The lower the better: Does simplicity lead to absurdity?

The need to estimate the impact of preventive strategies With the growing availability of results from large clinical trials, the effectiveness of cardiovascular prevention strategies is becoming better defined and the strategies are being used more widely. However, when the effect of a treatment strategy has been demonstrated, such as lowering blood pressure with antihypertensive drugs, many questions remain unanswered. What is the best blood pressure target to aim for? What is the best specific treatment? What is the impact of the treatment for individuals and different populations? Health policy decision-makers need answers to these questions to improve the health system and make it more cost efficient. One approach aiming to answer these questions is to model the impact of a treatment strategy for a specific national population. Such modelling requires a certain amount of information: (i) a reliable measure of the treatment effect of the strategy from randomised, controlled clinical trials and (ii) an accurate description of the target population, including all characteristics that are relevant for the impact prediction. In the case of the impact of blood pressure-lowering drugs, the relevant characteristics are the joint distributions of risk factors of cardiovascular events, including, amongst others, blood pressure, the structure of the cardiovascular risk in the target population, and the link between that risk structure and the joint distributions of risk factors. These elements are only partially known in a given national population. Transversal cohort studies ideally inform on the joint distributions of risk factors within the cohorts. Longitudinal studies of these cohorts inform on the link between the joint distributions of risk factors and the incidence of cardiovascular events, fatal or not: the most typical example is the Framingham study [1]. These prospective studies can be analysed together at the individual level, which increases the power of the analyses (e.g. allowing specific exploration in important strata such as by age and sex) [2]. Death statistics comprise core information that is most frequently available at the nation level, and can be used to extrapolate other information sources at that level [3]. The joint distributions of high blood pressure, age and sex worldwide Under the auspices of WHO, in this issue of the journal, Lawes et al. [4] have assembled the available information from epidemiological studies throughout the world, pertaining to the joint distributions of three risk factors: blood pressure, age and sex. This large database is summarized in a unified framework, helping us to appreciate the regional variations of these joint distributions around the world. In addition to offering a unique reference, this overview provides interesting new features, such as a steeper blood pressure increase with age in women than in men, and the identification of two regions where this phenomenon appears to be particularly large. One of its limitations is that the oldest age strata is under-represented. A model to compute the global burden of high blood pressure In an accompanying paper, Lawes et al. [5] used this database in conjunction with modelling techniques to estimate the global health burden due to high blood pressure. This population attributable burden is defined as ‘the proportional reduction in average disease risk over a specified time interval that would be achieved by eliminating the exposure of interest from the population while the distribution of other risk factors remain unchanged’. To calculate this burden, the authors used an alternative ‘ideal’ distribution as a comparator, which had a mean systolic blood pressure of 115 mmHg for the whole population at all ages. From the relative risks of cardiovascular events associated with blood pressure differences in specific age strata, it infers what would occur, over a given period of time, if the blood pressure was that of the ‘ideal’ distribution. This theoretical modelling provides interesting data on the repartition between ages and development status strata of the world regions: the so-called developed regions experience a burden of disability-adjusted life years due to an elevated blood pressure that is similar between three age groups (45–59, 60–69 and 70–79 years), and that for high mortality developing regions, the burden is highest in the 45–59 years age group. Furthermore, it is interesting to note that the model predicts that high blood pressure explains more ischaemic heart disease than stroke in some regions. However, overall, the model attributes 62% of strokes, and 49% of ischaemic heart disease events to high blood pressure. In our opinion, this is surprisingly high and unlikely to be true. Other approaches for informing prevention strategies Such high rates of events attributable to high blood pressure are worthy of discussion. This is particularly true if they are to be used to ‘gauge the potential for prevention strategies’. The modelling techniques used by Lawes et al. [5] have important limitations: (i) the mean for the counterfactual distribution is arbitrary; a systolic blood pressure of 115 mmHg corresponds to the lowest 10 or 20% of the distribution. It would be interesting to know the extent to which the fraction of estimated burden is sensitive to this arbitrary choice; (ii) other major contributors to cardiovascular risk, such as cholesterol or tobacco use, are not taken into account. Based on attributable burden estimates for those risk factors [6], the sum of all attributable burdens would be higher than the overall burden, shedding some doubt on the validity of using this arbitrary “ideal” distribution. We are especially concerned about the likelihood of misinterpretation of this analysis, which could confuse theoretical attributable burden with avoidable burden. The choice of the counterfactual distribution fits with an epidemiological perspective. However, the only way to properly estimate avoidable high blood pressure-related burden is to base it on a blood pressure reduction that is proven in randomised, controlled trials using available interventions. Thus, the avoidable burden must be modelled based on the results from randomized clinical trials rather than on observational data [7]. It has been reported that the relative risks from clinical trials fit those estimated from observational cohorts when the latter are corrected for regression dilution bias. However, as Lawes et al. [5] remind us, the correlation between blood pressure and risk is much steeper in younger than in older individuals, especially for stroke, and this was not the case in clinical trials. In addition, to make things more complex, recent approaches using individual patient data show that the blood pressure decrease does not explain all of the reduction in strokes caused by antihypertensive drugs [8]. The main reasons for considering the proposed model as a theoretical construction without real plausibility are that: (i) it is not possible to obtain a blood pressure reduction in a population to achieve the ‘ideal’ distribution. Randomized controlled trials show us that approximately 40% of patients in the treatment arm do not achieve the treatment target blood pressure (usually 140/90 mmHg); (ii) even if, in the future, we could achieve these reductions, data from trials suggest that the benefits from drugs are not as good as suggested by observational data. For example, the meta-regression by Staessen et al. [9] clearly shows that, for stroke and ischaemic heart disease events, the maximum reduction is obtained with a 15 mmHg decrease in systolic blood pressure, with no greater benefits and an increase of ischaemic heart disease events when systolic blood pressure is lowered by more than that. Is lower blood pressure invariably better? The ‘ideal’ distribution, with an average systolic blood pressure of 115 mmHg for a population comprising all ages, is unrealistic at best and absurd at worst, when seen in the light of the results obtained in clinical trials. The HOT trial is the largest study to have tested the lower-is-better hypothesis [10]. A common misinterpretation of the results from the HOT trial is that lower-is-better and that blood pressure ‘control’ is achievable, provided one adds as many drugs as are needed. However, this misinterpretation is based on a modelling analysis of the HOT trial as if it was an observational cohort. Closer scrutiny and proper analysis of the HOT trial shows that blood pressure targets are difficult to achieve in a large proportion of patients, and that there was no reduction in cardiovascular events in the two patient groups randomized to lower diastolic targets (< 85 mmHg and < 80 mmHg) compared to the group randomized to the standard diastolic target (< 90 mmHg). Why is the first interpretation so common? We propose three possible explanations. First, we think that practitioners prefer the simple recommendation “the lower the better” as opposed to a more complex one. Second, most practitioners are not able to critique the evidence, and are easily impressed and persuaded by modelling results from large clinical trials published in prestigious journals. Finally, the pharmaceutical industry clearly stands to benefit from ‘the lower-is-better’ concept and use every opportunity to reinforce it with practitioners and the public. Modelling: a tool for building our future What is the use of modelling? In reality, model building and answering questions using models comprises an important activity. These techniques allow us to devise predictions and hypotheses and to test them using available databases. When performed properly, modelling allows us to ask and answer new questions. When performed improperly, modelling does not provide the correct answers and, in its worst form, is used to supposedly confirm and propagate beliefs. Thus, modelling has the potential to have negative consequences. The consequences of an error, due to the model itself or to its interpretation, may be dramatic, particularly if, as in this case, the results have potential wide application at a global level.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.