Establishment and Evaluation of a Risk-prediction Model for HBP in Elderly Patients With NAFLD From a Health Management Perspective

Objective: Elderly patients with nonalcoholic fatty liver disease (NAFLD) are at a higher risk of developing high blood pressure (HBP) and having a low quality of life. This study established an effective, individualised, early HBP risk-prediction model and proposed health management advice for the ³60 patients with NAFLD in Shanghai, China. Methods: Questionnaire surveys, physical examinations, and biochemical tests were conducted on 7,319 cases of sample data. Risk factors were screened using the least absolute shrinkage and selection operator (Lasso) model and random forest (RF) model. A risk-prediction model was established using logistic regression analysis and dynamic nomogram was drawn. The model was evaluated for discrimination, calibration, and clinical applicability using receiver operating characteristic curves (ROC), calibration curves, decision curve analysis (DCA), net reclassication index (NRI), and external validation. Results: The results suggested the model showed moderate predictive ability. The area under curve (AUC) of internal validation was 0.707 (95% CI: 0.688-0.727), the external validation AUC was 0.688 (95% CI: 0.672-0.705). The calibration plots showed good calibration, the risk threshold of the decision curve was 30-56%, and the NRI value was 0.109. Conclusion: This HBP risk factor model may be used in clinical practice to predict the HBP risk in NAFLD patients.


Introduction
Nonalcoholic fatty liver disease (NAFLD) is an acquired metabolic disease characterised by fatty deposits in the liver [1] . Its prevalence is increasing with the standard of living of the population [2] , and it is gradually becoming the most common chronic liver disease worldwide, posing an increasing economic and medical burden [3,4] . NAFLD prevalence is higher in China than in some developed countries because of the variability of disease prevalence and lifestyle changes. A projection of the future burden of NAFLD shows China is expected to have 314 million patients with NAFLD by 2030 [5] .
High blood pressure (HBP) is a common cardiovascular disease with clinical diagnostic criteria of systolic blood pressure (SBP) ≥140 mmHg or diastolic blood pressure (DBP) ≥90 mm Hg [6] , resulting in 9.4 million deaths worldwide [7] . China ranks rst worldwide for deaths due to HBP, and the number of elderly people suffering from HBP is >60% [8,9] . Recent studies have found that the risk of hypertension is notably higher in patients with NAFLD than in other populations [10,11,12] , and there is a close relationship between the two. The prevalence of NAFLD combined with hypertension is approximately 39.34% [2,13,14], and the quality of life of patients with NAFLD-HBP is lower than either alone.
Used as a quantitative tool for assessing risks and bene ts, the clinical prediction models (CPMs) can provide medical practitioners with intuitive and scienti cally valid data; therefore, their application is becoming increasingly common [15] . However, most CPMs are established in developed countries, with only a few in developing and underdeveloped countries. Through a literature review, we found few reports on the establishment of the CPMs for HBP in China, and the CPMs in different regions are signi cantly different [7,16,17,18,19] . Therefore, it is of great practical value to establish CPMs for different regions and populations in China and propose health management strategies based on the research ndings.
The HBP risk-prediction model for elderly patients with NAFLD developed in this study could, to some extent, ll the research gap in terms of region and population. Using multiple indicators obtained from questionnaires, physical examinations, and biochemical tests predict the risk of concurrent HBP in elderly patients with NAFLD and can help medical practitioners identify and select high-risk individuals for nonmedical health interventions at an early stage to delay or prevent NAFLD patients from developing HBP.
Among the samples collected in 2017, the data analyses of the 5,287 samples in this study are shown in Table 1, illustrating the basic statistical description of the sample data for NAFLD and non-NAFLD cases.
The 24 variables in the analysis do not meet the normality by test due to the large sample size, mean ± standard deviation is used for description, and Chi-square test and t-test were used to analyse whether the differences between groups were statistically signi cant.  Figure 3, Table 2).

Model construction and prediction results
Using the seven indicators selected above as independent variables and "whether NAFLD patients have concurrent HBP" as dependent variables, a model was constructed using logistic regression to obtain a prediction model for the risk of concurrent HBP in elderly NAFLD patients: HBP risk in patients with NAFLD can be simply and visually predicted based on the nomogram drawn from this model ( Figure 4). For example, suppose a patient with NAFLD had the following characteristics: age 65 years, BMI 25.5 kg/m 2 , SBP 191 mmHg, ALT 106 U/L, UREA 6.2 mmol/L, UA 465.9 µmol/L, and NEUT 4.2⋅10 9 /L, with a score of 322, corresponding to a probability of 0.915, indicating that the risk of concurrent HBP in this NAFLD patient is 91.5%. In addition, we created an online app to predict the probability of concurrent HBP in elderly patients with NAFLD based on the model established in this study. The website is https://studentluo.shinyapps.io/DynNomapp/. This app can help clinicians diagnose HBP and community medical practitioners design reasonable and effective primary and secondary prevention management programs for NAFLD combined with HBP.

Model evaluation and comparison
To evaluate the discrimination, calibration, and clinical applicability of the risk nomogram established in this study, the model was validated. The AUC of internal validation set was 0.707 (95% CI: 0.688−0.727) and the AUC of external validation set was 0.688 (95% CI: 0.672−0.705) ( Figure 5). The calibration curves showed moderate agreement, and the results of the decision curves indicated that if the risk threshold of HBP in NAFLD patients was 30%−56%, predicting the risk of HBP using this nomogram is more bene cial than implementing an intervention-all-patients scheme ( Figure 6). In addition, we analysed each variable in the model and plotted the ROC curve; the results are shown in Figure 7, Table 3.  [29] . In a prospective study by Chen et al. that included 2,785 HBP patients, the variables in their established risk-prediction model were Age, BMI, SBP, DBP, and Fasting blood glucose (FBG) [30] . In addition, in a review study of HBP prediction models, 26 articles reporting on 48 prediction models were analysed (seven domestic studies; others from Europe, the United States, Korea, Japan, and Iran). The most common variables in the HBP prediction models included age, BMI, and blood pressure level, among others [16] . Most studies on HBP prediction models in China and abroad included the predictive variables of age, SBP, BMI (height, weight), etc., consistent with the results of this study.
Age is an irreversible risk factor for HBP development. Some studies have shown that HBP prevalence increases with an increase in age, and there is a strong correlation between biological ageing and elevated blood pressure [6,31,32,33] . SBP and DBP are diagnostic indicators commonly used for HBP. This study also found that SBP was one of the main predictors of concurrent HBP and NAFLD, and there is a correlation between BMI and HBP [34,35] . The results of an HBP prediction model study based on 135,715 individuals showed that a lower BMI reduced the risk of developing HBP [17] . Among patients with HBP in the United States, 49.5% are obese (BMI > 28) [36] . Moreover, there is strong evidence that obesity (including visceral obesity) is the most common risk factor for NAFLD [3,37] . The World Health Organization's BMI classi cation criteria are not suitable for application among Asians, and some studies have de ned the BMI classi cation for a Chinese population as follows: underweight (<18.5), normal (18.5−22.9), overweight (23.0−27.5), and obese (>27.5) [38] . Based on such criteria, the obesity rate of subjects with NAFLD and HBP in the present study was 73.8%.
Studies have shown that blood pressure is an important determinant of plasma UREA concentration. The higher the blood pressure, the higher the UREA concentration. In untreated HBP patients, elevated blood pressure may be caused by renal insu ciency [39] , and UREA abnormalities are also closely associated with kidney disease development [40,41] , suggesting that HBP is a cause or consequence of kidney disease [42] .
UA is associated with chronic kidney disease and cardiovascular disease and can be a predictor of developing HBP in adults [43,44] . It is positively correlated with the blood pressure level. An increase in UA implies an increased risk of future cardiovascular disease [21,45] . In a cross-sectional study of 33,570 patients with HBP by Ren et al., the established risk-prediction model included variables such as Gender, Age, Height, Weight, and UA [17] . Meanwhile, Rosa et al. proposed that UA is an important NAFLD predictor variable [46] .
The results of previous studies suggest that the development of cardiovascular diseases, including HBP, is closely associated with white blood cells (WBCs), a marker of in ammation [45,47,48] . An abnormal rise in WBC count may predict an increased risk of developing HBP, where NEUT is a major component of WBCs. A correlation between NEUT and HBP has also been noted in some studies, with higher NEUT values in patients with HBP [49] . In addition, some studies reported higher NEUT values in obese than in non-obese individuals [50,51] .
ALT is the main indicator abnormal liver function. It is present in the liver and is the main marker of liver injury [52] . The population in our study was mainly composed of patients with NAFLD, so the predictor ALT is included. A few studies have explored the association between ALT and HBP development, with equally few studies demonstrating association between ALT and HBP [53,54,55] . One of the cross-sectional studies conducted in Chinese adults suggested a positive linear correlation between ALT and HBP development.
Elevated blood pressure is strongly associated with increased dietary sodium intake [56,57,58] . In an article published in 2016, researchers conducted a dietary study across 20 Chinese provinces, which revealed that the per capita salt intake was 9.1 g/day and sodium intake was 5.4 g/day, both exceeding the recommended maximum daily salt intake (6 g/day) and sodium intake (2 g/day) [59] . Meanwhile, excessive intake of fats, sugars, and meat in the diet coupled with reduced physical activity are the main causes of increased BMI in the population [9] . A high-protein diet leads to excessive amino acids, which are oxidised into energy, in turn increasing UREA synthesis in the body [60] . To date, the basic treatment for excessive UA is dietary control, and a study by R et al. suggested that increased intake of vitamins, such as vitamin C and E, can reduce UA levels [43] . The Dietary Approaches to Stop Hypertension program (DASH), a healthy dietary model proposed by the United States in 1997, emphasises that adequate intake of vegetables, fruits, and low-fat (or skim) milk is essential, whereas excessive animal fats should be avoided. Some studies have con rmed the effectiveness of this approach [9,61,62] . Based on the results of the present study, a low sodium, low sugar, high vitamin C and E, and moderate polyunsaturated fatty acid (PUFA) and protein diet should be selected for individuals with high HBP risk in patients with NAFLD, based on DASH, while emphasising an increase in moderate physical exercise.
To date, interventions such as dietary control, physical exercise, and weight loss for patients with NAFLD remain one of the main basic therapies to delay or prevent the progression of the disease [63,64] . In addition, lifestyle changes play an important role in controlling blood pressure and preventing cardiovascular disease [65,66,67] and are suitable for the control of any stage of blood pressure in the HBP population (including HBP patients and high-risk populations). Particularly, interventions such as a rational diet, weight control, smoking and alcohol consumption cessation, and stress reduction can be adopted in the early stages for high-risk groups [33,68,69] . Considering the commonalities between NAFLD and HBP management, our study primarily analysed the correlation between risk factors and some dietary habits, aiming to provide a reference basis for health management based on dietary management to help physicians formulate individualised dietary interventions for individuals at high risk of HBP among NAFLD patients, and to raise patients' awareness of healthy dietary management through health education, adopt interactive follow-ups to timely access and provide feedback on the implementation of individualised dietary management and the real-time dynamics of key risk indicators.

Conclusion
This study used Lasso model based on 10-fold cross-validation and RF model to screen risk factor indicators (Age, SBP, BMI, ALT, UREA, UA, and NEUT) associated with NAFLD combined with HBP in the elderly population. Logistic regression models were used to obtain a simple, effective, and economical nomogram to predict the risk of comorbid HBP in NAFLD patients ≥60 years in Shanghai, China.
Based on the results, the main management approach for elderly patients with NAFLD at risk for HBP lies in lifestyle changes, including a healthy and balanced diet, weight loss, and regular physical activity. Health service providers can formulate and improve individualised intervention measures and risk monitoring systems to increase patients' quality of life and reduce the economic and medical burden imposed by NAFLD and HBP on the elderly, their families, and society.
This study still has potential limitations. First, this is a retrospective study, and some non-conforming data were eliminated, which may lead to selection deviation. This risk-prediction model has been validated in the external validation population, but it lacks further validation in other populations in other regions and countries. Therefore, the model needs to be externally evaluated in a wider elderly population and needs further prospective research. This will con rm whether the individual intervention based on the prediction results of the model can reduce the risk of HBP in patients with NAFLD and improve living standards of NAFLD patients.

Sample collection
The sample data for this retrospective study were obtained from the Zhangjiang, Pudong New Area, Shanghai. We relied on the Shanghai Collaborative Innovation Center of Traditional Chinese Medicine Health Service of the Shanghai University of Traditional Chinese Medicine and the Zhangjiang and Beicai Community Health Service Centers, Pudong New Area, Shanghai, to collect the relevant information regarding the participants belonging to the Han population who participated in the chronic disease health screening programs conducted at the two community health centers from April 2016 to July 2017 using questionnaires, physical examinations, and biochemical tests. This study was conducted in accordance with the basic principles of the Declaration of Helsinki, and the protocol was approved by the Ethics Committee of the Shanghai University of Traditional Chinese Medicine. Informed consent was obtained from the study subjects, and they signed an informed consent form.
The research group collected 6,664 cases in 2017, who were matched for age and sex, randomised, and did not have blood relationship among the study subjects. After data processing, study subjects with missing information, incomplete indicators, and <60 years of age were excluded. The 5,287 cases with valid data were divided into two groups, a NAFLD group (N = 2,845) and a non-NAFLD group (N = 2,442). The NAFLD group was the internal training set, and the external validation set had 2,031 cases, which were obtained after preprocessing the data collected by the research group in 2016.
Blood pressure was measured using the steps recommended by international guidelines, including a quiet rest for at least 5 minutes before measurement. An electronic sphygmomanometer (Biospace, Cheonan, Korea) was used to record each subject's SBP and DBP (mmHg).
The clinical diagnostic criteria in the "Guidelines for the diagnosis and management of nonalcoholic fatty liver disease (2010 Revision)" developed by the Chinese Society of Hepatology, CMA [20] , was referred to to determine the diagnostic criteria for NAFLD. The following three conditions had to be met: (1) No history of alcohol consumption or quantity of ethanol consumption < 140 g/week for men and <70 g/week for women.
(2) Exclusion of other causes that could lead to liver steatosis, including excessive alcohol consumption, viral hepatitis, drug-induced liver disease, autoimmune liver disease, hepatolenticular degeneration, and total parenteral nutrition.
(3) Imaging or histological ndings should have met the diagnostic criteria of fatty liver disease.
The Lasso model based on 10-fold cross-validation and RF model were used to screen the risk factors of HBP in patients with NAFLD. The Lasso model is suitable for variable screening of high-dimensional data, where the variables in the analysis are rst pooled and normalised to shrink the regression coe cients for the variables toward zero, and the predictor variables with non-zero coe cients at the best lambda value are screened with the smallest possible prediction error [21,22] . The RF model is based on a recursive feature elimination approach that uses signi cant variables in the hierarchical ranking to train the random forest and eliminate irrelevant variables, followed by the repeated training and elimination processes until there is no variable change [23] .
In this study, the outcome variable was whether HBP occurred under NAFLD conditions. As it is a dichotomous variable, logistic regression analysis was used to construct a prediction model containing the nal characteristic variables [15] . The Lasso model was intersected with the predictor variables screened using RF, and the screened common variables were included in the multivariate logistic regression analysis. The p > 0.05 predictor variables were gradually eliminated to obtain the nal characteristic variables of the risk-prediction model, and the results were visualised using forest plots.
A nomogram can simplify and visualise complex regression equations, integrate data with a model, comprehensively consider the in uential variables in the analysis, and graphically display the probability of the predicted outcome [24] . Therefore, we used the rms package in R software to plot the nomogram of the logistic regression model constructed in this study. The length of the line corresponding to each analysed variable in the plot indicated the degree of in uence of that variable on the predicted outcome [25] .

Model evaluation and comparison
We used the area under the ROC curve (AUC) to evaluate the discriminatory ability of the model in internal and external validation sets [26] . The value is generally between 0.5 and 1, and the closer to 1, the better the discriminatory ability of the model and the higher the predictive performance [18] . Appropriate calibration analysis is required when applying the prediction model constructed in this study to clinical decision-making [27] ; therefore, we used the calibration curve to determine the degree of agreement between the actual status and the predicted results of the model. Because there is a possibility of false positives or false negatives in the prediction results of this model, we used DCA to determine the net bene t of the prediction model in clinical applications [15] . The NRI proposed in 2008 to evaluate the improvement of a new model in risk-prediction ability [28] was applied in this study to compare the extent to which the newly constructed prediction model outperformed the prediction model constructed based on the traditional variables used in HBP diagnosis.
The ow chart of this study design is shown in Figure 1.

Declarations Data availability
The datasets generated and analysed during the current study are available from the corresponding author on reasonable request.      The area under the ROC curve is used to determine the predictive ability of the prediction model to distinguish outcomes, such as with or without disease. X-axis represents the false positive rate predicted    Model comparison based on NRI. The NRI was 0.109.