Study design and participants

Details of this cohort study have been published previously[8–10]. Briefly, the Chin-Shan Community Cardiovascular Cohort (CCCC) Study began in 1990 by recruiting 1703 men and 1899 women aged 35 years old and above, in the Chin-Shan township. A total 1816 participants (50.4%) with available serum 25(OH)D measurements was included into the study. Information about anthropometry, lifestyle, and health conditions was assessed by interview questionnaires in 2-year cycles up to 2011 and the validity and reproducibility of the collected data and measurements have been reported in detail elsewhere[9]. Incident CVD included coronary heart disease and stroke cases. Incident coronary heart disease cases were defined as fatal coronary heart disease and hospitalization due to nonfatal myocardial infarction or percutaneous coronary intervention and coronary bypass surgery. Non-fatal myocardial infarction and hospitalizations for coronary intervention and bypass graft surgery was ascertained by the combined information from patient interviews and medical record review. Incident stroke cases were ascertained according to the following criteria: a sudden neurological deficit of vascular origin that lasted longer than 24 hours, with supporting evidence from the image study and medical records. All-cause deaths were identified from the official certificate documents, further verified by house-to-house visits. The National Taiwan University Hospital Committee Review Board approved the study protocol.

Laboratory measurements

We performed the biochemical measurements and fatty acid profiles once in the baseline. The procedures involved in blood sample collection were previously reported.[11] Briefly, all venous blood samples were drawn after a 12-hour overnight fast, immediately refrigerated, and transported within 6 hours to the National Taiwan University Hospital. Serum samples were stored at -70°C before conducting batch assays to determine the levels of total cholesterol, triglycerides, and high density lipoprotein cholesterol (HDL-C). To ascertain these levels, standard enzymatic tests for serum cholesterol and triglycerides were employed (Merck 14354 and 14366, Germany, respectively). The supernatants were measured for HDL-C levels following the precipitation of specimens using a reagent of magnesium chloride phosphotungstate (Merck 14993). Low density lipoprotein concentrations were calculated as total cholesterol minus cholesterol in the supernatant obtained through precipitation (Merck 14992).[12]

The serum 25(OH)D levels were evaluated in the hospital laboratory using the 25(OH)-Vitamin D direct Elisa Kit (Immunodiagnostik; Bensheim, Germany), based on a competitive ELISA technique with a selected monoclonal antibody that recognizes 25(OH)D. The results were expressed, after point-to-point calculation, as nmol/L (with 1 ng/mL being equivalent to 2.5 nmol/L)[13], and the inter- and intra-assay CV was 7.0%,

Statistical analysis

Participants were categorized on the basis of 25(OH)D quartiles; the ANOVA and the chi-square tests were used to compare means and proportions among the quartiles of 25(OH)D concentration. In addition, the age, gender-adjusted partial Spearman correlation coefficients were estimated between 25(OH)D and various atherosclerotic risk factors.

We plotted the Kaplan-Meir survival curves for all-cause death and cardiovascular events according to categories and stratified according to baseline hypertension and diabetes status. The incidence rates of all-cause death and CVD events were calculated by dividing the number of cases by the number of person-years of the follow-up for each quartile. Quartiles of Vitamin D were entered as dummy variables in a Cox proportional hazards regression model to calculate the hazard ratio (HR) and 95% confidence interval (CI), taking the lowest quartile as the reference. Model 1 was adjusted for age groups (35–44, 45–54, 55–64, 65–74, > = 75 years old) and gender only. Model 2 included additional confounding factors: adjusted for the following baseline covariates: body mass index (<18, 18–20.9, 21–22.9, 23–24.9, > = 25 kg/m²), smoking (yes/no or abstinence), current alcohol consumption habits (yes/none, defined by > = 1 drink per day), education level (less than 9 years, at least 9 years), occupation (not employed, manual labor, or office job), and regular exercise (yes/no, defined by > = 30 min per day moderate intensity physical activity, such as walking or cycling). Model 3 included additional clinical variables: baseline hypertension (yes/no, defined by baseline systolic blood pressure > = 140 mmHg, diastolic blood pressure > = 90 mmHg, or on hypertensive medication), type 2 diabetes (yes/no, defined as baseline fasting glucose > = 126 mg/dL or on hypoglycemic medications), continuous LDL cholesterol, and HDL cholesterol values. Tests for trend were derived from the regression model with a single term representative of the medians of each quartile. Due to potential high-risk impacts, we performed subgroup analyses based on baseline hypertension and type 2 diabetes. Proportional hazard assumption was not rejected in these Cox models by plotting the log(-log(survival time)) versus log of survival time and including time dependent covariates.

To further investigate the role of 25(OH)D concentration to predict the risk, we compared the model including traditional risk factors and the model adding 25(OH)D and tested the prediction performance using calibration and discrimination ability. First, we assessed the goodness of fit for all models based on the Hosmer-Lemeshow test [14], which was a calibration measure to calculate how close the predicted risks were to the actual observed risks [15], and the results showed the calibration was good. Second, we compared the discrimination ability using the area under receiver operative characteristic curve (AUC). An AUC curve is a graph of sensitivity versus 1-specificity (or false-positive rate) for various cutoff definitions of a positive diagnostic test result [16]. Statistical differences in the AUCs were compared using the method of DeLong et al [17]. The AUC was a global summary measure for discrimination between individuals developing outcomes and those who did not[18]. Third, we compared the models by using the net reclassification improvement (NRI) and integrated discrimination improvement (IDI) statistics [19]. The NRI statistic was based on the reclassification tables and was calculated from a sum of differences between the ‘upward’ movement in categories for event subjects and the ‘downward’ movement in those for nonevent subjects. We presented the NRI according to the a priori risk categories of risk. The IDI can be interpreted as a difference between improvement in average sensitivity and any potential increase in average ‘one minus specificity’, and the statistic was a difference in Yates discrimination slopes with and without 25(OH)D information.

In addition, we used the natural cubic splines semi-parametric regression models to examine the relationship between 25(OH)D concentration and the risk of all-cause death and cardiovascular events. Natural cubic splines are smooth polynomial functions that can be used to fit data and accommodate potential changes in the direction of the association between the exposure and the outcome of interests, and are constructed of piecewise third-order polynomials which pass through a set of control points and its linear in this tail beyond the boundary knots[20,21]. A SAS macro named ‘lgtphcurv9’ written by Harvard scholars[22] was used which implements natural cubic spline methodology to fit potential non-linear dose-response curves in proportional hazards models. Likelihood ratio tests were performed to test non-linear and linear relationships.

All statistical tests were two-sided and P values < 0.05 were considered statistically significant. Analyses were performed with SAS version 9.2 (SAS Institute, Cary, NC) and Stata version 11 (Stata Corporation, College Station, Texas).

Previous literature and current study comparison

Evidence for an inverse association between vitamin D and all-cause death and cardiovascular events was still controversial. Most evidence showed an inverse effect existed for vitamin D as a role for cardiovascular events: Dobnig and colleagues[23] showed that lower 25(OH)D concentration quartiles were related to a higher all-cause death (hazard ratio, 2.08, 95% CI, 1.60–2.70) and cardiovascular death (hazard ratio, 2.22, 95% CI, 1.57–3.13) among high-risk patients. In addition, low Vitamin D concentration (less than 50 nmol/L) was highly prevalent (39%) and was associated with an additional 44% (95% CI, 1.01–2.01, p = 0.04) risk for cardiovascular risk[24] among Danish women. Schierbeck[24] also found that vitamin D deficiency (defined by serum 25(OH)D <50 nmol/L) was associated a 1.49 fold (95% CI, 1.16–1.92) for a composite endpoint, including death, heart failure, myocardial infarction and stroke among Danish women. Moreover, Liu and colleagues showed that 25(OH)D less than 50 nmol/L was associated with a 1.40 fold risk (95% CI, 1.17–1.68) for all-cause death (P<0.001)[5] among adults in USA. Among 18225 men in the Health Professionals Follow-up Study with 10 years of follow up, Giovannucci and colleagues[4] used the 2:1 matched nested case-control study design and showed that lower 25(OH) D concentration (< = 37.5 nmol/L was associated with a 2.42 fold risk (95% CI, 1.53–3.84) for acute myocardial infarction, compared with those of high 25(OH)D (>70 nmol/L), and a graded effect was found[4].

The relationship between vitamin D insufficiency with cardiovascular disease may be more evident among the high-risk group such as hypertension, diabetes mellitus and chronic renal disease[1]. A significant trend for higher risks for all-cause deaths and cardiovascular deaths were found in the lower quartiles among hypertensive participants[25]. In addition, a reverse linear association between 25(OH)D and all cause deaths as well as cardiovascular deaths were found[25], after adjustment for covariates such as blood pressure and medication history.

However, some negative results were available: Jassal and colleagues showed that a higher concentration than mean 25(OH)D concentration as 105 nmol/L did not relate cardiovascular mortality in a community[6]. In addition, vitamin D was a protective factor for cardiovascular mortality only in women, but not men among 3404 adult Swiss population[26]. Moreover, a cohort study based on 312 older Spanish (> = 85 years old) did not showed 25(OH)D was related to cardiovascular death[7]. Our current study showed that overall the association between 25(OH)D and outcomes were not significant; however, the protective effect was found among participants who were diabetic at baseline, and diabetes status had an effect modification to influence the association between 25(OH)D and all-cause death. Our findings implied that the protective effect of 25(OH)D was limited in some specific groups.

Proposed mechanism of vitamin D for CVD and all-cause death

Our findings supported that in high risk population such as type 2 diabetes, vitamin D indeed provided a protective role for all-cause death. A proposed mechanism for diabetes and cardiovascular disease was through obesity and endothelial dysfunction[27]. Vitamin D regulated macrophage cholesterol metabolism among diabetic patients to facilitate atherosclerosis process[28]. In addition, vitamin D deficiency was accompanied with activation of the renin-angiotensin-aldosterone system[29,30], and the associated increase of parathyroid hormone has been related to insulin resistance[31]. Using the knockout mouse model, Zhou and colleagues[32] clearly demonstrated that Vitamin D was an important regulator for renin-angiotensin system. This modulation of the renin-angiotensin system, in addition to decreasing blood pressure, improved the endothelial function and decreased inflammatory burden, thus limiting progression of atherosclerosis.

Clinical implication

Literature showed that a proposed 30 ng/mL (75 nmol/L) as the cutoff for vitamin D concentration was sufficient to provide regulatory effects for cardiovascular disease prevention [33]. Regarding with a cross-sectional study based on nation-wide survey in US, Judd and colleagues (2008) found that an optimal 25(OH)D concentration would attenuate the age-associated increase in systolic blood pressure: more than 80 nmol/L was associated with a 20% systolic blood pressure decrease, especially for white population[34]. In addition, seasonable variation of 25(OH)D concentration was observed among 1362 male Health Professionals Follow-up Study[35], and the serum 25(OH)D concentration was estimated as 82 nmol/L in summer and 75 nmol/L in winter among 262 healthy older women living in Taipei[36]. Our study did not provide seasonal variation; however, the sampling time was limited in summer time. The normal range of 25(OH)D was suggested as 50–100 nmol/L (20–40 ng/dL)[37], although some researchers argued that the lowest limit as 5 to 80 nmol/L[38]. Our study showed that a cutoff point at 50 nmol/L provided a discriminative measure for the risk for all-cause death.

Strengths and weakness

To the best of our knowledge, this is the first prospective cohort study based on ethnic Chinese to investigate the association and prediction of vitamin D and cardiovascular risk and all-cause death. Because of the prospective cohort design in this study, the baseline measurements of participants were not affected by information bias. In addition, the community-based population information could help to reduce the selection bias. Moreover, to control potential confounding factors, we incorporated important socioeconomic and lifestyle factors into the model. We also take consideration into the nonlinear relationship between vitamin D and outcomes. Furthermore, in addition to traditional multivariate adjusted modeling, we applied prediction measures, such as AUCs and integrated discrimination improvement, to demonstrate predictive ability of vitamin D. These measures provided comparative powers for novel biomarker prediction ability[39,40].

However, this study has several potential limitations. First, vitamin D was only once measured and no seasonable variation was explained, the association between vitamin D and outcome risk may be attenuated. Second, we did not examine the supplementary intake and seasonable variations so that we cannot explain the dietary habit role in the risk. Third, due to observational study, residual confounding factors cannot be extensively examined. Finally, potential information bias such as confounding by indication and missing was not completely excluded due to the cohort study design.