Accepted Manuscript

Development and validation of modified risk prediction models for cardiovascular

disease and its subtypes: The Hisayama Study

Takanori Honda, Daigo Yoshida, Jun Hata, Yoichiro Hirakawa, Yuki Ishida, Mao
Shibata, Satoko Sakata, Takanari Kitazono, Toshiharu Ninomiya

PII: S0021-9150(18)31433-3
DOI: 10.1016/j.atherosclerosis.2018.10.014
Reference: ATH 15752

To appear in: Atherosclerosis

Received Date: 31 May 2018

Revised Date: 17 September 2018
Accepted Date: 16 October 2018

Please cite this article as: Honda T, Yoshida D, Hata J, Hirakawa Y, Ishida Y, Shibata M, Sakata
S, Kitazono T, Ninomiya T, Development and validation of modified risk prediction models for
cardiovascular disease and its subtypes: The Hisayama Study, Atherosclerosis (2018), doi: https://
doi.org/10.1016/j.atherosclerosis.2018.10.014.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 1
ACCEPTED MANUSCRIPT
1 Development and validation of modified risk prediction models for cardiovascular

2 disease and its subtypes: The Hisayama Study

4 Takanori Hondaa, Daigo Yoshidaa,b, Jun Hataa,b,c, Yoichiro Hirakawaa,b,c, Yuki Ishidaa, Mao

5 Shibataa,b, Satoko Sakataa,c, Takanari Kitazonoc, Toshiharu Ninomiyaa,b

PT
6

RI
7 Department of Epidemiology and Public Health, Graduate School of Medical Sciences,

8 Kyushu University, 3-1-1 Maidashi, Higashi-ku, Fukuoka City, Fukuoka, 812-8582 Japan

SC
b
9 Center for Cohort Studies, Graduate School of Medical Sciences, Kyushu University, 3-1-1

10 Maidashi, Higashi-ku, Fukuoka City, Fukuoka, 812-8582 Japan

11
c

U
Department of Medicine and Clinical Science, Graduate School of Medical Sciences, Kyushu
AN
12 University, 3-1-1 Maidashi, Higashi-ku, Fukuoka City, Fukuoka, 812-8582 Japan
M

13

14 Yuki Ishida ishida-y@eph.med.kyushu-u.ac.jp

D

15 Toshiharu Ninomiya
TE

16

17 *Corresponding author
EP

18 Dr. Toshiharu Ninomiya

3-1-1 Maidashi, Higashi-ku, Fukuoka City, Fukuoka. nino@eph.med.kyushu-u.ac.jp

C

19
AC

20

21 Abstract

22 Background and aims: Predicting cardiovascular events is of practical benefit for disease

23 prevention. The aim of this study was to develop and evaluate an updated risk prediction

24 model for cardiovascular diseases and its subtypes.

25 Methods:A total of 2,462 community residents aged 40-84 years were followed up for 24
 2
ACCEPTED MANUSCRIPT
1 years. A Cox’s proportional hazards regression model was used to develop risk prediction

2 models for cardiovascular diseases, and separately for stroke and coronary heart diseases. The

3 risk assessment ability of the developed model was evaluated, and a bootstrapping method

4 was used for internal validation. The predicted risk was translated into a simplified scoring

5 system. A decision curve analysis was used to evaluate clinical usefulness.

PT
6 Results: The multivariable model for cardiovascular diseases included age, sex, systolic blood

RI
7 pressure, hemoglobin A1c, high-density lipoprotein cholesterol, low-density lipoprotein

8 cholesterol, smoking habits, and regular exercise as predictors. The models for stroke and

SC
9 coronary heart diseases incorporated both shared and unique variables. The developed models

10 showed good discrimination with little evidence of overfitting (optimism-corrected Harrell’s

11
U
C statistics 0.726-0.777) and calibrations (Hosmer-Lemeshow test, p = 0.44 – 0.90). The
AN
12 decision curve analysis revealed that the predicted risk-based decision-making would have
M

13 higher net benefit than either a CVD intervention strategy for all individuals or no individuals.

14 Conclusions: The developed risk prediction models showed a good performance and
D

15 satisfactory internal validity, which may help understand individual risk and setting
TE

16 personalized goals, and promote risk stratification in public health strategies for CVD

17 prevention.
C EP
AC
 3
ACCEPTED MANUSCRIPT
1 1. Introduction

2 Cardiovascular disease (CVD) is estimated to be one of the leading causes of death

3 throughout the world, and places a great burden on local communities. To reduce the burden

4 of cardiovascular mortality, it is important to detect a high-risk population for CVD and

5 modify the cardiovascular risk factors. In addition, the recent guidelines place emphasis on

PT
6 risk-based decision-making for determining the initial treatment (e.g., cholesterol-lowering

RI
7 treatment) [1,2]. Therefore, risk prediction algorithms based on multiple risk models would be

8 a clinically useful tool for estimating future risk of the development of CVD and its subtypes

SC
9 [3].

10 Several epidemiological studies conducted in Westerns populations have developed

11
U
risk prediction models for cardiovascular disease and its subtypes (i.e., stroke and coronary
AN
12 heart disease) [3,4]. On the other hand, it has been well-acknowledged that the absolute risk
M

13 of stroke and coronary heart disease and the magnitude of the influence of risk factors on

14 these diseases vary among ethnicities [5]. In addition, epidemiological evidence has suggested
D

15 that stroke and coronary heart disease share unique sets of risk factors [5,6]. Although there
TE

16 have been several prediction models for the future incidence of either stroke or coronary heart

17 disease in Asian populations [7-12], it is of value to review and compare the prediction
EP

18 models for the subtypes in the same cohort.

Our research group previously developed a risk prediction model for the
C

19
AC

20 development of overall CVD events using 14-year follow-up data from a cohort study of

21 Japanese community samples, but the model did not address the subtypes of CVD [13]. Since

22 we recently extended the follow-up period to 24 years, we can now develop separate risk

23 prediction models for stroke and coronary heart disease. The aim of this study was thus to

24 develop and validate an updated prediction model for overall CVD risk, and separate models

25 for stroke and coronary heart diseases, and to evaluate the performance and clinical usefulness
 4
ACCEPTED MANUSCRIPT
1 of the updated models using the data from the extended cohort.

3 2. Materials and methods

4 2.1. Participants

5 The Hisayama Study, an epidemiological study of cerebro- and cardiovascular diseases, was

PT
6 established in 1961 in the town of Hisayama, a suburban community adjacent to the city of

RI
7 Fukuoka, a metropolitan area of Kyushu Island in southern Japan. The population of the town

8 was approximately 7500 in 1988, and full community surveys of the residents have been

SC
9 repeated since 1961. In 1988, a screening survey for the present study was performed in the

10 town. A detailed description of this survey was published previously [14]. Briefly, 2,742 town

11
U
residents aged 40 years and over consented to participate in the examination and underwent a
AN
12 comprehensive assessment. After excluding 54 subjects aged ≥85 years, 99 who already had
M

13 CVD at baseline, one who died before starting follow-up, 80 who had already eaten breakfast

14 before providing a blood sample, 43 whose low-density lipoprotein (LDL) cholesterol values
D

15 could not be estimated, and 3 with missing information on one or more covariates, 2,462
TE

16 subjects were included in the present analysis.

17
EP

18 2.2. Ethical considerations

This study was conducted with the approval of the Kyushu University Institutional Review
C

19
AC

20 Board for Clinical Research. Written informed consent was provided by the study

21 participants.

22

23 2.3. Follow-up survey

24 The subjects were followed up prospectively from December 1988 to November 2012 by

25 repeated health examinations. A detailed description of the study methods has been published
 5
ACCEPTED MANUSCRIPT
1 previously [13,14]. In brief, the health status of any subject who had not undergone a regular

2 examination or who had moved out of town was checked yearly by mail or telephone. We also

3 established a daily monitoring system among the study team and local physicians or members

4 of the town’s Health and Welfare Office. When a subject died, an autopsy was performed at

5 the Departments of Pathology of Kyushu University. During the follow-up period, 977

PT
6 subjects died, of whom 667 (68.3%) underwent autopsy. No subject was lost to follow up.

RI
7

8 2.4. Ascertainment of endpoints

SC
9 The endpoints of the present analysis included overall CVD, stroke, and coronary heart

10 disease (CHD) [13]. CVD was defined as first-ever development of stroke or CHD. The

11
U
criteria for a diagnosis of CHD included first-ever acute myocardial infarction, silent
AN
12 myocardial infarction, sudden cardiac death within 1 h after the onset of acute illness, or
M

13 coronary artery disease followed by coronary artery bypass surgery or angioplasty. Acute

14 myocardial infarction was diagnosed when a subject met at least two of the following criteria:
D

15 (1) typical symptoms, including prolonged severe anterior chest pain; (2) abnormal cardiac
TE

16 enzymes more than twice the upper limit of the normal range; (3) evolving diagnostic

17 electrocardiographic changes; and (4) morphological changes, including local asynergy of

EP

18 cardiac wall motion on echocardiography, persistent perfusion defect on cardiac scintigraphy,

or myocardial necrosis or scars >1 cm long accompanied by coronary atherosclerosis at

C

19
AC

20 autopsy. Silent myocardial infarction was defined as myocardial scarring without any

21 historical indication of clinical symptoms or abnormal cardiac enzyme changes, and was

22 detected by electrocardiography, echocardiography, cardiac scintigraphy or autopsy. Stroke

23 was defined as a sudden onset of nonconvulsive and focal neurological deficit due to ischemia

24 or hemorrhage persisting for >24 h. The diagnosis of stroke was based on the clinical history,

25 neurological examination and all available clinical data, including brain CT/MRI and autopsy
 6
ACCEPTED MANUSCRIPT
1 findings.

3 2.5. Potential predictors

4 The height and weight were measured with the participant in light clothes without shoes, and

5 the body mass index (BMI; weight (kg)/height (m)2) was calculated. Obesity was defined as a

PT
6 BMI ≥25.0 kg/m2. Sitting blood pressure was measured three times at the right upper arm

RI
7 using a mercury sphygmomanometer after 5 minutes of rest; an average of three

8 measurements was used for the analysis. Plasma fasting glucose levels were measured by

SC
9 means of the glucose oxidase method. Diabetes was defined as fasting plasma glucose

10 concentration ≥126 mg/dL (7.0 mmol/L), 2-h post-load glucose concentration ≥200 mg/dL

11
U
(11.1mmol/L), or taking antidiabetic medication. Hemoglobin A1c (HbA1c) was measured by
AN
12 high pressure liquid chromatography. The value for HbA1c was corrected to the National
M

13 Glycohemoglobin Standardization Program equivalent value (%) with the following formula:

14 HbA1c (%) = 1.02 ×HbA1c (%) + 0.25%. Serum total cholesterol, high-density lipoprotein
D

15 (HDL) cholesterol and triglyceride levels were determined enzymatically. Non-HDL

TE

16 cholesterol concentrations were calculated by subtracting HDL cholesterol from the total

17 cholesterol. The LDL cholesterol level was estimated using the Friedewald formula.
EP

18 Information on smoking and drinking habits was obtained using a standard questionnaire and

was classified as either current or not. Regular exercise was defined as engaging in sports or
C

19
AC

20 other forms of athletic activity at least three times per week during leisure time.

21

22 2.6. Statistical analysis

23 We computed descriptive statistics for the entire cohort. Cox’s proportional hazards regression

24 models were used to predict the overall risk of CVD, as well as the separate risks of stroke

25 and CHD. Potential covariates included age, sex, obesity, systolic blood pressure, diabetes,
 7
ACCEPTED MANUSCRIPT
1 HDL cholesterol, LDL cholesterol, current smoking, drinking habits, and regular exercise.

2 Based on Richter’s method [15], a bootstrapping method was used to select variables for the

3 final multivariable models. The procedure consisted of selecting a group of “new” subjects

4 that was the same size as the “original” sample (n=2,462) from the original cohort with

5 allowance for replacement, and repeating this selection process 200 times. The proportional

PT
6 hazards regression model with a backward variable selection (p < 0.10) was repeatedly fitted

RI
7 in the 200 bootstrapping samples. A multivariable model that included the covariates

8 employed in ≥ 40% of the repetitions (i.e., 80 times) was fitted to the original dataset, which

SC
9 constructed the final model. Alternative models were fitted by substituting LDL cholesterol

10 for non-HDL cholesterol and/or HbA1c for diabetes as predictors to promote utility in various

11 clinical settings.
U
AN
12 The prediction of risk of diseases from a proportional hazards regression model can
M

13 be written as:

(∑ ∑ ̅)
14 ̂ =1− ( ) ,
D

15 where S0(t) is a baseline survivor function at time t. The values of ∑ were taken as the
TE

16 individual risk estimates (IREs) for presentation in a summary table. ∑ ̅ is a baseline risk

17 estimate calculated as the sum of the product of coefficients and mean values of each variable.
EP

18 Discrimination was assessed by Harrell’s overall C statistics. For internal validation,

C

19 the optimism of the developed models was estimated in line with Harrell’s method by using
AC

20 the 200 bootstrapping samples [16]. The estimated optimism was then subtracted from the C

21 statistics calculated in the original cohort to obtain the optimism-corrected C statistics. The

22 calibration at 10-year follow-up was assessed by illustrating calibration plots and by a

23 Hosmer-Lemeshow χ2 test with 8 degrees of freedom [3].

24 In line with Sullivan’s approach [17], the risk prediction models were translated into

25 point-based risk score sheets. The following age groups were designated: 40-49, 50-59, 60-69,
 8
ACCEPTED MANUSCRIPT
1 70-79, and 80-84 years. Systolic blood pressure (SBP) and HDL and LDL cholesterols were

2 categorized as follows: SBP: <120, 120-129, 130-139, 140-149, 150-159, ≥160 mmHg; HDL

3 cholesterol: ≥ 60, 40-59, <40 mg/dL; LDL cholesterol: <120, 120-139, ≥140 mg/dL. The

4 midpoints of each category were set as reference values. Reference values in the lowest and

5 highest categories were set as the nearest round figures to the median value in that category.

PT
6 For each variable, a category that represents lower risk was set as a reference group (0

RI
7 points)—e.g., the age category of 40-49, women, systolic blood pressure of <120 mmHg, and

8 so on.

SC
9 We calculated “vascular age” to facilitate interpretation of the predicted risk by

10 providing supportive information for practical use. Vascular age (or heart age) was originally

11
U
conceptualized in the Framingham Heart Study as the age of a person with the same predicted
AN
12 risk but with all other risk factor levels in “optimal” ranges [3]. In this study, however, we
M

13 modified this concept as with all other risk factor levels in average ranges: thus, vascular age

14 in the present study referred to the age of a person with the same predicted risk but with all
D

15 other risk factor levels in “average” ranges in the corresponding age group. To calculate
TE

16 vascular age as a function of the predicted risk, a parametric survival model with Weibull

17 distribution was used separately for men and women. To facilitate understanding of the
EP

18 integrated algorithms for overall CVD, stroke, and CHD, clinical examples with different risk

profiles were presented.

C

19
AC

20 Finally, clinical usefulness was evaluated using a decision curve analysis [18].

21 Decision curve analysis is a novel statistical technique to assess clinical value by comparing

22 the benefits of correctly detecting people who will develop CVD and its subtypes with the

23 harms due to a false-positive classification (i.e., incorrect detection of a disease-free

24 individual as positive) [19]. The net benefit of a risk equation at a given risk threshold is

25 given by calculating the difference between the proportion of true positives and the proportion
 9
ACCEPTED MANUSCRIPT
1 of false positives multiplied by the odds defined by the risk threshold value [18]. Decision

2 curves were generated by estimating the net benefit of making a decision about the initiation

3 of interventions based on the prediction models, as well as point-based estimations, by

4 different cut-offs of predicted 10-year risks for developing the diseases (i.e., threshold

5 probability). The net benefits were compared with alternative strategies assuming that no or

PT
6 all individuals would undergo interventions. In addition, we also generated a decision curve

RI
7 based on the Framingham general cardiovascular risk profile [3]. To calculate individual risks

8 for the three endpoints using the Framingham risk profile, we used regression coefficients

SC
9 reported in the original paper [3] and 10-year baseline survival rates among our study

10 population estimated by proportional hazards regression models.

11
U
All analyses were performed using SAS version 9.4. Two-sided p values of < 0.05
AN
12 were considered statistically significant.
M

13

14 3. Results
D

15 Baseline characteristics of the study sample are shown in Table 1. During the 24-year
TE

16 follow-up, 489 incident cases of CVD, 324 incident cases of stroke, and 216 incident cases of

17 CHD were documented. The univariable- and multivariable-adjusted hazard ratios and beta
EP

18 regression coefficients of potential predictors and overall CVD, stroke, and CHD are shown in

Supplementary table 1 and Table 2, respectively. The final multivariable model for CVD
C

19
AC

20 included age, sex, systolic blood pressure, the presence of diabetes, HDL and LDL cholesterol,

21 smoking habits, and regular exercise as predictors. These variables were identical to those

22 selected in ≥40% of bootstrapping samples (Supplementary table 2). Finally, the risk

23 prediction model for CVD to compute IRE was as follows:

24 =

25 (0.067 × !" # $%) + (0.559 × 1 if + , % ) + (0.013 × %#% .,!/ 0,..1 $ %%2$ !" ++3 ) +

26 (0.412 × 1 !6 1! 0 %) + (−0.009 × 378 /ℎ., % $., !" + /18) +

 10
ACCEPTED MANUSCRIPT
1 (0.003 × 878 /ℎ., % $., !" + /18) + (0.338 × 1 !6 /2$$ " %+.< $) +

2 (−0.420 × !6 ℎ =!" $ 2, $ $/!% ) .

3 Using the risk prediction model for CVD, the 10-year risks (p̂ ) of overall CVD events of

4 study participants were estimated by calculating an IRE as follows:

(>?@ A.BCD)
5 ̂ = 1 − 0.936 .

PT
6 The risk prediction models for stroke and coronary heart disease were developed in a

7 similar manner. Systolic blood pressure, diabetes, and smoking habits were selected in both of

RI
8 the separate models for stroke and CHD. A regular exercise habit was included in the model

SC
9 for stroke but not that for CHD. On the other hand, HDL and LDL cholesterol levels were

10 only included in the model for CHD, and not in the model for stroke. Obesity and drinking

11 habits were not included in any models.

U
AN
12 The value of Harrell’s C statistics for the risk prediction model for overall CVD was

13 0.743 (standard error, SE 0.011) in the original cohort, and the value of the
M

14 optimism-corrected C statistics was 0.737. Similarly, the crude and optimism-corrected C

D

15 statics were 0.733 (SE 0.013) and 0.726 for the stroke model, and 0.784 (SE 0.014) and 0.777
TE

16 for the CHD model, respectively. Calibration plots demonstrated that the predicted risk and

17 observed frequency were highly linearly correlated, and the Hosmer-Lemeshow tests
EP

18 indicated a good calibration of these models (p=0.44 – 0.90, Supplementary figure 1).

19 Alternative models that included non-HDL cholesterol (instead of LDL cholesterol) and/or
C

20 HbA1c (instead of diabetes) are shown in Supplementary tables 3-5. As with the original
AC

21 model described above, these alternative models showed satisfactory performances.

22 Table 3 summarizes the simplified scoring system based on the three multivariable

23 models. Total scores ranged from 0-62, 0-24, and 0-40 for the CVD, stroke, and CHD models,

24 respectively. The IREs, simplified scores, and corresponding vascular age according to the

25 predicted risks of each endpoint are summarized in Table 4. For example, if an individual has

26 an IRE of 6.5 for CVD or scores 24 points in the simplified CVD score, then the 10-year
 11
ACCEPTED MANUSCRIPT
1 predicted risk for CVD for this individual is about 12%, and the vascular age is around 63

2 (man) or 70 (woman). For further reference, predicted risks according to the simplified scores

3 for each endpoint are shown in Supplementary figure 2. Sex-specific plots of vascular age

4 were made against 0.5% intervals of the predicted risk of overall CVD (Supplementary figure

5 3). The risk estimates for clinical examples for two men of the same age with different risk

PT
6 profiles are presented in Supplementary table 6. In this example, a man (A) had unfavorable

RI
7 metabolic profiles characterized by elevated blood lipid and glucose levels and a smoking

8 habit, but he also had normal systolic blood pressure and regular exercise habits. Another man

SC
9 (B) had better glucose and lipid profiles but also had markedly elevated blood pressure and no

10 exercise habit. The 10-year risk for overall CVD and vascular age were similar between these

11
U
individuals, but the risk for stroke and CHD were 8% and 17% for subject A, and 12% and
AN
12 8% for subject B, respectively.
M

13 Finally, we addressed the benefit of deciding the intervention strategy with the risk

14 prediction model (i.e., the predicted risk-based decision-making) compared with either a
D

15 treat-all or treat-none strategy. Figure 1 illustrates the decision curves for the endpoints of
TE

16 interest. Our prediction models were the most predictive of all endpoints, followed by

17 point-based scoring and/or the Framingham risk profile. For CVD, the risk prediction models
EP

18 had higher net benefit than either the treat-all or treat-none strategy at the cut-off of predicted

10-year risks below around 0.3. Similarly, net benefits were higher in the strategy based on
C

19
AC

20 the risk prediction models than in either the treat-all or treat-none strategy at the cut-off of

21 predicted risk below around 0.15 for stroke and coronary heart disease.

22

23 4. Discussion

24 Identifying risk factors and foreseeing the risk for the development of chronic

25 diseases is one of the critical objectives of modern epidemiology. Using the 24-year follow-up
 12
ACCEPTED MANUSCRIPT
1 data from the Hisayama Study, we developed new risk prediction models for estimating the

2 future incidence of CVD and its subtypes—namely, stroke and CHD. Some predictors that

3 were included in the models for stroke and CHD were shared but some were not. These

4 models showed a good discrimination and calibration, and satisfactory internal validity.

5 Moreover, the present findings support the idea that decision-making based on the predicted

PT
6 risk model would be of benefit. The findings from the present study may provide insight into

RI
7 the practical usefulness of risk stratification tools for the primary prevention of cardiovascular

8 events.

SC
9 To date, several prediction models and risk charts for the Japanese population have

10 been developed for cardiovascular death [20], and incident CVD and its subtypes [7–11]. The

11
U
variables included in the CVD model of the present study were generally similar to those
AN
12 included in these previous studies, as well as those for other populations [21]. We observed
M

13 that cholesterol levels and sex had a stronger impact on coronary heart diseases than on stroke,

14 which was in line with the previous findings from the Japan Public Health Center-based
D

15 cohort study [8]. Although several models, especially those developed early, employed total
TE

16 cholesterol levels as a predictor of CVD events [3,4,20], we employed the HDL and LDL

17 cholesterol levels as potential predictors rather than total cholesterol. Recent lipid guidelines,
EP

18 such as the Japan Atherosclerosis Society Guidelines for the Prevention of Atherosclerotic

Diseases 2012 [2] and the 2013 American College of Cardiology/American Heart Association
C

19
AC

20 Guidelines [1], recommend monitoring and management of atherogenic lipid as a primary

21 target for CVD prevention. Given the current emphasis on lipid screening as a more inclusive

22 index than total cholesterol [22], we believed that it would be more reasonable to use HDL

23 and LDL cholesterol levels as predictors in a routine care setting. In cases in which a fasting

24 blood sample was not available, we also presented alternative models that included non-HDL

25 cholesterol instead of LDL cholesterol. We previously reported that non-HDL cholesterol was
 13
ACCEPTED MANUSCRIPT
1 predictive of coronary heart disease in the general population of Japanese adults [23]. In

2 addition, HbA1c was employed in additional analyses instead of the presence of diabetes,

3 which may be useful to visualize the impact of glucose-lowering treatment on prevention of

4 CVD and to set more segmentalized treatment goals.

5 We found that the developed models for stroke and CHD shared several important

PT
6 predictors; however, the impacts of the predictors differed between the two models. In

RI
7 addition, the two models comprised unique variables for predicting particular outcomes. This

8 would allow individuals to take into account the relative contributions of predictors in order to

SC
9 better understand their current risk and to set personalized goals. For example, systolic blood

10 pressure accounted for a little more of the total score in the stroke model than in the CHD

11
U
model. In addition, regular exercise was strongly associated with a reduced risk of stroke but
AN
12 not CHD, while lipid levels were associated with CHD but not stroke risk. Thus the risk of
M

13 stroke and CHD would differ according to the patient profiles, and the prediction model

14 would allow us to identify the factor(s) contributing to an elevation or a reduction in the risk
D

15 status for either disease in each patient. This specificity would support a decision-making
TE

16 process that balances the risks and preferences at the individual patient level [24].

17 The prediction models might also be used to promote risk stratification in public
EP

18 health strategies for the primary prevention of CVD events. The results showed that

decision-making based on the predicted risk model had higher net benefits than did an
C

19
AC

20 intervention strategy for CVD for either none or all of the individuals. For example, the

21 decision curve for CVD showed that when 8% was set as a cut-off for the predicted 10-year

22 risk, the net benefit of applying the model was 0.039 and that for the treat-all strategy was

23 0.008. The difference (i.e., 0.031) implied that, for every 1,000 individuals in which the model

24 was applied, an additional 31 true-positive cases could be detected without increasing the

25 false-positive rate. Thus the public health sectors would benefit from screening individuals at
 14
ACCEPTED MANUSCRIPT
1 high risk of CVD or its subtypes by setting an appropriate cut-off for the 10-year risk of the

2 disease, and providing individual health guidance and education, and/or initiating treatment

3 such as cholesterol-lowering therapy.

4 The strengths of this study included the accurate diagnosis of subtypes, very high

5 follow-up rate, long duration of follow-up, and detailed examinations of the model

PT
6 performances. In addition, there has been, to our knowledge, no study examining the practical

RI
7 usefulness of a risk model for cardiovascular diseases using decision curve analysis in an

8 Asian population. Several limitations should also be noted. First, the external validity of the

SC
9 models was not confirmed in this study. Moreover, the number of CVD events—and

10 especially CHD events—was insufficient for a more detailed examination of the model

11
U
performances under various conditions. Thus, the generalizability and the accuracy of the
AN
12 established model among populations from other countries/regions may be limited. However,
M

13 the risk factors included in the final models were generally similar to those included in

14 previously published CVD risk prediction models worldwide [21]. Replication studies will be
D

15 needed to confirm the external validity and to evaluate the performance and clinical
TE

16 usefulness in other populations. Secondly, our observations were based on a single

17 measurement of the risk factors at baseline. It is possible that the level of risk factors was
EP

18 changed by modifications to lifestyle or medication during follow-up. This limitation could

have led to misclassification of the levels of risk factors included in the risk model, which in
C

19
AC

20 turn could have reduced the predictive ability of the developed models, biasing the results

21 toward the null hypothesis. On the other hand, most prediction models tend to use information

22 gathered at a single time-point, because this condition is more applicable to a primary care

23 setting, rather than data gained from repeated measurements.

24 In conclusion, we developed new risk prediction models for CVD and its subtypes,

25 which showed good discrimination and calibration, and satisfactory internal validity. The
 15
ACCEPTED MANUSCRIPT
1 developed prediction models may help individuals to understand their current risk and to set

2 personalized goals, in addition to promoting risk stratification in public health strategies for

3 the primary prevention of CVD events.

5 Conflict of interest

PT
6 The present project was conducted with the cooperation of DeNA Co., Ltd. Funders had no

RI
7 role in the study design, conduct of the study, data collection, or preparation of the report.

SC
9 Financial support

10 This study was supported in part by Grants-in-Aid for Scientific Research (A) (JP16H02644

11
U
and JP16H02692), (B) (JP16H05850, JP16H05557, JP17H04126, and JP18H02737) and (C)
AN
12 (JP16K09244, JP17K09114, JP17K09113, JP17K01853, JP18K07565, and JP18K09412) and
M

13 Grants-in-Aid for Early-Career Scientists (JP18K17925 and JP18K17382) from the Ministry

14 of Education, Culture, Sports, Science and Technology of Japan; by Health and Labour
D

15 Sciences Research Grants of the Ministry of Health, Labour and Welfare of Japan
TE

16 (H29-Junkankitou-Ippan-003, and H30-Shokuhin-[Sitei]-005); and by grants from the Japan

17 Agency for Medical Research and Development (JP18dk0207025, JP18ek0210082,

EP

18 JP18gm0610007, JP18ek0210083, JP18km0405202, JP18ek0210080, and JP18fk0108075).

In addition, this study was sponsored by DeNA Co., Ltd. (Tokyo, Japan). The funders had
C

19
AC

20 no role in the design and conduct of the study; collection, analysis, and interpretation of the

21 data; preparation or review of the manuscript; and decision to submit the manuscript for

22 publication.

23

24 Author contributions

25 TH contributed to study design, statistical analyses, interpretation, and drafting manuscript;

 16
ACCEPTED MANUSCRIPT
1 DY, JH, YH, YI, MS, and SS contributed to data collection, interpretation, and revision; TK

2 contributed to critical revision; TN contributed to study design, statistical analysis,

3 interpretation, and critical revision. All authors have read the final version of the manuscript

4 and approved before submission for publication.

PT
6 Acknowledgements

RI
7 The authors thank the residents of the town of Hisayama for their participation in the survey

8 and the staff of the Division of Health and Welfare of Hisayama for their cooperation with this

SC
9 study. We would like to gratefully and sincerely thank Professor Yoshinao Oda, Professor

10 Toru Iwaki, and the colleagues from Department of Anatomic Pathology and Department of

11
U
Neuropathology, Graduate School of Medical Sciences, Kyushu University, who provided
AN
12 insight and expertise of the autopsy findings that greatly assisted the research.
M

13

14 References
D

15 [1] N.J. Stone, J.G. Robinson, A.H. Lichtenstein, C.N. Bairey Merz, R.H. Eckel, et al.,
TE

16 2013 ACC/AHA guideline on the treatment of blood cholesterol to reduce

17 atherosclerotic cardiovascular risk in adults: a report of the American College of

EP

18 Cardiology/American Heart Association Task Force on Practice Guidelines,

Circulation. 129[25 suppl 2] (2014) S1–S45.

C

19
AC

20 [2] T. Teramoto, J. Sasaki, S. Ishibashi, S. Birou, H. Daida, et al., Executive Summary of

21 the Japan Atherosclerosis Society (JAS) Guidelines for the Diagnosis and Prevention

22 of Atherosclerotic Cardiovascular Diseases in Japan-2012 Version, J. Atheroscler.

23 Thromb. 21 (2014) 291–295.

24 [3] R.B. D’Agostino, R.S. Vasan, M.J. Pencina, P.A. Wolf, M. Cobain, et al., General

25 cardiovascular risk profile for use in primary care: the Framingham heart study,
 17
ACCEPTED MANUSCRIPT
1 Circulation. 117 (2008) 743–753.

2 [4] R.M. Conroy, K. Pyörälä, A.P. Fitzgerald, S. Sans, A. Menotti et al., SCORE project

3 group. Estimation of ten-year risk of fatal cardiovascular disease in Europe: the

4 SCORE project. Eur. Heart. J. 24 (2003) 987–1003.

5 [5] E. Palomeras Soler, V. Casado Ruiz, Epidemiology and risk factors of cerebral

PT
6 ischemia and ischemic heart diseases: similarities and differences, Curr. Cardiol. Rev.

RI
7 6 (2010) 138–149.

8 [6] K.W. Giang, L. Björck, M. Novak, G. Lappas, L. Wilhelmsen, et al., Stroke and

SC
9 coronary heart disease: predictive power of standard risk factors into old age -

10 Long-term cumulative risk study among men in Gothenburg, Sweden, Eur. Heart J. 34

11 (2013) 1068–1074.
U
AN
12 [7] H. Yatsuya, H. Iso, K. Yamagishi, Y. Kokubo, I. Saito, et al., Development of a
M

13 point-based prediction model for the incidence of total stroke: Japan Public Health

14 Center Study, Stroke 44 (2013) 1295–1302.

D

15 [8] H. Yatsuya, H. Iso, Y. Li, K. Yamagishi, Y. Kokubo, et al.. Development of a risk

TE

16 equation for the incidence of coronary artery disease and ischemic stroke for

17 middle-aged Japanese - Japan Public Health Center-Based Prospective Study, Circ J.

EP

18 80 (2016) 1386–95.

[9] K. Nishimura, T. Okamura, M. Watanabe, M. Nakai, M. Takegami, Predicting

C

19
AC

20 coronary heart disease using risk factor categories for a Japanese urban population, and

21 comparison with the Framingham risk score: the Suita Study, J. Atheroscler. Thromb.

22 (2014) 784–798.

23 [10] M. Matsumoto, S. Ishikawa, K. Kayaba, T. Gotoh, N. Nago, et al., Jichi Medical

24 School (JMS) Cohort Study Group, Risk charts illustrating the 10-year risk of

25 myocardial infarction among residents of Japanese rural communities: the JMS Cohort
 18
ACCEPTED MANUSCRIPT
1 Study, J. Epidemiol. 19 (2009) 94–100.

2 [11] S. Ishikawa, M. Matsumoto, K. Kayaba, T. Gotoh, N. Nago, et al., Jichi Medical

3 School (JMS) Cohort Study Group, Risk charts illustrating the 10-year risk of stroke

4 among residents of Japanese rural communities: the JMS Cohort Study, J. Epidemiol.

5 19 (2009) 101–106.

PT
6 [12] K.L. Chien, T.C. Su, H.C. Hsu, W.T. Chang, P.C. Chen, et al., Constructing the

RI
7 prediction model for the risk of stroke in a Chinese population: report from a cohort

8 study in Taiwan. Stroke 41 (2010) 1858–64.

SC
9 [13] H. Arima, K. Yonemoto, Y. Doi, T. Ninomiya, J. Hata, et al., Development and

10 validation of a cardiovascular risk prediction model for Japanese: the Hisayama Study,

11 Hypertens. Res. 32 (2009) 1119–22.

U
AN
12 [14] T. Usui, M. Nagata, J. Hata, N. Mukai, Y. Hirakawa, et al., Serum non-high-density
M

13 lipoprotein cholesterol and risk of cardiovascular disease in community dwellers with

14 chronic kidney disease: the Hisayama Study, J. Atheroscler. Thromb. (2017) 706–715.
D

15 [15] B. Richter, L. Koller, P.J. Hohensinner, G. Zorn, M. Brekalo, et al., A multi-biomarker

TE

16 risk score improves prediction of long-term mortality in patients with advanced heart

17 failure, Int. J. Cardiol. 168 (2013) 1251–1257.

EP

18 [16] F.E. Harrell Jr, K.L. Lee, D.B. Mark, Multivariable Prognostic Models: Issues in

developing models, evaluating assumptions and adequacy, and measuring and reducing
C

19
AC

20 errors, Stat. Med. 28 (1996) 361–387.

21 [17] L.M. Sullivan, J.M. Massaro, R.B. D’Agostino, Presentation of multivariate data for

22 clinical use: the Framingham Study risk score functions, Stat. Med. 23 (2004) 1631–

23 1660.

24 [18] A.J. Vickers, E.B. Elkin, Decision curve analysis: a novel method for evaluating

25 prediction models, Med. Decis. Mak. 26 (2008) 565–574.

 19
ACCEPTED MANUSCRIPT
1 [19] A.J. Vickers, B. Van Calster, E.W. Steyerberg, Net benefit approaches to the

2 evaluation of prediction models, molecular markers, and diagnostic tests, BMJ 352

3 (2016) i6

4 [20] NIPPON DATA80 Research Group, Risk assessment chart for death from

5 cardiovascular disease based on a 19-year follow-up study of a Japanese representative

PT
6 population., Circ. J. 70 (2006) 1249–1255.

RI
7 [21] J.A.A.G. Damen, L. Hooft, E. Schuit, T.P.A. Debray, G.S. Collins, et al., Prediction

8 models for cardiovascular disease risk in the general population: systematic review,

SC
9 BMJ 353 (2016) i2416.

10 [22] H.E. Bays, A lipidologist perspective of global lipid guidelines and recommendations,

11
U
part 1: lipid treatment targets and risk assessment, J. Clin. Lipidol. 10 (2016) 228–239.
AN
12 [23] T. Imamura, Y. Doi, T. Ninomiya, J. Hata, M. Nagata, et al., Non-high-density
M

13 lipoprotein cholesterol and the development of coronary heart disease and stroke

14 subtypes in a general Japanese population: the Hisayama Study, Atherosclerosis. 233

D

15 (2014) 343–348.
TE

16 [24] T. Yu, D. Vollenweider, R. Varadhan, T. Li, C. Boyd, M.A. Puhan, Support of

17 personalized medicine through risk-stratified treatment recommendations - an

EP

18 environmental scan of clinical practice guidelines, BMC Med. 11 (2013) 7.

C

19
AC

20 Keywords

21 risk prediction, prospective study, cardiovascular disease, stroke, coronary heart disease

22

23 Abbreviations

24 BMI, body mass index; CVD, cardiovascular diseases; CHD, coronary heart diseases; HDL

25 cholesterol, high-density lipoprotein cholesterol; IRE, individual risk estimates; LDL

 20
ACCEPTED MANUSCRIPT
1 cholesterol, low-density lipoprotein cholesterol; SD, standard deviation; SE, standard error

PT
RI
U SC
AN
M
D
TE
C EP
AC
 21
ACCEPTED MANUSCRIPT
1 Figure 1. Decision curves for cardiovascular disease, stroke, and coronary heart disease.

3 Gray thin line: assuming treatment of all subjects; horizontal line (on Y=0): assuming

4 treatment of no subjects; black thick line: multivariable model-based prediction; black thin

5 line: point-based scoring estimation; dotted line: prediction based on the Framingham general

PT
6 risk profile. Threshold probability refers to the cut-offs of predicted 10-year risks for

RI
7 development of the respective diseases, for use in the decision to initiate

8 intervention/treatment.

SC
9

10

11
U
AN
M
D
TE
C EP
AC
 22
ACCEPTED MANUSCRIPT
1 Table 1. Baseline characteristics of study subjects
Mean (SD) or prevalence

Age, years 58.3 (11.1)

Sex, % men 42.1

Body mass index, kg/m2 22.9 (3.1)

Obesity, % 23.5

PT
Systolic blood pressure, mmHg 132.4 (20.3)
Diastolic blood pressure, mmHg 77.5 (11.2)

Diabetes, % 11.8

RI
HbA1c, % level 5.9 (0.8)

SC
Total cholesterol, mg/dL 207.4 (42.0)
HDL cholesterol, mg/dL 50.6 (11.7)

Non-HDL cholesterol, mg/dL 156.8 (41.0)

U
LDL cholesterol, mg/dL 134.0 (39.5)
AN
Triglycerides, mg/dL 114.1(62.9)
Current drinker, % 30.3

Current smoker, % 25.0

M

Regular exercise, % 10.2

2 SI conversion factors: to convert mg/dL values to mmol/L, multiply total cholesterol, HDL cholesterol,
D

3 non-HDL cholesterol, and LDL cholesterol values by 0.0259, and triglyceride values by 0.0113.
TE

4 SD, standard deviation; HbA1c, hemoglobin A1c; HDL cholesterol, high-density lipoprotein cholesterol;

5 LDL cholesterol, low-density lipoprotein cholesterol.

EP

6
C

7
AC
 ACCEPTED MANUSCRIPT

23

Table 2. Multivariable models for the development of cardiovascular diseases, stroke, and coronary heart diseases
Cardiovascular disease Stroke Coronary heart disease
HR (95%CI) β p value HR (95%CI) β p value HR (95%CI) β p value

PT
Number of events 489 324 216
Age (per year) 1.07 (1.06-1.08) 0.067 <0.001 1.07 (1.05-1.08) 0.064 <0.001 1.08 (1.06-1.10) 0.077 <0.001

RI
Men (vs female) 1.75 (1.41-2.17) 0.559 <0.001 1.26 (0.97-1.62) 0.228 0.081 2.79 (2.02-3.85) 1.025 <0.001
Systolic blood pressure (per 1
1.01 (1.01-1.02) 0.013 <0.001 1.01 (1.01-1.02) 0.013 <0.001 1.01 (1.01-1.02) 0.011 0.001

SC
mmHg)
Diabetes (vs nondiabetic) 1.51 (1.20-1.90) 0.412 <0.001 1.56 (1.18-2.07) 0.444 0.002 1.63 (1.17-2.28) 0.489 0.004

U
HDL cholesterol (per 1 mg/dL) 0.991 (0.983-0.999) -0.009 0.03 0.986 (0.975-0.998) -0.014 0.03
LDL cholesterol (per 1 mg/dL) 1.003 (1.000-1.005) 0.003 0.03 1.005 (1.001-1.008) 0.005 0.005

AN
Current smoker (vs. non-smoker) 1.40 (1.12-1.76) 0.338 0.003 1.45 (1.09-1.93) 0.373 0.01 1.38 (1.00-1.90) 0.321 0.050
Regular exercise (vs. no or

M
0.66 (0.48-0.90) -0.420 0.008 0.61 (0.41-0.91) -0.49 0.02
irregular)
Baseline survival function

D
at 10-year follow-up 0.936 0.955 0.978

TE
at 20-year follow-up 0.841 0.889 0.939
Baseline risk estimate 5.866 5.693 6.486
EP
C statistics 0.743 0.733 0.784
Optimism-corrected C statistics 0.737 0.726 0.777
C

Cox’s proportional hazard regression models with a backward selection method (p<0.1) were fitted to 200 bootstrapping samples, and variables selected ≥ 80 times
were included in the final models.
AC

SI conversion factors: to convert mg/dL values to mmol/L, multiply HDL cholesterol and LDL cholesterol values by 0.0259.

HR, hazard ratio; 95% CI, 95% confidence interval; HDL cholesterol, high-density lipoprotein cholesterol; LDL cholesterol, low-density lipoprotein cholesterol.
 24
ACCEPTED MANUSCRIPT
Table 3. Simplified point-based scoring system for cardiovascular disease
Overall
Variable Levels Stroke CHD
CVD
40-49 years 0 0 0
50-59 years 7 3 5
Age 60-69 years 15 7 10
70-79 years 22 10 15
80-84 years 28 12 19

Women 0 0 0

PT
Sex
Men 6 1 7

<120 mmHg 0 0 0

RI
120-129 mmHg 2 1 1
Systolic blood pressure
130-139 mmHg 4 2 2
140-159 mmHg 6 3 3
160 mmHg - 9 4 4

SC
No 0 0 0
Diabetes
Yes 5 2 3

U
≥60 mg/dL 0 0
HDL cholesterol 40-59 mg/dL 2 1
AN
<40 mg/dL 3 3

<120 mg/dL 0 0
LDL cholesterol 120-139 mg/dL 1 1
M

≥140 mg/dL 2 2

No 0 0 0
Current smoker
D

Yes 4 2 2

No 5 3
TE

Regular exercise
Yes 0 0
EP

Total score range 0-62 0-24 0-40

C

SI conversion factors: to convert mg/dL values to mmol/L, multiply HDL cholesterol and LDL cholesterol
AC

values by 0.0259.

CVD, cardiovascular diseases; CHD, coronary heart diseases; HDL cholesterol, high-density lipoprotein

cholesterol; LDL cholesterol, low-density lipoprotein cholesterol.

 ACCEPTED MANUSCRIPT

25

Table 4. Summary of risk calculation/scoring and vascular age according to predicted 10-year risks

Cardiovascular disease Stroke Coronary heart disease

PT
Simplified Vascular age Simplified Predicted Simplified
Predicted risk IRE Predicted risk IRE IRE
score (Men, women) score risk score
4% 5.38 8 49, 57 4% 5.57 5 4% 7.09 13

RI
6% 5.80 14 54, 61 6% 5.99 7 6% 7.51 16
8% 6.10 18 58, 65 8% 6.29 9 8% 7.81 19

SC
10% 6.33 22 60, 67 10% 6.52 11 10% 8.04 21
12% 6.53 24 63, 70 12% 6.71 12 12% 8.23 22
14% 6.69 27 65, 71 14% 6.88 13 14% 8.40 24

U
16% 6.84 29 67, 73 16% 7.02 14 16% 8.54 25

AN
18% 6.97 31 68, 75 18% 7.15 15 18% 8.67 26
20% 7.08 32 70, 76 20% 7.27 16 20% 8.79 27
25% 7.34 36 73, 79 25% 7.53 18 25% 9.05 29

M
30% 7.55 39 75, 81 30% 7.74 19 30% 9.26 31
35% 7.74 41 78, 83 35% 7.93 20 35% 9.45 33

D
40% 7.91 44 80, 85 40% 8.10 21 40% 9.62 34
45% 8.07 46 82, 87 45% 8.26 22 45% 9.78 35

TE
50% 8.22 48 84, 89 50% 8.40 23 50% 9.92 36
IRE (individual risk estimate) values were calculated by summing the products of each predictor variable value and the corresponding regression beta coefficient.
EP
Values for the IRE and simplified score were the nearest value corresponding to the predicted risk for each outcome in the same line, and vascular age corresponds to
that for overall cardiovascular disease.
C
AC
 0.080 Prediction model
Point-based scoring
Net benefit
0.060 ACCEPTED MANUSCRIPT Framingham

0.040

0.020

0.000

-0.020

PT
0 0.1 0.2 0.3 0.4 0.5
Threshold probability

RI
[B] Stroke

SC
0.060
0.050
Prediction model
Net benefit

U
0.040 Point-based scoring
AN
Framingham
0.030
0.020
M

0.010
D

0.000
TE

-0.010
0 0.1 0.2 0.3 0.4 0.5
EP

Threshold probability
C

0.030 [C] Coronary heart disease

AC

0.025
Prediction model
Net benefit

0.020 Point-based scoring

Framingham
0.015
0.010
0.005
0.000
-0.005
0 0.1 0.2 0.3 0.4 0.5
 ACCEPTED MANUSCRIPT
Highlights

3 to 5 bullet points (maximum 85 characters, including spaces, per bullet point).

We developed risk prediction models for cardiovascular disease and its

PT
subtypes.

Our risk prediction models exhibited good performance and internal validity.

RI
Predicted risk-based decision-making can be beneficial in primary

SC
prevention.

U
AN
M
D
TE
C EP
AC