Introduction
Hypertension is a major global health issue due to its high prevalence and importance as a modifiable risk factor for cardiovascular disease and premature mortality all over the world [
1]. It may be asymptomatic up to the occurrence of clinical complications [
2] and is also hard to manage effectively because of the lack of awareness and adherence to the treatment [
3,
4].
According to global burden of disease (GBD) 2017, high systolic blood pressure (SBP) is the first leading risk factor for early death and disability, accounting for 10.4 million deaths and 218 million DALYs [
5]. The number of people with raised blood pressure has increased worldwide, mainly in low- and middle-income countries [
6]. Factors including population growth, aging, and behavioral risk factors, such as unhealthy diet, tobacco use, lack of physical activity, excess weight, and exposure to persistent stress, are attributable to the growing prevalence of hypertension [
7]. Hence, one of the global non-communicable disease (NCD) targets, adapted by the World Health Assembly in 2013, is a 25% reduction in the prevalence of high blood pressure, defined as systolic blood pressure ≥ 140 mL/Hg and diastolic blood pressure ≥ 90 mL/Hg, by 2025 [
8].
Evidence has shown that the risk of hypertension incidence depends on some clinical factors such as blood pressure, age, and BMI [
9‐
11]. Therefore, an individual approach, based on risk stratification and targeted treatment of non-hypertension people who are at high risk for high blood pressure, may be more desirable [
12], which requires a simple tool based on the prediction model. To apply such a model, ideally, the model has to be based on demographic and medical variables that are easily plain and available to non-specialized individuals and health care providers [
13]. Thus, a risk assessment tool would be useful to identify high-risk individuals who should be targeted for early interventions to prevent or postpone the development of hypertension. Such models have potential public health implications and clinical applications in the prevention of hypertension [
14].
Accordingly, several models to predict the risk of new-onset hypertension have been developed in different populations [
12,
13,
15‐
20]. Framingham hypertension risk score is a well-known and straightforward model for predicting hypertension in adults; it includes only seven simple factors and, with a c-statistic of 0.788, has a good performance in estimating the 4-year risk of developing hypertension among participants in the Framingham study. However, further testing beyond the cohort in which the risk score was developed is necessary before its implementation in a new population [
21]. We aimed to assess the predictive ability of the Framingham hypertension risk score in a Middle Eastern population-based cohort study.
Discussion
In this population-based cohort study of non-hypertensive adults aged 20 to 69 years, we applied the Framingham hypertension risk function to predict the 3-year absolute risk of incident hypertension. In the TLGS population, HRs of risk factors for incident hypertension events were significantly similar to those obtained in the Framingham study. The only difference of potential importance that we noticed was a different result in the contribution of sex to the risk of hypertension and a slightly lower hazard ratio for systolic blood pressure.
Our study, in contrast to the Framingham study [
10], showed that women were less likely to be hypertensive compared to men (HR = 0.809). In line with our study findings, some previous studies demonstrated that among individuals with the same age until the sixth decade of life, men have a higher incidence of hypertension compared to women [
24,
27‐
29]. Sex differences can be attributed to biological and behavioral factors [
30]. Although the biological differences between men and women are the same in the two communities of the TLGS and Framingham, behavioral factors, including smoking, physical activity, alcohol consumption, and other culturally related behaviors (e.g. due to religious beliefs) are different [
31,
32]. For example, Iranian women smoke less and consume less alcohol, and are less educated and more likely housekeeper. These behaviors may protect them against hypertension.
In the current study, we showed that the Framingham hypertension risk score has a high ability to discriminate individuals who developed hypertension and those did not in the TLGS cohort (c-index = 0.82). This risk score systematically underestimated the risk of hypertension; however, it was able to be corrected by the process of recalibration and model revision. We indicated that both recalibrated and revised models have proper calibration for predicting the risk of incident hypertension. The ratio of the predicted to observed risks across the entire score deciles also confirmed the improvement of the revised Framingham risk score. (The ratio improved from 0.69 (95% CI, 0.68–0.70)) for the original Framingham risk score to 0.96 (95% CI, 0.95–0.97) for the revised risk score without any overlap between CIs).
We did recalibration as the first and the most straightforward step of updating a prediction model in a new population to address systematic over-or under-estimation of the risk [
21,
25]. We also did model revision since it is a more complicated and extensive approach to updating a prediction model to modify the equation for differences in baseline incidence and the associations between the outcome and risk factors [
21]. In this way, we addressed the significantly different HR for sex in the TLGS compared to that in the Framingham population. This point affected the performance of the Framingham prediction model in our community slightly.
Given that current ADA guidelines recommend a BP goal of < 140/90 mmHg for most patients with diabetes [
33], we also assessed the validity of the Framingham risk score by including individuals with diabetes; however, the validity findings provided no marked difference.
The predictive performance of the Framingham hypertension prediction model has been tested on different populations. Consistent with our findings, the results from the MESA study showed that the Framingham risk score provides good discrimination but underestimates the risk of incident hypertension in some ethnic groups. Still, it could be corrected using a recalibration process [
34]. Also, the performance of the Framingham risk prediction model was assessed in a younger population (age 18–30 years); the model in the CARDIA population performed well but systematically underestimated the risk [
35]. In contrast, the 5-year predictive ability of the Framingham risk score in the Whitehall II study was reasonable, given both calibration and discrimination, but slightly overestimated hypertension risk among individuals < 50 years old. They showed that reclassification based on the Whitehall model, i.e., the model with the same variables in the Framingham but new beta coefficients in the Whitehall II population, does not improve the prediction [
36].
Bozorgmanesh, et al. have developed a point-score system for predicting incident hypertension in the TLGS study [
37]. The c-index for this prediction model was 0.73 among women and 0.74 among men. This is substantially lower than found for the Framingham model in our evaluation. A reason may be that their model did not include the family history of hypertension as a well-known predictor for incident hypertension [
12].
It has been demonstrated that a targeted preventive strategy in individuals at high risk of developing hypertension is an effective strategy for the prevention of hypertension [
38,
39].
Prediction models for CVDs, e.g. WHO CVD risk scores, are planned to be routinely used in primary health care using data on routinely measured conventional risk factors. Since these data are common to the hypertension prediction model, joining CVD and hypertension risk predictions in primary care can be an opportunity at no extra cost for NCD prevention programs.
Strengths and limitations
The strength of the current study is that it included a large population-based cohort of both sexes. This study also has several limitations. First, TLGS is comprised of urban adults in Tehran, and generalizing the results to mainly rural individuals should be done with caution. Also, generalization beyond the Middle East may be limited. Second, we defined the incidence of hypertension based on blood pressure measurements taken on a single visit, which may be less accurate than several measurements to confirm hypertension diagnosis; however, it is a common method in observational studies. Finally, like other population-based cohort studies, selection bias due to excluding missing data is a concern. We repeated all the analyses using multiple imputations and the results did not change (data not shown). However, since we followed the original study’s exclusion criteria for the Framingham model, we excluded missing covariates as they did.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.