ORIGINAL ARTICLE
Prediction models for the risk of new-onset
hypertension in ethnic Chinese in Taiwan
K-L Chien
1,2, H-C Hsu
2, T-C Su
2, W-T Chang
3, F-C Sung
4, M-F Chen
2and Y-T Lee
2,4 1Institute of Preventive Medicine, College of Public Health, National Taiwan University, Taipei, Taiwan; 2Department of Internal Medicine, National Taiwan University Hospital, Taipei, Taiwan;3Department ofEmergency Medicine, National Taiwan University Hospital, Taipei, Taiwan and4China Medical University
Hospital, Taichung, Taiwan
Prediction model for hypertension risk in Chinese is still lacking. We aimed to propose prediction models for new-onset hypertension for ethnic Chinese based on a prospective cohort design on community, which re-cruited 2506 individuals (50.8% women) who were not hypertensive at the baseline (1990–91). Total 1029 cases of new-onset hypertension developed during a median of 6.15 (interquartile range, 4.04–9.02) years of follow-up. In the clinical model, gender (2 points), age (8 points), body mass index (10 points), systolic blood pressure (19 points) and diastolic blood pressure (7 points) were assigned. The biochemical measures, including white blood count (3 points), fasting glucose (1 point), uric acid (3 points), additional to above clinical variables, were constructed. The areas under the
receiver operative characteristic curves (AUCs) were 0.732 (95% confidence interval (CI), 0.712–0.752) for the point-based clinical model and 0.735 (95% CI, 0.715– 0.755) for the point-based biochemical model. The coefficient-based models had a good performance (AUC, 0.737–0.741). The point-based clinical model had a similar net reclassification improvement as the coefficient-based clinical model (P¼ 0.30), and had a higher improvement than the point-based biochemical model (P¼ 0.015). We concluded that the point-based clinical model could be considered as the first step to identify high-risk populations for hypertension among Chinese.
Journal of Human Hypertension advance online publication, 8 July 2010; doi:10.1038/jhh.2010.63
Keywords: cohort study; prediction model; community screening
Introduction
Identifying individuals who are at high risk of hypertension will improve the efficiency of primary prevention strategies. Recent clinical trials have demonstrated that body weight control and lifestyle intervention in individuals with pre-hypertensive status can substantially delay hypertension devel-opment,1providing a rationale for the identification
of high-risk individuals so as to implement early lifestyle intervention strategies to prevent hyperten-sion. Routinely available and easily collected clinical information and lifestyle-related factors have been found to be effective in identifying hypertension risk in prevalent and incident cases.2–8 However, the evidence on prediction
models providing absolute risk for hypertension risk is relatively scanty and these prediction models have also been developed, primarily in Cauca-sians.2,4–5 Moreover, previous studies based on
hypertension prediction models were limited
because of short follow-up periods,3–4,9an inability
to incorporate laboratory data,4 multiple
biomar-kers,10 limited validation11 and a lack of simple
algorithm usage (Supplementary Table S1). Further-more, the self-reporting of hypertension incidence may invalidate the accuracy of incidence rates.4
Therefore, we constructed the prediction models for hypertension risk using a community-based cohort of middle-aged and elderly ethnic Chinese in Taiwan as the following strategies. First, we incor-porated gender, age, body mass index (BMI), systolic (SBP) and diastolic blood pressures (DBP) as the clinical model and included white blood cell count, fasting glucose and uric acid12–13as the biochemical
models. Second, we proposed two different scoring systems: regression coefficient-based scores14 and
point-based scores.15 Finally, we tested the
perfor-mance measures of these prediction models and compared the available models.
Materials and methods
Study design and study participants
Details of this cohort study have been previously published.16–17 Briefly, the Chin-Shan Community
Cardiovascular Cohort study began in 1990 by
Received 29 December 2009; revised 8 April 2010; accepted 24 April 2010
Correspondence: Dr Y-T Lee, Department of Internal Medicine, National Taiwan University Hospital, Taipei 100, Taiwan. E-mail: ytlee@ntu.edu.tw
recruiting 1703 men and 1899 women of Chinese ethnicity aged 35 years old and above from Chin-Shan township. Information about anthropometry, lifestyle and medical conditions was assessed by the interview questionnaires in 2-year cycles for the initial 6 years and the validity and reproducibility of the collected data and measurements have been reported in detail elsewhere.17The response rate of
the cohort participants was 85.7% at the end of the study.
BMI was calculated as weight (in kilograms)/ height (in metres).2 Family history was defined by
first-degree hypertension. Blood pressure was mea-sured twice in the right arm by a mercury sphygmo-manometer with the subject seated comfortably and the arms supported and positioned at the level of the heart. The average of the blood pressure measure-ments was used as previously described.18–19Family
history of hypertension was coded as the prevalent hypertension among first relatives.2 Smoking habit
was defined by current smoking status. Drinking history was defined as a binary variable using the frequency of drinking habits. Regular physical activity was coded as daily exercise habits.
Measurement of biochemical markers
The procedure for blood collection has been re-ported elsewhere.20–21 Briefly, all venous blood
samples drawn after a 12-h overnight fast were immediately refrigerated and transported within 6 h to the National Taiwan University Hospital. Serum samples were then stored at –70 1C before batch assay for levels of total cholesterol, triglycerides and high-density lipoprotein cholesterol. Standard enzymatic tests for serum cholesterol and triglycerides were used (Merck 14354 and 14366, Merck KGaA, Darmstadt, Germany). Glucose levels were measured in the supernatant by enzymatic assay (Merck 3389) in an Eppendorf 5060 autoanalyzer (Eppendorf Corp., Hamburg, Germany). The peripheral blood cell analysis was measured using a blood cell counter (Sysmex Cell Counter NE-8000, TOA Med-ical Electronics Co. Ltd, Kobe, Japan). Plasma uric acid concentrations were assayed with commercial kits (Merck Chem. Co, Darmstadt, Germany) placed in an Eppendorf 5060 autoanalyzer (Eppendorf Corp.).22
Follow-up strategy
We collected events and blood samples from the participants at baseline (1990–1991), the first fol-low-up period (1992–1993) and the fourth folfol-low-up period (1997–1998).16 We measured blood pressure
and collected data on anti-hypertensive medication in the serial follow-up visit periods. Participants with baseline hypertension (defined by SBP or DBP X140/90 mm Hg or a history of anti-hypertensive medication use in the baseline period, 1990–91) were excluded from this investigation. A total of
2506 participants were included. We calculated the cumulative incidence rates of hypertension bien-nially in the first three periods (1992–93, 94–95 and 96–97) and the 2000–2001 period was the end of the study. The response rates in all periods were relatively high, from 86 to 96%.
Definition of hypertension and associated risk factors We defined the incident hypertension categories according to the criteria established by the Seventh Joint National Committee. Normotensive was de-fined as SBP o120 mm Hg and DBP o80 mm Hg. Hypertension was defined as SBP X140 mm Hg or DBP X90 mm Hg, and individuals on anti-hyperten-sive medications in the follow-up periods were also included as incident cases. Individuals with a fasting blood sugar level 4126 mg per 100 ml and/or use of oral hypoglycaemia agents or insulin injec-tions were defined as diabetes mellitus.15,23
Statistical analysis
The basic clinical and biochemical measures were listed according to the status of developing hyper-tension or not. We used the multivariate Weibull model to construct the prediction models because the Weibull model is suitable for interval-censored data.4 We specified the stepwise method for best
subset selection, by choosing variables entering in or removing from the model using a significant level as 0.05. We forced gender into the biochemical model for completeness. We constructed the parsimonious model for predicting the risk of hypertension according to two categories of covariate. First, the clinical model included gender, age, BMI, SBP and DBP, which were obtained from questionnaires and physical examinations and were statistically asso-ciated with the risk of hypertension. Second, the biochemical model included white blood cell count, fasting glucose and uric acid additionally. The lifestyle factors, such as smoking, drinking alcohol and physical activity, were also tested but the likelihood ratio tests showed that adding these variables into the model did not improve prediction beyond the parsimonious models. Furthermore, adding family history of hypertension did not increase appreciably the prediction measures; there-fore, we decided to exclude family history in the model. We still incorporated the family history in constructing the models from the John Hopkins8
and Framingham cohorts4 when comparing their
performance.
We constructed coefficient-based24 and
point-based models15,25for predicting the risk of
hyperten-sion using the clinical and biochemical variables. With regard to the coefficient-based model, the risk scores were derived from the estimated coefficients and were calculated to absolute risk in the Weibull model.4For the point-based model, the absolute risk
was summed by the derived point scores, which was
acquired by categorizing the covariates.25The details
were described in the Supplementary Materials. To enhance the comparability of our models with those from other studies, we compared the predic-tion models with available predicpredic-tion models, including John Hopkins8and Framingham cohorts4
(Supplementary Table S2), 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,26
which was a calibration measure to calculate how close the predicted risks were to the actual observed risks,27 and the results showed the calibration was
good (Supplementary Figure S1), except John Hop-kins model (Supplementary Table S3). Second, we compared the discrimination ability using the area under receiver operative characteristic curve (AUC). An AUC curve is a graph of sensitivity vs 1-specificity (or false-positive rate) for various cutoff definitions of a positive diagnostic test result.28
Statistical differences in the AUCs were compared using the method of DeLong et al.29The AUC was a
global summary measure for discrimination between individuals developing hypertension and those who did not.30We also conducted an internal validation
of the simple points model and obtained a bias-corrected estimate of AUC using a fivefold cross-validation procedure,31and the overall performance
by averaging the AUC estimates obtained from the five different partitions were similar to those of all data sets. Third, we compared the models by using the net reclassification improvement (NRI) and integrated discrimination improvement statistics.32
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 non-event subjects. We presented the
NRI according to the a priori risk categories of hypertension risk as 20, 40 and 60% risk. The integrated discrimination improvement can be inter-preted 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 between the new and old models. Finally, we plotted the Bland–Altman plot of the difference between point-based and coefficient-based risks vs the average of these two risks to compare the patterns between the point-based and coefficient-based risks.33
All statistical tests were two sided and P-values o0.05 were considered statistically significant. Analyses were performed with SAS version 9.1 (SAS Institute, Cary, NC, USA) and Stata version 9.1 (Stata Corporation, College Station, TX, USA).
Results
Patient characteristics
The mean and proportions of various clinical and biochemical measures are listed in Table 1. Compared with those who did not develop hyper-tension, participants with new-onset hypertension were likely to have a family history of hypertension, to be alcohol drinkers and older, and to have diabetes and a higher BMI, blood pressure, white blood cell count, fasting glucose and uric acid levels. A total of 2506 individuals (50.8% women) who were not hypertensive at the baseline (1990) were followed up and 1029 cases of new-onset hypertension developed during a median 6.15 (interquartile range, 4.04–9.02) years of period. Table 2 shows the parsimonious models using multivariate Weibull models for the prediction models.
Table 1 The means and proportions of various clinical and biochemical measures in the study participants according to the status of developing hypertension or not (n ¼ 2506)
Characteristic Unit New-onset HT () New-onset HT (+) P-value
n ¼ 1477 n ¼ 1029
Gender Men 48.8 50.0 0.55
Women 51.3 50.1
Family history of hypertension — 22.9 27.6 0.007
Smoking history — 37.6 37.0 0.78
Drinking history — 28.2 34.4 0.001
Regular physical activity habit — 13.3 14.1 0.59
Type II diabetes history — 8.3 13.8 o0.0001
Age Year 51.5 12.1 54.0 11.7 o0.0001
Body mass index kg m2 22.4 3.0 23.9 3.4 o0.0001
SBP mm Hg 112.2 10.6 120.4 10.1 o0.0001
DBP mm Hg 70.9 7.7 75.5 7.4 o0.0001
White blood cell count 1000 ml1 6.1 1.6 6.4 1.8 o0.0001
Fasting glucose mg per 100 ml 105.4 24.8 111.0 31.8 o0.0001
Uric acid mg per 100 ml 5.3 1.6 5.7 1.6 o0.0001
Abbreviations: DBP, diastolic blood pressure; HT, hypertension; SBP, systolic blood pressure.
Point-based prediction models and performance measures
The clinical and biochemical point-based score chart to estimate the risk of hypertension is shown in Tables 3 and 4. In the clinical model, gender (2 points), age (8 points), BMI (10 points), SBP (19 points) and DBP (7 points) were assigned. The biochemical measures, including white blood count (3 points), fasting glucose (1 point) and uric acid (3 points), in addition to the above clinical variables, were constructed. These risk charts allowed a manual estimation of the annual and cumulative risk of developing hypertension for each individual, as shown in Tables 3 and 4 for 1-, 4-, 5- and 10-year predicted risk. By using the clinical point-based risk chart, we determined that 50% of the sample had a o20% risk, 33% had a 20–40% risk, 13% had a 20–60% risk and 4% had a 460% risk of incident hypertension during a 5-year follow-up interval. The AUCs were 0.732 (95% confidence interval (CI), 0.712–0.752) for the point-based clinical model and 0.735 (95% CI, 0.715–0.755) for the point-based biochemical model, indicating a good discrimina-tion ability (Figure 1 and Supplementary Table S4). The AUC difference between the clinical and biochemical models was not significantly different (P ¼ 0.17). Comparing reclassification measures between the point-based and coefficient-based mod-els (Table 5), we found that the point-based clinical model had a similar NRI as the coefficient-based clinical model (NRI, 2.0%, P ¼ 0.30), and had a higher improvement than the point-based biochem-ical model (NRI, 3.7%, P ¼ 0.015). Finally, the Bland–Altman plot showed that compared with the coefficient-based model, the point-based prediction model overestimated the clinical risk (estimated coefficient, 0.0579±0.0037, Po0.001),
yet underestimated the biochemical risk (estimated coefficient, –0.0430±0.0044, Po0.001) (Figure 2).
Discussion
In the ethnic Chinese cohort data, we have con-structed the point-based prediction models using clinical and biochemical measures. The clinical model, which contained age, gender, BMI, SBP and DBP and had a better prediction performance, was suggested for further application in mass screening for the risk of hypertension. The availability of the manual risk charts to predict future risk of hyperten-sion, as has been the case for the prediction of coronary heart disease,34would improve the
predic-tion of hypertension risk, identify high-risk popula-tions and enhance preventive strategies.
Clinical risk factors
To our knowledge, this is the first hypertension prediction model specifically developed for an ethnic Chinese population. Several hypertension prediction models have previously been developed in various populations. Among 3202 Iranian dia-betic patients with 2.9 years of follow-up,3gender,
age at diagnosis of diabetes, BMI, fasting glucose and glycosated haemoglobin concentrations were associated with hypertension risk. Restriction to only type II diabetes patients might limit the generalizability to primary prevention in a general population. In the Framingham Heart Study cohort, Parikh et al.4proposed a prediction model on 1717
adult Caucasians without diabetes nor hypertension on a median 3.8-year follow-up period. The pro-posed prediction model included gender, age,
Table 2 Estimated coefficient, s.e., RR, 95% CI and significant levels in the clinical and biochemical models for the risk of hypertension, based on the Weibull regression model
Covariate Coefficient s.e. RR 95% CI P-value
Clinical model
Sex, women vs men 0.124 0.038 0.88 0.82 0.95 0.001
Age, +1 years 0.011 0.002 1.01 1.01 1.01 o0.0001
Body mass index (kg m2) 0.043 0.006 1.04 1.03 1.06 o0.0001
SBP (mm Hg) 0.029 0.002 1.03 1.02 1.03 o0.0001
DBP (mm Hg) 0.014 0.003 1.01 1.01 1.02 o0.0001
Biochemical model
Sex, women vs men 0.037 0.043 0.964 1.048 0.887 0.39
Age, +1 years 0.010 0.002 1.010 1.014 1.007 o0.0001
Body mass index (kg m2) 0.036 0.006 1.036 1.049 1.024 o0.0001
SBP (mmHg) 0.028 0.002 1.029 1.034 1.024 o0.0001
DBP (mm Hg) 0.013 0.003 1.013 1.019 1.007 o0.0001
White blood cell count (1000 ml1) 0.035 0.011 1.036 1.059 1.013 0.002
Fasting glucose (mg per 100 ml) 0.001 0.001 1.001 1.003 1.000 0.030
Uric acid (mg per 100 ml) 0.038 0.013 1.039 1.065 1.014 0.002
Abbreviations: CI, confidence interval; DBP, diastolic blood pressure; RR, relative risk; SBP, systolic blood pressure.
The Weibull regression uses an opposite metric to other proportional hazard models and results in opposite signs and interpretation of regression coefficients. The Weibull scale parameters are 0.592 and 0.589, and intercepts 8.173 and 8.604 for the clinical and biochemical models. 4
SBP, DBP, family history of hypertension, BMI and the interaction between age and DBP for the prediction model, with an AUC of 0.788 (95% CI, 0.733–0.803). Our clinical model was similar to this Framingham data; however, the interaction between age and DBP did not reach a significant level (P ¼ 0.74). Accordingly, we did not include the interaction items of age and DBP into our prediction model. In addition, we examined the role of body weight change in the first 2 years, and found that the 2-year one-unit BMI change increased risk of hypertension by 10%.35These results are consistent
with available evidence on weight change.36
How-ever, we did not include serial BMI change in the prediction model because serial measures increased the difficulty for the applicability of the prediction model.
Table 3 The simple points system according to the clinical model and the total points (left) and predicted risk (%) (right) for hypertension in the study participants
Risk factor Categories Points
Sex Men 2 Women 0 Age (year) 35–39 0 40–44 1 45–49 2 50–54 3 55–59 4 60–64 5 65–69 6 70–74 7 X75 8
Body mass index (kg m2) o18 0
18–19.9 2 20–21.9 3 22–23.9 5 24–25.9 6 26–27.9 8 X28 10 SBP (mm Hg) o105 0 105–109 3 110–114 5 115–119 10 120–124 11 125–129 14 130–134 16 135–139 19 DBP (mm Hg) o65 0 65–69 2 70–74 3 75–79 4 80–84 5 85–89 7 Point
total risk (%)1-year risk (%)4-year risk (%)5-year risk (%)10-year
0 0.3 3.0 4.4 13.4 1 0.3 3.3 4.8 14.6 2 0.4 3.6 5.2 15.9 3 0.4 3.9 5.7 17.2 4 0.4 4.3 6.2 18.7 5 0.5 4.7 6.8 20.3 6 0.5 5.1 7.4 22.0 7 0.6 5.6 8.1 23.8 8 0.6 6.1 8.8 25.7 9 0.7 6.7 9.6 27.8 10 0.7 7.3 10.4 29.9 11 0.8 7.9 11.4 32.3 12 0.9 8.7 12.4 34.7 13 0.9 9.4 13.5 37.3 14 1.0 10.3 14.6 40.0 15 1.1 11.2 15.9 42.8 16 1.2 12.2 17.3 45.8 17 1.4 13.3 18.8 48.8 18 1.5 14.4 20.3 52.0 19 1.6 15.7 22.0 55.2 20 1.8 17.0 23.8 58.5 21 1.9 18.5 25.8 61.8 22 2.1 20.1 27.8 65.1 23 2.3 21.7 30.0 68.4 24 2.5 23.5 32.4 71.7 25 2.8 25.4 34.8 74.9 26 3.0 27.5 37.4 77.9 27 3.3 29.7 40.1 80.9 28 3.6 32.0 43.0 83.7 29 4.0 34.4 45.9 86.2 30 4.3 37.0 49.0 88.6 31 4.7 39.6 52.1 90.7 32 5.2 42.5 55.3 92.6 33 5.6 45.4 58.6 94.2 34 6.2 48.4 61.9 95.6 35 6.7 51.6 65.3 96.7 36 7.3 54.8 68.6 97.6 37 8.0 58.0 71.8 98.3 38 8.7 61.4 75.0 98.9 39 9.5 64.7 78.1 99.3 40 10.4 68.0 81.0 99.5 41 11.3 71.3 83.8 99.7 42 12.3 74.5 86.3 99.8 43 13.4 77.6 88.7 99.9 44 14.5 80.5 90.8 100 45 15.8 83.3 92.6 100 46 17.2 85.9 94.3 100
Abbreviations: DBP, diastolic blood pressure; SBP, systolic blood pressure.
Table 4 The simple points system according to the biochemical model (left) and predicted risk (%) (right) for hypertension in the study participants
Risk factor Categories Points
Sex Men 1 Women 0 Age (year) 35–39 0 40–44 1 45–49 2 50–54 3 55–59 4 60–64 5 65–69 6 70–74 7 X75 8
Body mass index (kg m2) o18 0
18–19.9 1 20–21.9 3 22–23.9 4 24–25.9 5 26–27.9 7 X28 9 SBP (mm Hg) o105 0 105–109 3 110–114 5 115–119 10 120–124 11 125–129 14 130–134 16 135–139 20 DBP (mm Hg) o65 0 65–69 2 70–74 3 75–79 4 80–84 5 85–89 7 WBC (1000 ml1) o5.1 0 5.1–5.9 1 6.0–7.0 1 X7.1 3
Fasting glucose (mg per 100 ml) o95 0
95–101 0
102–110 0
X111 1
Uric acid (mg per 100 ml) o4.4 0
4.4–5.2 1
5.3–6.4 2
X6.5 3
The high incidence rate in our study participants may be attributed to the following reasons: First, we repeated blood pressure three times during the 6-year period, which would increase the incidence rates. Second, the 6-year incidence of hypertension in the US communities was 28 and 30% in African-American men and women aged 50–64 years, respectively,37indicating whites had a lower
hyper-tension rate. In fact, the original Framingham Heart
Study showed that the lifetime risk for developing hypertension were 90% in both 55- and 65-year-old participants who were free of hypertension at base-line during the 1976–1998,38 indicating a high
incidence rate in middle and elderly individuals. In addition, inclusion of diabetes cases may be one explanation for a high hypertension incidence. A 55-year-old men with 22 kg m2 BMI and
120/80 mm Hg would developed an 81% probability of the 10-year hypertension incidence.
Our data showed that the association between DBP and the risk of hypertension did not change significantly by age group. The likelihood ratio test comparing the model with the interaction terms of age and DBP and without those terms did not reach significant level (P ¼ 0.18). Our findings did not support previous evidence on the bimodal effect of age as an effect modifier for DBP on the risk of hypertension, as reported in previous studies.4,39
The possible explanation was due to truncated age distribution and high incident hypertension rates in our study participants.
Biochemical risk factors
For biochemical measure, Wang et al.9demonstrated
that some biomarkers were related to new-onset hypertension in the 1456 adults from a Framingham offspring cohort for 3 years of follow-up. Our biochemical models included white blood cell count, fasting glucose and uric acid, which may be obtained in a laboratory-based mass screening. Our findings did not support a family history of hypertension as a significant role for further devel-oping hypertension in our population. In another study based on young John Hopkins medical students for 450 years of follow-up, family history of hypertension was associated with a multivariate 1.8- to 2.4-fold risk for hypertension.2 The
non-significant association in our population was partly attributed to the relatively older age of the partici-pants. Although our prediction model did not include family history of hypertension, family history may be more valuable in risk prediction among younger adults than in older adults.2,8
In addition, our data did not support plasma lipid levels, including total cholesterol, triglyceride, high-density lipoprotein and low-high-density lipoprotein cholesterol, as significant predictors for hyperten-sion risk, in contrast to the findings from 16130 women for 10.8 years of follow-up, which showed that hyperlipidemia had a 1.34-fold risk of hypertension.5 In this Women’ Health study, lack
of controlling baseline blood pressure and other biochemical variables, such as inflammatory mar-kers and uric acid, may overestimate the role of lipids in the risk of hypertension, although apoli-poprotein B, liapoli-poprotein(a) and C-reactive protein were included.10Furthermore, the ethnic difference
in the metabolic syndrome components should be taken into consideration. We also incorporated uric
Point total 1-year risk (%) 4-year risk (%) 5-year risk (%) 10-year risk (%) 0 0.2 2.2 3.2 10.0 1 0.2 2.4 3.5 10.9 2 0.3 2.6 3.8 11.8 3 0.3 2.9 4.2 12.9 4 0.3 3.1 4.5 14.0 5 0.3 3.4 4.9 15.1 6 0.4 3.7 5.4 16.4 7 0.4 4.1 5.9 17.8 8 0.4 4.4 6.4 19.3 9 0.5 4.8 7.0 20.8 10 0.5 5.2 7.6 22.5 11 0.6 5.7 8.2 24.3 12 0.6 6.2 9.0 26.2 13 0.7 6.8 9.7 28.3 14 0.7 7.4 10.6 30.5 15 0.8 8.0 11.5 32.8 16 0.9 8.7 12.5 35.2 17 0.9 9.5 13.6 37.7 18 1.0 10.3 14.7 40.4 19 1.1 11.2 16.0 43.1 20 1.2 12.2 17.3 46.0 21 1.3 13.3 18.8 49.0 22 1.5 14.4 20.3 52.1 23 1.6 15.6 22.0 55.2 24 1.7 16.9 23.7 58.4 25 1.9 18.3 25.6 61.7 26 2.1 19.8 27.6 64.9 27 2.3 21.5 29.7 68.2 28 2.5 23.2 32.0 71.3 29 2.7 25.0 34.4 74.5 30 2.9 27.0 36.9 77.5 31 3.2 29.1 39.5 80.4 32 3.5 31.3 42.2 83.1 33 3.8 33.7 45.1 85.7 34 4.2 36.1 48.0 88.0 35 4.5 38.7 51.1 90.2 36 5.0 41.4 54.2 92.0 37 5.4 44.2 57.4 93.7 38 5.9 47.2 60.6 95.1 39 6.4 50.2 63.8 96.3 40 7.0 53.3 67.1 97.3 41 7.6 56.4 70.3 98.0 42 8.3 59.7 73.4 98.6 43 9.0 62.9 76.5 99.1 44 9.8 66.1 79.4 99.4 45 10.6 69.4 82.2 99.6 46 11.6 72.5 84.8 99.8 47 12.6 75.6 87.3 99.9 48 13.6 78.6 89.5 99.9 49 14.8 81.4 91.4 100 50 16.0 84.1 93.2 100 51 17.4 86.6 94.7 100 52 18.8 88.9 95.9 100
Abbreviations: DBP, diastolic blood pressure; SBP, systolic blood pressure; WBC, white blood cell.
Table 4 Continued 6
acid concentrations in our biochemical model, because uric acid has been strongly associated with new-onset hypertension.40Among the 2062 men free
of hypertension in the Normative Aging Study for a 21-year follow-up period, uric acid increased new-onset hypertension risk by 1.05-fold,40similar to our
estimate in the multivariate model. We believe that most of the variables included in our models were feasible in the usual care practice in primary prevention settings. In addition, our data showed that as age at study initiation increased, the incidence rates of hypertension risk and the net 20/10 mm Hg blood pressure change increased sig-nificantly, especially for women, indicating that age at onset of observation initiation is a critical factor in evaluating change of blood pressure over time. Moreover, we stratified the incident hypertension into two subtypes: isolated systolic hypertension (SBP X140 mm Hg and DBP o90 mm Hg) and diastolic hypertension (defined as isolated diastolic and more commonly mixed diastolic/systolic hypertension), and we found that the cumulative rates of isolated systolic hypertension increased
significantly as age increased; however, diastolic hypertension rates decreased as age progressed. The largest percentage of new-onset hypertension among elderly participants was attributed to isolated systolic hypertension.
The lack of lifestyle risk factors in the prediction model in our study, compared with previous data,41
may be explained by the following reasons. First, potential measurement misclassification in asses-sing lifestyle risk factors could induce measurement errors and non-differential misclassifications redu-cing the power for predicting hypertension risk. Second, attenuation of lifestyle factor effects was induced by mediating factors, such as blood pres-sure, obesity and biochemical markers. When included in the prediction models, these clinical and biochemical variables made lifestyle factors less significant for hypertension outcome. Although our prediction models did not include lifestyle factors, we still emphasize the importance of lifestyle factors for the risk of hypertension.41 As the baseline
distribution of lifestyle factors in the study sample was identified as having the highest hypertension
Figure 1 Receiver-operating characteristic curves for various models applied to the study population. Green, coefficient-based biochemical (AUC, 0.741); dark blue, coefficient-based, clinical (AUC, 0.737), red, based, biochemical (AUC, 0.735); grey, point-based, clinical (AUC, 0.732); yellow, Framingham (AUC, 0.709); blue, John Hopkins, (AUC, 0.707); black, reference.
Table 5 Summary of statistics comparing risk prediction algorithms to prediction based on the models
Model comparison NRI( %) 95% CI P-value IDI (%) 95% CI P-value
Point-based vs coefficient-based, clinical 2.0 1.5 5.4 0.30 0.6 0.3 0.8 o0.0001
Clinical point-based vs biochemical point-based 7.0 3.7 10.3 0.0002 1.0 0.7 1.3 o0.0001
Coefficient-based vs point-based, biochemical 3.7 0.7 6.7 0.015 0.9 0.6 1.2 o0.0001
Biochemical coefficient-based vs clinical coefficient-based 1.3 4.2 1.7 0.40 0.5 0.2 0.8 0.001 Abbreviations: CI, confidence interval; IDI, integrated discrimination improvement; NRI, net reclassification improvement.
NRI with a priori 5-year cumulative risk categories according too20, 20–40, 40–60 and X60%.
risk by the prediction models, the potential oppor-tunities for lifestyle intervention during primary prevention should be identified early through pre-diction model screening.
Strengths and limitations
To our knowledge, this is the first hypertension prediction model specifically developed for an ethnic Chinese population. Owing to the large sample size, the estimates from our prediction models were found to be stable as demonstrated by the internal validation study. In addition, the use of a community-based population could reduce the possibility of selection bias. In addition, the
constructed simple points system and predicted risk was still available among the participants without diabetes (Supplementary Tables S5 and S6) and cardiovascular disease at baseline. However, several potential limitations of this study should be mentioned. First, the point-based models were inferior to the coefficient-based model based on NRI and integrated discrimination improvement values, although the AUCs were similar. Second, we did not include extensive biomarker data in the model and the blood pressure ascertainment was performed once every 2 years. Third, we did not separate genders and did not include lifestyle factors in the prediction model. Lack of lifestyle factors in the model may decrease the application of
Figure 2 Bland–Altman plot of the difference between point-based and coefficient-based risks vs the average of these two risks in the clinical (upper) and biochemical (lower) models. (Black circle: subject developed hypertension; grey circle: subject did not develop hypertension.)
the prediction model in primary prevention. Finally, our study participants were middle and elderly Chinese population and the community is geogra-phically unique, and so the external generalization to general population of our results was unknown. Furthermore, prevention strategies would be more effective on relatively young population. Therefore, further validation studies for these prediction models are warranted.
In conclusion, we have constructed the clinical and biochemical prediction models for predicting the 10-year incidence of hypertension among ethnic Chinese people. We recommend the point-based clinical model as the first step to identify high-risk populations for hypertension because of its simplicity and easily obtained measures in clinical practice and may be helpful to identify high-risk populations and improve prevention and treatment strategies for Chinese populations.
Conflict of interest
The authors declare no conflict of interest.
Acknowledgements
We thank the staff of the Department of Internal Medicine, National Taiwan University Hospital, and the participants of the CCCC study for their contributions. This study was partly supported by the National Science Council (NSC 97-2314-B-002-130-MY3, 97-3112-B-002-034-). This study was also supported in part by Taiwan Department of Health Clinical Trial Research Center of Excellence (DOH 99-TD-B-111-004).
References
1 Mancia G, De Backer G, Dominiczak A, Cifkova R, Fagard R, Germano G et al. 2007 Guidelines for the management of arterial hypertension: the Task Force for the Management of Arterial Hypertension of the European Society of Hypertension (ESH) and of the European Society of Cardiology (ESC). Eur Heart J 2007; 28: 1462–1536.
2 Wang NY, Young JH, Meoni LA, Ford DE, Erlinger TP, Klag MJ. Blood pressure change and risk of hyperten-sion associated with parental hypertenhyperten-sion: the Johns
Hopkins Precursors Study. Arch Intern Med 2008; 168: 643–648.
3 Janghorbani M, Amini M. Hypertension in type 2 diabetes mellitus in Isfahan, Iran: incidence and risk factors. Diabetes Res Clin Pract 2005; 70: 71–80. 4 Parikh NI, Pencina MJ, Wang TJ, Benjamin EJ, Lanier
KJ, Levy D et al. A risk score for predicting near-term incidence of hypertension: the Framingham Heart Study. Ann Intern Med 2008; 148: 102–110.
5 Sesso HD, Buring JE, Chown MJ, Ridker PM, Gaziano JM. A prospective study of plasma lipid levels and hypertension in women. Arch Intern Med 2005; 165: 2420–2427.
6 Mellen PB, Bleyer AJ, Erlinger TP, Evans GW, Nieto FJ, Wagenknecht LE et al. Serum uric acid predicts incident hypertension in a biethnic cohort: the athero-sclerosis risk in communities study. Hypertension 2006; 48: 1037–1042.
7 Sesso HD, Buring JE, Rifai N, Blake GJ, Gaziano JM, Ridker PM. C-reactive protein and the risk of develop-ing hypertension. JAMA 2003; 290: 2945–2951. 8 Pearson TA, LaCroix AZ, , Mead LA, , Liang KY. The
prediction of midlife coronary heart disease and hyper-tension in young adults: the Johns Hopkins multiple risk equations. Am J Prevent Med 1990; 6: 23–28. 9 Wang TJ, Gona P, Larson MG, Levy D, Benjamin EJ,
Tofler GH et al. Multiple biomarkers and the risk of incident hypertension. Hypertension 2007; 49: 432–438.
10 Paynter NP, Cook NR, Everett BM, Sesso HD, Buring JE, Ridker PM. Prediction of incident hypertension risk in women with currently normal blood pressure. Am J Med 2009; 122: 464–471.
11 Kivimaki M, Batty GD, Singh-Manoux A, Ferrie JE, Tabak AG, Jokela M et al. Validating the Framingham Hypertension Risk Score: results from the Whitehall II study. Hypertension 2009; 54: 496–501.
12 Krishnan E, Kwoh CK, Schumacher HR, Kuller L. Hyperuricemia and incidence of hypertension among men without metabolic syndrome. Hypertension 2007; 49: 298–303.
13 Sundstrom J, Sullivan L, D’Agostino RB, Levy D, Kannel WB, Vasan RS. Relations of serum uric acid to longitudinal blood pressure tracking and hypertension incidence. Hypertension 2005; 45: 28–33.
14 Schulze MB, Shai I, Manson JE, Li T, Rifai N, Jiang R et al. Joint role of non-HDL cholesterol and glycated haemoglobin in predicting future coronary heart disease events among women with type 2 diabetes. Diabetologia 2004; 47: 2129–2136.
15 Chien K, Cai T, Hsu H, Su T, Chang W, Chen M et al. A prediction model for type 2 diabetes risk among Chinese people. Diabetologia 2009; 52: 443–450. 16 Chien KL, Hsu HC, Sung FC, Su TC, Chen MF, Lee YT.
Incidence of hypertension and risk of cardiovascular events among ethnic Chinese: report from a commu-nity-based cohort study in Taiwan. J Hypertens 2007; 25: 1355–1361.
17 Lee YT, Lin RS, Sung FC, Yang CY, Chien KL, Chen WJ et al. Chin-Shan Community Cardiovascular Cohort in Taiwan: baseline data and five-year follow-up morbidity and mortality. Clin Epidemiol 2000; 53: 836–846.
18 Chien KL, Hsu HC, Su TC, Chen MF, Lee YT, Hu FB. Apolipoprotein B and non-high-density lipoprotein cholesterol and risk of coronary heart disease in Chinese. J Lipid Res 2007; 48: 2499–2505.
What is known about topic
KHypertension incidence was associated with clinical and biochemical factors.
KAvailable prediction models were constructed in Caucasians. What this study adds
KA simple point-based prediction model incorporating clinical measures was useful for Chinese.
KGender, age, body mass index, SBP and DBP were assigned in this prediction model.
19 Chien KL, Hsu HC, Chen PC, Su TC, Chang WT, Chen MF et al. Urinary sodium and potassium excretion and risk of hypertension in Chinese: report from a community-based cohort study in Taiwan. J Hypertens 2008; 26: 1750–1756.
20 Chien KL, Lee YT, Sung FC, Su TC, Hsu HC, Lin RS. Lipoprotein (a) level in the population in Taiwan: relationship to sociodemographic and atherosclerotic risk factors. Atherosclerosis 1999; 143: 267–273. 21 Chien KL, Hsu HC, Sung FC, Su TC, Chen MF, Lee YT.
Hyperuricemia as a risk factor on cardiovascular events in Taiwan: the Chin-Shan Community Cardiovascular Cohort Study. Atherosclerosis 2005; 183: 147–155. 22 Fossati P, Prencipe L, Berti G. Use of
3,5-dichloro-2-hydroxybenzenesulfonic acid/4-aminophenazone chromogenic system in direct enzymic assay of uric acid in serum and urine. Clin Chem 1980; 26: 227–231. 23 Chien KL, Chen MF, Hsu HC, Chang WT, Su TC, Lee YT et al. plasma uric acid and the risk of type 2 diabetes in a Chinese community. Clin Chem 2008; 54: 310–316.
24 Schulze MB, Hoffmann K, Boeing H, Linseisen J, Rohrmann S, Mohlig M et al. An accurate risk score based on anthropometric, dietary, and lifestyle factors to predict the development of type 2 diabetes. Diabetes Care 2007; 30: 510–515.
25 Sullivan LM, Massaro JM, D’Agostino Sr RB. Presenta-tion of multivariate data for clinical use: the Framing-ham Study risk score functions. Stat Med 2004; 23: 1631–1660.
26 Hosmer Jr DW, Lemeshow S. The multiple logistic regression model. In: David W Hosmer Jr, Stanley Lemeshow (eds). Applied Logistic Regression. 1st edn. John Wiley & Sons: New York, 1989, pp 25–37. 27 McGeechan K, Macaskill P, Irwig L, Liew G, Wong TY.
Assessing new biomarkers and predictive models for use in clinical practice: a clinician’s guide. Arch Intern Med 2008; 168: 2304–2310.
28 Hanley JA, McNeil BJ. A method of comparing the areas under receive operating characteristic curves derived from the same cases. Radiology 1983; 148: 839–843. 29 DeLong ER, DeLong DM, Clarke-Pearson DL.
Compar-ing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988; 44: 837–845.
30 Cook NR. Use and misuse of the receiver operating characteristic curve in risk prediction. Circulation 2007; 115: 928–935.
31 Efron B, Tibshirani RJ. An introduction to the Bootstrap. Chapman & Hall/CRC: New York, 1994. p 456.
32 Pencina MJ, D’ Agostino Sr RB, D’ Agostino Jr RB, Vasan RS. Evaluating the added predictive ability of a new marker: from area under the ROC curve to reclassification and beyond. Stat Med 2008; 27: 157–172.
33 Bland JM, Altman DG. Comparing methods of measurement: why plotting difference against standard method is misleading. Lancet 1995; 346: 1085–1087.
34 D’Agostino Sr RB, Vasan RS, Pencina MJ, Wolf PA, Cobain M, Massaro JM et al. General cardiovascular risk profile for use in primary care: the Framingham Heart Study. Circulation 2008; 117: 743–753.
35 Chen PC, Sung FC, Su TC, Chien KL, Hsu HC, Lee YT. Two-year change in body mass index and subsequent risk of hypertension among men and women in a Taiwan community. J Hypertens 2009; 27: 1370–1376.
36 Czernichow S, Mennen L, Bertrais S, Preziosi P, Hercberg S, Oppert JM. Relationships between changes in weight and changes in cardiovascular risk factors in middle-aged French subjects: effect of dieting. Int J Obes Relat Metab Disord 2002; 26: 1138–1143. 37 Apostolides AY, Cutter G, Daugherty SA, Detels R,
Kraus J, Wassertheil-Smoller S et al. Three-year incidence of hypertension in thirteen US commu-nities. On behalf of the Hypertension Detection and Follow-up Program cooperative group. Prev Med 1982; 11: 487–499.
38 Vasan RS, Beiser A, Seshadri S, Larson MG, Kannel WB, D’Agostino RB et al. Residual lifetime risk for developing hypertension in middle-aged women and men: the Framingham Heart Study. Jama 2002; 287: 1003–1010.
39 Franklin SS, Gustin IV W, Wong ND, Larson MG, Weber MA, Kannel WB et al. Hemodynamic patterns of age-related changes in blood pressure. The Framing-ham Heart Study. Circulation 1997; 96: 308–315. 40 Perlstein TS, Gumieniak O, Williams GH, Sparrow D,
Vokonas PS, Gaziano M et al. Uric acid and the development of hypertension: the normative aging study. Hypertension 2006; 48: 1031–1036.
41 Forman JP, Stampfer MJ, Curhan GC. Diet and lifestyle risk factors associated with incident hypertension in women. Jama 2009; 302: 401–411.
Supplementary Information accompanies the paper on the Journal of Human Hypertension website (http:// www.nature.com/jhh)
