Recent development of risk-prediction models for incident hypertension: An updated systematic review

Search strategy

The research strategy, study selection and analysis methods used in this study followed the guidelines from the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) Statement[6] (S1 Table). We conducted a complete literature search in both PubMed and Embase to retrieve all published reports about hypertension prediction models using the keywords “hypertension”, “high blood pressure”, “prediction model”, and “risk score”. The search strategy was (((prediction model[Title/Abstract]) OR risk score)) AND ((hypertension[Title/Abstract]) OR high blood pressure[Title/Abstract]). The last search was conducted on September 5, 2016. The related references from those retrieved reports were also searched manually to identify any additional published reports. For those identified articles that were not available online, we contacted the authors directly to request copies.

Inclusion and exclusion criteria

All the retrieved reports were screened independently for inclusion by two researchers from this study. The titles and abstracts of retrieved papers were used as the primary review content for inclusion verification. However, if questioned or unclear, the full article was reviewed prior to inclusion decision. The study’s inclusion criteria include: 1. Reporting a risk assessment tool, e.g., an equation or a risk score system; 2. Predicting the risk incidence of essential hypertension; 3. Published in English-language journals; 4. Conducted in subjects 18 years old or older; 5. Reporting quantitative measures of model performance (preferred but not necessarily required). Exclusion criteria include: 1. Studies only describe association between risk factors and incident hypertension; 2. Simulation studies; 3. Studies predict gestation-related hypertension; 4. Unpublished research data.

Data extraction and synthesis

Any discrepancy of the independently collected data from the two researchers was resolved by group discussion among all participating project investigators. The following data were extracted from each study: study design, subject characteristics, number of subjects in derivation and validation cohorts, number of subjects who developed hypertension, number of candidate variables considered, variables included in the final model and statistical method used for development of the model. We extracted the area under the curve (AUC) of the receiver operating characteristic or C-statistic to assess the discrimination ability of each model. We also collected the value of Hosmer–Lemeshow χ², and the p value of the corresponding test statistic, to assess model calibration ability. Due to the wide spread of differences in risk factors, population, study design, and sufficiency of data, it was impossible to perform meta-analysis in our current study. Instead, we opted to conduct a narrative synthesis of the evidence. However, to provide a nice summary graph, we applied the random effects model meta-analysis to combine the estimates of the AUC from studies with enough data and assessed the between-study heterogeneity, with the use of the Stata statistical software version 12.0(http://www.stata.com/). The data used in meta-analysis was transformed in the way of double arcsine transformations to addresses the problems of confidence limits and variance instability. The potential publication bias was assessed with funnel plot, as well as Begg's and Egger’s test. A P value <.05 indicated significant publication bias.

Performance of prediction models

The performance of prediction models is shown in Table 2. AUC ranged from 0.64 to 0.97, and C-statistic ranged from 60% to 90%. The results of pooling 35 models in meta-analysis(S1 File) show the value of AUC as 0.767, 95%CI(0.742, 0.792). The calibration was assessed by Hosmer–Lemeshow χ², suggesting that these models had good calibration ability.

Validation of prediction models

The prediction models of 7 studies[9,10,12,13,15,26,28] were validated in internal cohorts through split samples, with C-statistics ranging from 0.79 to 0.9. Three models were externally validated. The SHIP risk model[15] from northeast Germany was validated by data from the Danish INTER99, comprising 2887 participants, and it performed well, with an AUC of 0.77 (P = 0.74) and the Hosmer–Lemeshow χ² test of 40.6 (P = 2×10⁻⁶). The KoGES risk score from Korea was externally validated by a large nationwide Korean cohort[33]. The discrimination (AUC = 0.733) and calibration (Hosmer–Lemeshow χ² = 14.85, P = 0.062) of this model were both good. The Framingham model was externally validated by 7 studies[12,15,24,33–36] from different countries (S3 Table).

Meta-analysis

Results from pooling 35 models in the meta-analysis showed that the AUC was 0.767, 95% CI(0.742, 0.792) indicating the performance of prediction models was well. Fig 2 shows the forest plots of analysis. As expected, the heterogeneity between studies(I-squared = 99.5%, Estimate of between-study variance Tau-squared = 0.0055) was significant(S1 file). Publication bias was evaluated with Funnel plot (Fig 3). The results(P>0.05) indicated no significant publication bias.

Population of studies

Most of these models were derived from American, European or East Asian populations; only one study was carried out in India, and the other 2 were in Iran. It is perceivable that systematic underestimation or overestimation of risk may occur when applying a model constructed from one particular cohort to a distinct ethnic population with different characteristics (the selection of predictors and the genetic background). We found that most prediction models were established in developed countries, and only a few were established in developing or undeveloped countries. Thus, it is imperative to establish reliable predictive models in those countries or regions to help reduce the incidence of hypertension and cardiovascular events caused by high blood pressure.

Predictors included

The most commonly used predictors include age, BMI, SBP, DBP, etc., which are easy to obtain in clinical practice. A few studies also take blood biochemistry factors or anthropometric parameters[25,27] as predictors (Table 2), which are also part of the routine lab test results in a general physical examination. The biochemistry factors used as predictors include blood glucose, triglycerides, high-density lipoprotein cholesterol (HDL-C), and fibrinogen. It has been reported that the level of blood glucose is associated with high blood pressure[37]. Triglyceride, cholesterol and HDL-C are also known to contribute to blood hyperviscosity[38] and vascular sclerosis, which could lead to the rise of the BP. Since hypertension is also considered as a metabolic disease, the changes of blood biochemical factors could provide important and valuable information for the accuracy of certain hypertension prediction models.

It is well known that the interaction between environmental and genetic factors contributes to the development of hypertension. Theoretically, the prediction models should contain both environmental and genetic predictors. Most of the SNPs used to construct GRS were from GWAS (S2 Table). In the Finnish study[16], results showed that GRS were significantly associated with BP but weakly associated with BP increase and incident hypertension; in contrast, in Hispanic Americans[11], GRS was constructed by 2 SNPs on chromosome 3 alone, and when GRS was added into the model, the improvement of predicting capability measured by the AUC was minor. In a Korean population[26], GRS was constructed by 4 SNPs based on GWAS, which was independently associated with the risk of incident hypertension. Among the 4 SNPs, rs17249754 was the same predictor as that selected in 2 Chinese genetic studies[19,21], and rs1378942 was the same as that used in the Swedish genetic study[14]. However, adding GRS into models with traditional risk factors did not significantly improve the discrimination ability. In the Swedish study[14], when adding cGRS (derived from a simple, unweighted count method) into the traditional model, AUC was marginally, but not significantly, improved (from 0.662 to 0.664). In the 29 SNPs that constructed cGRS, one (rs1378942) was the same as that selected in Korean study[26] and two (rs16998073, rs11191548) were the same as those selected in 2 Chinese studies[19,21]. A couple of factors may contribute to these unfavorable observations. First, since hypertension is a known multigene disease, a limited number of SNPs as representative predictors may not fully reflect the overall contribution and weight of all genetic variants. Second, it is possible that some of those included SNPs were selected without fully considering their potential interactions with other genetic variants or environmental factors.

In contrast, in a Chinese study[21], adding GRS constructed by 22 carefully selected SNPs to the traditional predictors produced an ideal result, as the C-index value improved significantly (C-index change = 0.3%–0.5%; all p < 0.05). Among the 22 uncorrelated (r² < 0.5) SNPs, 10 were associated with SBP or DBP from published GWAS data obtained from an East Asian population, and 19 SNPs had been identified and verified in a Chinese population. These results clearly suggested that the contribution and value of GRS to a hypertension prediction model heavily depends on the selection of SNPs. Since hypertension is a disease of polygenic inheritance, the selection of SNPs used for GRS construction is thus critical. Using GWAS results as the only source for SNP selection is inadequate, as the characteristics of SNPs obtained from one particular GWAS may not necessarily be suitable for other ethnic populations. More appropriate SNP selection should come from the genetic research results in the same ethnic group. Other SNP selection considerations for GRS construction should include a sufficient number of SNPs, causal relationship between the select genes and disease development, gene-gene or gene-environment interactions, and proper statistical methods to include or exclude gene loci.

At the present stage, genetic markers for predicting hypertension can be of great interest for researchers and basic scientists (and possibly for drug companies), but may not hold much interest for patients. Once genetic factors are included in prediction models, patients cannot use the model for self-assessment, clinicians could face problems explaining the model, and cost for genetic tests can be high. These problems may be resolved with the development of gene-function and gene-sequencing research.

Model validation

Seven studies validated their prediction models using internal validation. All studies indicated good discriminatory ability and calibration, suggesting that the models could be applied in the original population with satisfactory performance. A Framingham prediction model was validated in external populations by 7 studies (S3 Table). It performed well in a study of African-American and Caucasians in the US[35], a German study[15] and a British study[12]. In a large nationwide Korean cohort,[33]the AUC was acceptable, but this model underestimated hypertension incidence (p<0.001) in Korea. In the Multi-Ethnic Study of Atherosclerosis (MESA)[34], including Caucasian, African-American, Hispanic, and Asian (primarily of Chinese descent) participants, the Framingham model showed better discrimination ability than SBP alone or age-specific DBP categories. However, the difference between the observed and the predicted hypertension risks (Hosmer-Lemeshow goodness of fit p<0.001) in the MESA study was significant. In contrast, the discrimination (C-statistics = 0.5 to 0.6) and calibration ability (p<0.0001) in rural Chinese was poor[36], whereas poor agreement (χ² = 29.73, p = 0.0002) underestimated the risk of hypertension in Koreans[24]. The distinct performance in different populations was partially attributed to the various levels of risk predictors and inherited variables. These differences suggested that a model derived from one particular population could not be directly applied to a distinct population, and the fittest model for one particular population is that derived from the same population.

Heterogeneity

The meta-analysis showed the heterogeneity was significant. The included variables, study designs, number of participants, populations, statistical methods, and follow-up times were different from each other, which might be the source of heterogeneity. We attribute this to the specialization of prediction models, which need to be built for various populations, because no one model could be applied to all people.

Clinical implications

Currently, a large issue regarding hypertension prediction models/scores is that nobody uses these scores in daily life or clinical practice beyond research publications. Some people even question whether hypertension needs to be predicted, as it can be easily measured with noninvasive, cheap, and accessible methods. The function of these models is not only to predict the occurrence of hypertension but, more meaningfully, to remind patients and physicians to pay attention to BP. What is more, it has been proved that the process of progression from normotensive or prehypertension to hypertension can be delayed or prevented by proper and timely clinical interventions. It is urgent and meaningful for people to conduct timely interventions. The importance of prediction models/scores needs to be widely disseminated by authorities or the media to promote their application.

Strengths and limitations of existing models

Most of the current predictors are data commonly collected in routine clinical practice, which are relatively easier for both health-care providers and patients to access. Some models are in the form of risk scores, which may still have room to improve but are also convenient to use in routine clinical practice. Furthermore, several models took GRS into account, which could contribute significantly to their prediction accuracy of hypertension. Since the performance of all these prediction models was accepted as good, the application of these models in clinical practice is very promising.

In contrast, several limitations of these prediction models are also noted. First, since not all these studies were specifically designed or conducted for generating prediction models, the clinical data collection may not be complete, and quality of data collected to inform these models also varies greatly; thus, prediction accuracy is a concern. Second, the enrollment number of participants was low in some studies and may not represent the true characteristics of the general population. Third, the various levels of risk predictors and inherited variables between populations made the models inapplicable for other general people. Fourth, a justified method in selecting the suitable SNPs is lacking. Fifth, since most of the BP data were obtained in hospital or clinic settings, the “white coat effect” may influence the outcome of the BP measurement. Sixthly, none of these models have been shown to improve outcomes in prospective research. Lastly, only a few models were indeed validated by internal cohorts, and only 3 were validated in external cohorts. The validation in internal cohort is more or less considered as a repeat of the original cohort and thus may be overoptimistic in its prediction performance results.

Perspective of future research

It is obvious that the current prediction models might not be perfect, but they do provide a solid foundation for future studies. Of course, more studies on prediction models of hypertension should be conducted with large enrollment numbers, complete data collection, experienced or well-trained investigators, and appropriate statistical analysis. With the development of genetic research, more hypertension-associated SNPs will be found, and a standard protocol in gene loci selection as a candidate prediction factor will be needed. Indeed, before any models are used as guidelines, they need to be validated in various cohorts and adjusted accordingly.