A diagnostic prediction model for hypertension in Han and Yugur population from the China National Health Survey (CNHS) : Chinese Medical Journal

Secondary Logo

Journal Logo

Original Articles

A diagnostic prediction model for hypertension in Han and Yugur population from the China National Health Survey (CNHS)

Yu, Chengdong1; Ren, Xiaolan2; Cui, Ze3; Pan, Li1; Zhao, Hongjun2,4; Sun, Jixin3; Wang, Ye1; Chang, Lijun2; Cao, Yajing3; He, Huijing1; Xi, Jin’en2; Zhang, Ling5; Shan, Guangliang1

Editor(s): Ni, Jing

Author Information
Chinese Medical Journal 136(9):p 1057-1066, May 5, 2023. | DOI: 10.1097/CM9.0000000000001989



Hypertension is the leading preventable risk factor for morbidity and mortality of cardiovascular disease (CVD).[1] Almost 1.38 billion people worldwide (31.1% of the global adult population) were hypertensive in 2010.[2] The guidelines stressed the importance of hypertension prevention and management to reduce the burden of CVD.[3] In China, recent evidence demonstrated that the prevalence of hypertension was 23.2% (about 244.5 million) among Chinese adults and another 41.3% (about 435.3 million) had prehypertension.[4] However, compared with the developed countries, the awareness in China was much lower (82.1% vs. 46.9%),[4,5] indicating hundreds of millions of Chinese are not yet aware of this dangerous and invisible killer.

Yugur is a minority of Gansu Province located in northwest China. Compared with Han people, the dietary pattern of Yugur people tends to be traditionally rich in beef, mutton, and alcohol, which might increase the risk of hypertension. Our previous study had confirmed that the prevalence of hypertension in Yugur was higher than in Han.[6] Therefore, identifying individuals who are at high risk of hypertension and establishing a risk prediction model will help to improve the efficiency of primary prevention strategies.

Prediction models are developed to estimate the probability or risk that a specific disease or condition is present (diagnostic models) or that a specific event will occur in the future (prognostic models).[7] So far, many prognostic models for hypertension have been developed or validated in the Caucasian and Asian general populations.[8–11] However, the Framingham prediction models have been validated as not so suitable for Chinese as they were based on Caucasian.[12] For Chinese people, a famous hypertension prognostic model was established by the Chin-Shan Community Cardiovascular Cohort Study in Taiwan;[10] however, lifestyle risk factors were not included in this model. Based on a large cross-sectional study in central China, Ren et al[13] established a diagnostic model for hypertension using the logistic regression model, but the results were not visualized, making it not convenient to calculate the risk probability. Du et al[14] established a diagnostic model for hypertension prediction in northern China using a cross-sectional data, but the model was not fully calibrated or validated.

Till now, no study has attempted to develop a prediction model for hypertension for Han and Yugur people in the Gansu Province of China. Therefore, this study aimed to develop and validate a well visualized diagnostic model that incorporated objective demographic, lifestyle factors, and anthropometric parameters of Han and Yugur people, and further assess the clinical utility and impact of the model.[15,16]


Ethics approval

The research was approved by the Institutional Review Board of the Institute of Basic Medical Sciences, Chinese Academy of Medical Sciences (No. 029-2013), and conducted in accordance with the ethical principles of the Declaration of Helsinki. Informed consent was obtained from all individual participants included in the study.

Study design and participants

The data of this observational study were based on the China National Health Survey in Gansu and Hebei provinces,[17] which was conducted in 2016 to 2017 by the Institute of Basic Medical Sciences, Chinese Academy of Medical Sciences, School of Basic Medicine, Peking Union Medical College. Details of the sampling procedures have been described in a study by Yu et al.[6] In brief, a total of 9699 eligible Han and Yugur participants aged 20 to 80 years, who had been living in Gansu/Hebei Province for not <1 year were recruited. Participants who had no information about the history of hypertension and blood pressure measurements were excluded.

Measures and procedures

Blood pressure was measured three times using the Omron digital blood pressure measuring device (HEM-907, Japan) on the right arm with the participant seated, and the average of the three measurements was used as mean blood pressure for analysis. Participants were classified as hypertensive if mean systolic blood pressure ≥140 mmHg and/or mean diastolic blood pressure ≥90 mmHg and/or in case of physician-diagnosed hypertension and/or self-report of current antihypertensive medication use.[18]

Family history of hypertension (FHH) was assessed based on the self-report of participants about whether any of his/ her family members (living or deceased) had hypertension diagnosed by a physician. Provided options were “yes”, “no”, and “don’t know”. If a participant selected “yes”, he/she was asked to ascertain the presence of hypertension in grandfather, grandmother, mother, father, siblings, and offspring. In addition, FHH was grouped into the following generations: first generation-siblings, second generation-parents, and third generation-grandparents. According to the presence of hypertension in the three generations, FHH was categorized as FH0 (negative), FH1 (only one generation had hypertension), and FH2+ (two or three generations had hypertension). Those who were unaware (“don’t know”) of their family history were regarded as having no FHH. Information regarding offspring was not considered. The definitions of other variables are detailed in the Supplementary Materials, https://links.lww.com/CM9/A936.

Statistical analysis

Continuous variables were described as mean (standard deviation), and categorical variables were expressed as number (percentage). The differences between the training set and internal/external validation set were compared using independent samples t tests for continuous variables and chi-squared tests for categorical variables. Details of the sample size calculation are provided in Supplementary Materials, https://links.lww.com/CM9/A936.

Division of datasets

To ensure independence, participants of the training set and internal validation set were chosen from independent survey sites nonrandomly. Finally, the internal validation set consisted of all the 1402 Han participants in Gaotai county and all the 522 Yugur participants in Minghua township. The rest of the participants in each population of Gansu Province (2845 Han and 1108 Yugur) comprised the training set. All the Han participants in Hebei Province comprised the external validation set. The flowchart is shown in Supplementary Figure 1, https://links.lww.com/CM9/A936.

Feature selection

As the least absolute shrinkage and selection operator (Lasso) regression is a statistical technique that selects the most important predictive factors, the method is suitable for regression of high-dimensional data. We performed Lasso logistic regression in the training set using the “glmnet” package.[19] The dependent variable in the model was hypertension, whereas the independent variables included continuous predictors (age and body mass index [BMI]) and categorical predictors (ethnicity, gender, current residence, marital status, educational level, smoking status, alcohol consumption, physical activity, annual income, heart rate, and FHH). Dummy variables were created for categorical variables.

Development of an individualized prediction model

Using the features selected by the Lasso method, a multivariable binary logistic regression model was applied to develop a diagnostic model for hypertension in the training set. To provide the clinician with a quantitative tool to predict the individual probability of hypertension, a nomogram was constructed based on multivariable logistic analysis using the “rms” package.[20] Subsequently, the calculated probabilities of individuals in the training set were classified into three risk levels (low risk, intermediate risk, and high risk) based on the two cutoff points determined by the “findcut” function.[21] Furthermore, a simple and user-friendly website was developed to calculate the individualized risk of hypertension based on the “Dynnom” package.[22]

Performance of the model in the training set

Harrell's C-index was calculated among predicted and actual outcomes to evaluate the discriminatory abilities. The model was subjected to bootstrapping validation (1000 bootstrap resamples) to estimate a relative corrected C-index. To estimate calibration, the Hosmer-Lemeshow goodness of fit test was performed.[23] To identify the ranges of the probability of miscalibration in the model, the GiViTI calibration belts (see more details in Supplementary Materials, https://links.lww.com/CM9/A936) were plotted using the “ givitiR” package, together with its statistical confidence (80% and 95%).[24]

Validating the performance of the model

Discrimination and calibration were also taken into account to estimate the validity of the model in internal and external validation sets. The logistic regression model formed in the training set was applied to all participants of validation sets, and the probability for each individual was calculated. Logistic regressions in validation sets were performed using the probabilities as the independent or dependent factor. Finally, the C-indexes and GiViTI calibration belts were derived based on the regression analyses.

In addition, the internal validation set was split into two sets (Han and Yugur) by ethnicity to validate the performance of the model in Han or Yugur people separately.

Clinical use

To evaluate the clinical usefulness of the model, decision curve analysis was conducted by quantifying the net benefits (NBs) at different threshold probabilities in validation datasets using the “rmda” package. The NB were calculated using the following formula[16]:Netbenefit=(true-positivecounts)/n(false-positivecounts)/n)[pt/(1pt)]

where the true- and false-positive counts are the number of participants with true- and false-positive results, n is the total number of participants, and pt is the threshold probability where the expected benefit of treatment is equal to the expected benefit of avoiding treatment. Considering the specific difficulty of NB is that it is in units of B, and the maximum NB is prevalence (P), it motivates the metric standardized NB, sNB = NB/P, also known as the relative utility.[15] Because the sNB always has a maximum value of 1.0, providing a sense of large and small on a percent scale, sNB is slightly easier to interpret than NB.

Comparison with a published diagnostic model

Du et al[14] developed a diagnostic prediction model for hypertension based on a cross-sectional study in Northern China in 2018. To determine whether our model had a better performance, we calculated the probabilities in training, internal, and external validation sets using Du's and our model separately, and then evaluated the differences of area under receiver operating characteristic (ROC) curve (AUC) by performing the DeLong's test.

Two-sided P value <0.05 was considered as statistical significance. The statistical analysis was conducted using R software (version 4.1.1; http://www.r-project.org). Details of relevant packages are shown in Supplementary Materials, https://links.lww.com/CM9/A936.


Demographic characteristics

A total of 9699 participants were finally included in our study. Characteristics of participants in training and validation sets are shown in Table 1. The prevalence of hypertension had no significant difference between training and internal validation sets (29.4% vs. 30.6%, P= 0.349), while it was higher in the external validation set than in training set (39.9% vs. 29.4%, P< 0.001). Aside from ethnicity, BMI between training and internal validation sets, and smoking status between training and external validation sets, the obvious differences were found between training and validation sets in terms of socio-demographic, lifestyle, and physical examination characteristics (all P< 0.05). Considering the independence of training and validation sets, these differences justified the use as training or validation sets.

Table 1 - Demographic characteristics of 9699 participants in training and validation sets.
Training set Internal validation set External validation set
Parameters (N = 3953) (N = 1924) (N = 3822) P Internal P External
Ethnicity 0.470 -
 Han 2845 (72.0) 1402 (72.9) 3822 (100.0)
 Yugur 1108 (28.0) 522 (27.1) 0 (0.0)
Age (years) 48.9 (12.5) 51.4 (11.2) 47.9 (13.8) <0.001 0.001
 20–29 337 (8.5) 76 (4.0) 481 (12.6)
 30–39 563 (14.2) 202 (10.5) 680 (17.8)
 40–49 1105 (28.0) 552 (28.7) 853 (22.3)
 50–59 1098 (27.8) 569 (29.6) 913 (23.9)
 60–69 660 (16.7) 447 (23.2) 673 (17.6)
 70–80 190 (4.8) 78 (4.1) 222 (5.8)
Gender <0.001 0.041
 Male 1580 (40.0) 953 (49.5) 1615 (42.3)
 Female 2373 (60.0) 971 (50.5) 2207 (57.7)
BMI (kg/m2) 24.2 (3.4) 24.1 (3.4) 25.3 (5.3) 0.376 <0.001
 <24.0 1962 (50.3) 963 (51.0) 1706 (44.8)
 24.0–27.9 1454 (37.3) 704 (37.2) 1051 (27.6)
 ≥28.0 485 (12.4) 223 (11.8) 1052 (27.6)
Annual income (CNY) <0.001 <0.001
 <20,000 968 (24.6) 1147 (59.8) 1683 (44.3)
 ≥20,000 2972 (75.4) 772 (40.2) 2115 (55.7)
Current residence <0.001 <0.001
 Urban 2936 (74.3) 500 (26.0) 1518 (39.7)
 Rural 1017 (25.7) 1424 (74.0) 2302 (60.3)
Marital status <0.001 <0.001
 Single 218 (5.5) 45 (2.3) 134 (3.5)
 Married 3492 (88.3) 1775 (92.3) 3509 (91.8)
 Divorced or widowed 243 (6.1) 104 (5.4) 179 (4.7)
Educational level <0.001 <0.001
 Low 1167 (29.6) 1007 (52.6) 1188 (31.1)
 Medium 1548 (39.2) 689 (36.0) 1994 (52.2)
 High 1231 (31.2) 220 (11.5) 637 (16.7)
Smoking status <0.001 0.465
 None 2736 (69.2) 1197 (62.2) 2616 (68.4)
 Former or current 1217 (30.8) 727 (37.8) 1206 (31.6)
Alcohol consumption <0.001 <0.001
 None 1629 (41.2) 687 (35.7) 1929 (50.5)
 Former or current 2324 (58.8) 1235 (64.3) 1893 (49.5)
Physical activity <0.001 <0.001
 Low 604 (15.3) 171 (8.9) 874 (22.9)
 Moderate 2869 (72.6) 1242 (64.7) 2454 (64.2)
 High 479 (12.1) 508 (26.4) 492 (12.9)
FHH <0.001 0.012
 FH0 1500 (37.9) 949 (49.3) 1334 (34.9)
 FH1 1582 (40.0) 671 (34.9) 1640 (42.9)
 FH2+ 871 (22.0) 304 (15.8) 847 (22.2)
Heart rate (bpm) <0.001 0.001
 <80 2294 (58.1) 1001 (52.1) 2077 (54.4)
 ≥80 1655 (41.9) 921 (47.9) 1740 (45.6)
Hypertension 0.349 <0.001
 No 2792 (70.6) 1336 (69.4) 2296 (60.1)
 Yes 1161 (29.4) 588 (30.6) 1526 (39.9)
The data are presented as n (%) or mean (standard deviation). Differences in the characteristics between the two groups were examined using independent samples t-test for continuous variables and Chi-squared test for categorical variables. Counts may not add up to the total due to missing values, and percentages may not sum to 100% due to rounding. PInternal means the significance of the difference between the training set and internal validation set, and PExternal means the significance of the difference between the training set and external validation set. BMI: Body mass index; bpm: Beat per minute; CNY: Chinese Yuan; FHH: Family history of hypertension; FH0: Negative family history; FH1: Only one generation had hypertension; FH2 +: Two or three generations had hypertension.

Feature selection

In the training set, the 17 variables, with dummy variables, were included in the Lasso logistic regression model, and finally, were reduced to eight potential predictors with nonzero coefficients (age, female, BMI, rural, annual income ≥20,000 Chinese Yuan (CNY), heart rate ≥80 beats per min, FH1, and FH2+) with a cross-validated error within one standard error of the minimum (the optimal lambda was 0.016). The coefficient profile plot and 10-fold cross-validated error plot of the Lasso regression model are shown in Figure 1.

Figure 1:
Factor selection using the Lasso binary logistic regression model. (A) Lasso coefficients of 17 candidate variables (including dummy variables) in the training set. (B) Identification of the optimal penalization coefficient (λ) in the Lasso model was achieved by 10-fold cross-validation and the minimum criterion in the training set. The left vertical line represents the minimum error, and the right vertical line represents the cross-validated error within one standard error of the minimum. Lasso: Least absolute shrinkage and selection operator.

Development of the diagnostic model

Subsequently, multivariable logistic regression analysis identified these factors as independent predictors [Table 2]. The model that incorporated the above independent predictors was developed and presented as a nomogram [Figure 2]. In addition, the risk classification system for hypertension was also developed according to the prediction probabilities of each individual estimated by the model. Based on the two cutoff points determined by “findcut” function, participants were classified into low risk (probability ≤10.3%), intermediate risk (10.3% <probability ≤55.1%), and high risk (probability >55.1%) subgroups [Figure 2]. The prevalences of hypertension in each dataset and each subgroup are shown in Supplementary Figure 2, https://links.lww.com/CM9/A936. Furthermore, to facilitate the use of the nomogram for clinicians and participants, we developed an operation interface on a web page (https://chris-yu.shinyapps.io/hypertension_risk_prediction/) to calculate the exact probabilities of hypertension.

Table 2 - Odds ratio (95% CI) and ß-coefficient estimated by logistic regression analysis using the training set.
Parameters ß-coefficient OR (95% CI) P value
Intercept −9.58 <0.001
Age (years) 0.08 1.08 (1.07, 1.09) <0.001
Gender (ref = Male) 1.00
 Female −0.56 0.57 (0.48, 0.67) <0.001
BMI (kg/m2) 0.18 1.20 (1.17, 1.23) <0.001
Annual income (CNY) (ref = “<20,000”) 1.00
 ≥20,000 −0.27 0.77 (0.63, 0.93) 0.008
Current residence (ref = Urban) 1.00
 Rural 0.22 1.25 (1.03, 1.52) 0.024
Heart rate (beat per min) (ref = “<80”) 1.00
 ≥80 0.19 1.21 (1.03, 1.42) 0.022
FHH (ref = FH0) 1.00
 FH1 0.72 2.06 (1.70, 2.50) <0.001
 FH2+ 1.57 4.83 (3.89, 6.00) <0.001
BMI: Body mass index; CI: Confidence interval; CNY: Chinese Yuan; FHH: Family history of hypertension; FH0: Negative family history; FH1: Only one generation had hypertension; FH2+: Two or three generations had hypertension; OR: Odds ratio.

Figure 2:
Nomogram with risk classification for predicting hypertension in Han and Yugur people. FH0, FH1, and FH2+ respectively mean 0,1, 2, or more of generation(s) with FHH. BMI: Body mass index; bpm: Beat per minute; CNY: Chinese Yuan; FHH: Family history of hypertension.

Performance of the model in the training set

In terms of the prediction of hypertension, the C-index was 0.802 (95% confidence interval [CI]: 0.788, 0.817) for the model and was confirmed to be 0.800 via bootstrapping validation, which exceeded 0.75 and thus indicated that the model was suitable and sufficiently accurate. The GiViTI calibration belt [Figure 3A] showed that the 80% and 95% CI of the belt did not cross the 45° diagonal bisector (P = 0.135), suggesting that the calibration of the model on the development of participants was acceptable. The Hosmer-Lemeshow statistical value for the model was 12.4 (P = 0.134), which suggested that there was no departure from a perfect fit.

Figure 3:
GiViTI calibration belts of the model in different sets. (A) Calibration in training set, (B) Calibration in internal validation set, (C) Calibration in external validation set, (D) Calibration in Han people of the internal validation set, (E) Calibration in Yugur people of the internal validation set. If the 80% or 95% CI of the belt did not cross the 45° diagonal bisector, the calibration of the model was acceptable. CI: Confidence interval.

Validation of the model

Internal validation

In internal validation set, the C-index was 0.789 (95% CI: 0.768, 0.810), and the corrected value was 0.789 via bootstrapping validation. The 95% CI of the GiViTI calibration belt did not cross the 45° diagonal bisector (Figure 3B; P= 0.114). The Hosmer-Lemeshow test also yielded a nonsignificant statistic (P = 0.244), indicating adequate model fit.

Subsequently, the internal validation set was split into Han and Yugur sets. The C-indexes and 95% CIs were 0.797 (0.772, 0.822) and 0.770 (0.728, 0.812), and the corrected values were 0.798 and 0.770 for Han and Yugur, respectively. Both in Han and Yugur sets, the 95% CIs of the GiViTI calibration belt did not cross the 45° diagonal bisector (Figure 3D; P= 0.166 and Figure 3E; P= 0.428), and the Hosmer-Lemeshow tests yielded nonsignificant statistics (P = 0.341, P= 0.878).

External validation

In external validation set, the C-index was 0.829 (95% CI: 0.816, 0.842) for the model, and the corrected value was 0.829 via bootstrapping validation. The 80% and 95% CI of the GiViTI calibration belt did not cross the 45° diagonal bisector (Figure 3C; P= 0.681). The Hosmer-Lemeshow test also yielded a nonsignificant statistic (P = 0.280).

Clinical use

The decision curve analysis for the model is shown in Figure 4. The threshold probability for hypertension was plotted on the x-axis and the y-axis measured the standard NB. The decision curves [Figure 4A] showed that, both in internal and external validation sets, using the model to predict hypertension added more benefit than either the treat, all policy or the treat, none policy. In addition, Figure 4B also shows a similar result in Han and Yugur people derived from the internal validation set.

Figure 4:
Decision curve analysis for the diagnostic model in validation sets. The y-axis displays standardized net benefit, and the x-axis displays threshold probability (Pt). (A) Decision curves in internal and external validation sets, (B) Decision curves in Han and Yugur people derived from the internal validation set.

For example, if we assumed that pt = 40%, thus, in the external validation set, the sNB using the model was 0.473, indicating that the model achieves almost half of the maximum possible achievable utility. Furthermore, we could say that the model offers the same sNB to the population as a policy that resulted in intervention for 47.3% of cases and no controls.

The clinical impact curve for the model of internal validation set [Supplementary Figure 3A, https://links.lww.com/CM9/A936] showed that the total number who would be deemed high risk was 341 of 1000 participants, and 188 of those would be true positives (cases) at the risk threshold of 40%. The corresponding numbers for the external validation set, Han and Yugur people derived from the internal validation set were 265/380 [Supplementary Figure 3B, https://links.lww.com/CM9/A936], 183/337 [Figure 3C in Appendix, https://links.lww.com/CM9/A936], and 203/351 [Supplementary Figure 3D, https://links.lww.com/CM9/A936], respectively.

True- and false-positive rates against the risk thresholds for the model are shown in Supplementary Figure 4, https://links.lww.com/CM9/A936. The figures show information similar to that of ROC curves and also show the risk thresholds corresponding to each true- and false-positive rate.

Comparison with a previously published diagnostic model

For comparison, we calculated the prediction probabilities according to the formula of Du's and our model and evaluated the AUCs. The ROC curves are presented in Figure 5. In training [Figure 5A], internal validation set [Figure 5B], and external validation set [Figure 5C], our model showed higher AUCs than the previous one (all DeLong's test P< 0.001). In Han [Figure 5D] and Yugur [Figure 5E- people derived from the internal validation set, our model also presented higher AUCs than Du's, but the difference in Yugur was not significant (P = 0.094).

Figure 5:
ROC analysis between the current model and a previously published diagnostic model. (A) ROC curves in the training set, (B) ROC curves in the internal validation set, (C) ROC curves in the external validation set, (D) ROC curves in Han people of the internal validation set, (E) ROC curves in Yugur people of the internal validation set. ROC: Receiver operating characteristic.


In this study, a diagnostic model with risk classification was developed and validated for the individualized prediction of hypertension in Gansu Province. The model incorporated eight items (age, female, BMI, rural, annual income ≥20,000 CNY, heart rate ≥80 beats per min, FH1, and FH2+). In addition, a nomogram and a simple and user-friendly website were also developed to calculate the exact probabilities of hypertension for Han and Yugur people to facilitate the use of the model. To our knowledge, this is the first study to establish a diagnostic model visualized as an easy-to-use nomogram and a website with risk classification to distinguish hypertension individually from the general population in Gansu Province. The model showed good discrimination and calibration in the training and validation sets.

For the selection of potential predictive factors, 17 candidate features were reduced to eight predictors by examining the association between predictors and hypertension by shrinking the regression coefficients with the Lasso method. This method uses L1-norm and executes both automatic variable selection and continuous shrinkage simultaneously,[25,26] thus, it is the preferred option to reduce model dimensionality and avoid overfitting when fewer predictors are desired in health-related researches.[27] Several studies compared Lasso and classical statistical methods, and the results showed that Lasso regression had better identification of predictors and better performance for prediction model selection than classical regression.[28,29]

The model was constructed according to the sociodemo-graphic, lifestyle, and physical examination characteristics, which were easily accessible to the general population, especially those living in rural areas. When participants intend to calculate their risk of hypertension, they just need to use the nomogram or the website based on the common factors (eg, age, gender, BMI, and FHH), and this avoids being unable to estimate the risk because of the poor accessibility of special measures, such as blood biochemical tests.[30] Our previous study found an increase in the risk of hypertension with the increasing number of relative generations affected by hypertension.[31] Thus, we further grouped the FHH based on the number of generations for predicting with more details. Du et al[14] established a diagnostic model for hypertension prediction in northern

China, but the FHH was not included in the model, and the model was not fully calibrated or validated either. Compared with Du's model, our model had higher AUCs in training and validation sets. Xue et al[32] developed a nomogram to identify the determinants of hypertension in patients with type 2 diabetes mellitus (T2DM) and to quickly calculate the probability of hypertension in individuals with T2DM, thus, the model could not be generalized to the general population.

The data in this study were split into two groups non-randomly by location: one to develop the prediction model and one to evaluate its predictive performance. This split type is a stronger design for evaluating model performance than the type of randomly split because it allows for nonrandom variation between the two datasets.[33] Arguably this type is commonly referred to as external validation, although it may be considered as an intermediary between internal and external validations.[7] This means that it develops and validates a prediction model on all available data in the primary set (meaning it still belongs to internal validation) and achieves the effect of external validation through nonrandomly split.[34,35] In addition, the data of Han people from Hebei Province were used as the external validation set. In our study, although significant differences of most candidate predictors existed between training and validation sets after nonrandomly splitting, the model still showed good discrimination and calibration, implying that the model we developed was robust for prediction.

To ascertain the individual need for further treatment or care based on their real need is the most important and final argument for the model use. However, the performance, discrimination, and calibration of the diagnostic model could not capture the clinical consequences of a particular level of discrimination or type of miscalibra-tion.[36,37] Thus, to demonstrate the clinical usefulness, we evaluated whether the model-assisted decision making would derive clinical benefit for individuals by decision curve analysis. It was based upon the decision-theoretical framework that accounted for both the benefits of an intervention and the costs of the intervention to participants who cannot benefit.[15] By deriving the standardized NB of a model across different threshold probabilities, this method offered insight into clinical consequences and identified the range of threshold probabilities in a valuable model and the magnitude of benefit.[16] The decision curves in the current study showed that using the model to predict hypertension could add more benefit than either the treat all policy or the treat none policy.

To our knowledge, this is the first study that developed and validated the hypertension diagnostic model with risk classification specifically for Han and Yugur people in Gansu Province. The estimate from our model was found to be robust as demonstrated by internal and external validations. However, several potential limitations of this study should be mentioned. First, we only selected the common predictors that were convenient for participants and available in our study, yet did not include other common predictors, such as salt intake and psychosocial factors. As salt sensitivity is a strong independent risk factor for hypertension,[38] the lack of salt intake data may affect the efficiency of the final model. Second, our model was developed in the context of the Han and Yugur people, and internal and external validations were performed. However, the predicted results in other different ethnic populations are uncertain. Third, our study participants did not include the participants whose age <20 years old or >80 years old, and pregnant women; thus, the external generalization of our results to these populations was unknown. Furthermore, appropriate prevention strategies would be more effective on relatively younger people. Therefore, further validation studies for these prediction models are warranted.

In conclusion, we developed and validated a diagnostic model with risk classification for hypertension prediction in Han and Yugur people in Gansu Province, and further developed a nomogram and a simple and user-friendly website (https://chris-yu.shinyapps.io/hypertension_risk_prediction/), which could be conveniently used to facilitate the individualized prediction of hypertension in the general population. Because of the simplicity and easily obtained measures, the nomogram and website with risk classification may be helpful to identify high-risk populations and improve prevention and treatment strategies for Han and Yugur people.


We sincerely express our gratitude to Biao Zhang, Guodong Xu, Haiying Gong, Fen Dong, Guoju Li, Yanlong Li, Xiaoyang Wang, Ya Tuo, Bin Liu, Huiru Ren, Qianqian Liu, Jiangling Gao, Jigang Yu, and all the staff of Gansu and Hebei Centers for Disease Control for support with the collection of demographic data. We are also grateful to all the study participants.


This work was supported by CAMS Innovation Fund for Medical Sciences (Nos. 2020-I2M-2-009,2020-I2M-2-003).

Conflicts of interest



1. Jeffrey DS, Ashkan A, Emmanuela G, Stephen SL. Global, regional, and national comparative risk assessment of 84 behavioural, environmental and occupational, and metabolic risks or clusters of risks for 195 countries and territories, 1990-2017: A systematic analysis for the Global Burden of Disease Study 2017. Lancet 2018; 392:19231994. doi: 10.1016/S0140-6736(18)32225-6.
2. Mills KT, Bundy JD, Kelly TN, Reed JE, Kearney PM, Reynolds K, et al. Global disparities of hypertension prevalence and control. Circulation 2016; 134:441450. doi: 10.1161/CIRCULATIO-NAHA.115.018912.
3. Chobanian AV, Bakris GL, Black HR, Cushman WC, Green LA, Izzo JJ, et al. the seventh report of the joint national committee on prevention, detection, evaluation, and treatment of high blood pressure: the JNC 7 report. JAMA 2003; 289:25602572. doi: 10.1001/jama.289.19.2560.
4. Wang Z, Chen Z, Zhang L, Wang X, Hao G, Zhang Z, et al. Status of hypertension in china: results from the china hypertension survey, 2012-2015. Circulation 2018; 137:23442356. doi: 10.1161/CIR-CULATIONAHA.117.032380.
5. Yoon SS, Gu Q, Nwankwo T, Wright JD, Hong Y, Burt V. Trends in blood pressure among adults with hypertension. Hypertension 2015; 65:5461. doi: 10.1161/HYPERTENSIONAHA.114.04012.
6. Yu C, Zhao H, Pan L, Zhang J, Wang X, Chang L, et al. The additive interaction between body mass index and hypertension family history in Han and Yugur: the China National Health Survey (CNHS). Int J Hypertens 2019; 2019:19. doi: 10.1155/2019/8268573.
7. Collins GS, Reitsma JB, Altman DG, Moons K. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): The TRIPOD statement. Ann Intern Med 2015; 162:55L63. doi: 10.1186/s12916-014-0241-z.
8. 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; 148:102110. doi: 10.7326/0003-4819-148-2-200801150-00005.
9. Otsuka T, Kachi Y, Takada H, Kato K, Kodani E, Ibuki C, et al. Development of a risk prediction model for incident hypertension in a working-age Japanese male population. Hypertens Res 2015; 38:419425. doi: 10.1038/hr.2015.41.
10. Chien KL, Hsu HC, Su TC, Chang WT, Sung FC, Chen MF, et al. Prediction models for the risk of new-onset hypertension in ethnic Chinese in Taiwan. J Hum Hypertens 2010; 25:294303. doi: 10.1038/jhh.2010.63.
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; 54:496501. doi: 10.1161/HYPERTENSIONAHA.109.132373.
12. Zheng L, Sun Z, Zhang X, Li J, Hu D, Chen J, et al. Predictive value for the rural Chinese population of the Framingham hypertension risk model: results from Liaoning province. Am J Hypertens 2013; 27:409414. doi: 10.1093/ajh/hpt229.
13. Ren Z, Rao B, Xie S, Li A, Wang L, Cui G, et al. A novel predicted model for hypertension based on a large cross-sectional study. Sci Rep 2020; 10:10615doi: 10.1038/s41598-020-64980-8.
14. Du M, Yin S, Wang P, Wang X, Wu J, Xue M, et al. Self-reported hypertension in Northern China: a cross-sectional study of a risk prediction model and age trends. BMC Health Serv Res 2018; 18:475doi: 10.1186/s12913-018-3279-3.
15. Kerr KF, Brown MD, Zhu K, Janes H. Assessing the clinical impact of risk prediction models with decision curves: guidance for correct interpretation and appropriate use. J Clin Oncol 2016; 34:25342540. doi: 10.1200/JCO.2015.65.5654.
16. Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making 2006; 26:565574. doi: 10.1177/0272989X06295361.
17. He H, Pan L, Pa L, Cui Z, Ren X, Wang D, et al. Data resource profile: the China National Health Survey (CNHS). Int J Epidemiol 2018; 47:17341735. doi: 10.1093/ije/dyy151.
18. Liu L. Hypertension WGOC. 2018 Chinese guidelines for the management of hypertension (in Chinese). Chin J Cardiovasc Med 2019; 24:2456. doi: 10.3969/j.issn.1007-5410.2019.01.002.
19. Tibshirani R. Regression shrinkage and selection via the Lasso. J R Statis Soc Series B-Methodol 1996; 58:267288. doi: 10.1111/j.2517-6161.1996.tb02080.x.
20. Grimes DA. The nomogram epidemic: resurgence of a medical relic. Ann Intern Med 2008; 149:273275. doi: 10.7326/0003-4819-149-4-200808190-00010.
21. Chang C, Hsieh M, Chang W, Chiang AJ, Chen J. Determining the optimal number and location of cutoff points with application to data of cervical cancer. PLoS One 2017; 12:e176231doi: 10.1371/journal.pone.0176231.
22. Jalali A, Alvarez-Iglesias A, Roshan D, Newell J. Visualising statistical models using dynamic nomograms. PLoS One 2019; 14:e225253doi: 10.1371/journal.pone.0225253.
23. Kramer AA, Zimmerman JE. Assessing the calibration of mortality benchmarks in critical care: The Hosmer-Lemeshow test revisited. Crit Care Med 2007; 35:20522056. doi: 10.1097/01.CCM.0000275267.64078.B0.
24. Nattino G, Finazzi S, Bertolini G. A new test and graphical tool to assess the goodness of fit of logistic regression models. Stat Med 2016; 35:709720. doi: 10.1002/sim.6744.
25. Wang S, Nan B, Rosset S, Zhu J. Random lasso. Ann Appl Stat 2011; 5:468485. doi: 10.1214/10-AOAS377.
26. Zou H, Hastie T. Regularization and variable selection via the elastic net. J R Statis Soc SeriesB-Statis Methodol 2005; 67:301320. doi: 10.1111/j.1467-9868.2005.00503.x.
27. Pavlou M, Ambler G, Seaman SR, Guttmann O, Elliott P, King M, et al. How to develop a more accurate risk prediction model when there are few events. BMJ 2015; 351:h3868doi: 10.1136/bmj.h3868.
28. Göbl CS, Bozkurt L, Tura A, Pacini G, Kautzky-Willer A, Mittlbock M. Application of penalized regression techniques in modelling insulin sensitivity by correlated metabolic parameters. PLoS One 2015; 10:e141524doi: 10.1371/journal.pone.0141524.
29. Kim SM, Kim Y, Jeong K, Jeong H, Kim J. Logistic LASSO regression for the diagnosis of breast cancer using clinical demographic data and the BI-RADS lexicon for ultrasonography. Ultrasonography 2017; 37:3642. doi: 10.14366/usg.16045.
30. Ma L, Zhao H, Yu M, Wen Y, Zhao T, Yan M, et al. Association of epoxide hydrolase 2 gene Arg287Gln with the risk for primary hypertension in Chinese. Int J Hypertens 2020; 2020:2351547doi: 10.1155/2020/2351547.
31. Yu C, Pan L, Ren X, Zhang J, Zhao H, Wang X, et al. Association between family history and hypertension in Han population of Gansu Province (in Chinese). Chin J Dis Control Prev 2018; 22:875879. doi: 10.16462/j.cnki.zhjbkz.2018.09.002.
32. Xue M, Liu L, Wang S, Su Y, Lv K, Zhang M, et al. A simple nomogram score for screening patients with type 2 diabetes to detect those with hypertension: a cross-sectional study based on a large community survey in China. PLoS One 2020; 15:e236957doi: 10.1371/journal.pone.0236957.
33. Moons KG, Altman DG, Reitsma JB, Ioannidis JPA, Macaskill P, Steyerberg EW, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med 2015; 162:W1W73. doi: 10.7326/M14-0698.
34. Wang J, Zhang Z, Zhou Z, Gu H. Clinical prediction models: model validation. Chin J Evid Based Cardiovasc Med 2019; 11:141144. doi: 10.3969/j.issn.1674-4055.2019.02.04.
35. Steyerberg EW, Harrell FE. Prediction models need appropriate internal, internal-external, and external validation. J Clin Epidemiol 2015; 69:245247. doi: 10.1016/j.jclinepi.2015.04.005.
36. Localio AR, Goodman S. Beyond the usual prediction accuracy metrics: reporting results for clinical decision making. Ann Intern Med 2012; 157:294295. doi: 10.7326/0003-4819-157-4-201208210-00014.
37. Van Calster B, Vickers AJ. Calibration of risk prediction models. Med Decis Making 2014; 35:162169. doi: 10.1177/0272989X14547233.
38. Pilic L, Pedlar CR, Mavrommatis Y. Salt-sensitive hypertension: mechanisms and effects of dietary and other lifestyle factors. Nutr Rev 2016; 74:645658. doi: 10.1093/nutrit/nuw028.

Hypertension; Diagnostic model; Decision curve; Yugur people; The China National Health Survey

Supplemental Digital Content

Copyright © 2023 The Chinese Medical Association, produced by Wolters Kluwer, Inc. under the CC-BY-NC-ND license.