined by the timing of the most recent screening supports the robustness of the findings, reverse causality cannot be fully excluded. Fourth, although classification diagnostics supported the adequacy of the selected trajectory model, group assignment in trajectory modeling is probabilistic and may be influenced by model-selection decisions. Finally, the 6-year observation window may not have fully captured the cumulative vascular effects of longer-term smoking exposure, as cardiovascular risks associated with smoking can persist over decades [33, 34]. Future studies with longer prediagnostic exposure histories are needed to determine whether prolonged smoking patterns confer additional vascular risk.

Smoking trajectories over the 6 years preceding the index date were associated with cardiovascular and cerebrovascular disease, with progressively higher odds in groups characterized by greater smoking intensity and persistence. Associations were stronger for ischemic than for hemorrhagic outcomes. Repeated smoking information from population-based screening programs may characterize prediagnostic exposure patterns and identify sustained high-risk smoking profiles for earlier cessation intervention.