Heterogeneity in Blood Pressure Response to 4 Antihypertensive Drugs: A Randomized Clinical Trial

Key Points

Question  Is there a potential for personalized drug therapy in hypertension, and, if so, what is the magnitude of the benefit of personalization?

Findings  In this randomized, double-blind, repeated crossover trial, the blood pressure response to treatments varied substantially between individuals. It was estimated that personalized treatment choice would on average lead to 4.4 mm Hg–lower systolic blood pressure than a fixed choice.

Meaning  There is heterogeneity in blood pressure response to drug therapy for hypertension, of a magnitude that warrants further research.

Importance  Hypertension is the leading risk factor for premature death worldwide. Multiple blood pressure–lowering therapies are available but the potential for maximizing benefit by personalized targeting of drug classes is unknown.

Objective  To investigate and quantify the potential for targeting specific drugs to specific individuals to maximize blood pressure effects.

Design, Setting, and Participants  A randomized, double-blind, repeated crossover trial in men and women with grade 1 hypertension at low risk for cardiovascular events at an outpatient research clinic in Sweden. Mixed-effects models were used to assess the extent to which individuals responded better to one treatment than another and to estimate the additional blood pressure lowering achievable by personalized treatment.

Interventions  Each participant was scheduled for treatment in random order with 4 different classes of blood pressure–lowering drugs (lisinopril [angiotensin-converting enzyme inhibitor], candesartan [angiotensin-receptor blocker], hydrochlorothiazide [thiazide], and amlodipine [calcium channel blocker]), with repeated treatments for 2 classes.

Main Outcomes and Measures  Ambulatory daytime systolic blood pressure, measured at the end of each treatment period.

Results  There were 1468 completed treatment periods (median length, 56 days) recorded in 270 of the 280 randomized participants (54% men; mean age, 64 years). The blood pressure response to different treatments varied considerably between individuals (P < .001), specifically for the choices of lisinopril vs hydrochlorothiazide, lisinopril vs amlodipine, candesartan vs hydrochlorothiazide, and candesartan vs amlodipine. Large differences were excluded for the choices of lisinopril vs candesartan and hydrochlorothiazide vs amlodipine. On average, personalized treatment had the potential to provide an additional 4.4 mm Hg–lower systolic blood pressure.

Conclusions and Relevance  These data reveal substantial heterogeneity in blood pressure response to drug therapy for hypertension, findings that may have implications for personalized therapy.

Trial Registration  ClinicalTrials.gov Identifier: NCT02774460

The number of people with hypertension in the world has doubled in the last 30 years.¹ Despite global access to multiple classes of highly effective blood pressure (BP)–lowering drugs,² only 1 in 4 women and 1 in 5 men with hypertension reach treatment targets.¹ While most hypertension guidelines advocate combination pharmacotherapy, many patients in routine care continue to be treated with monotherapy, with adverse effects and nonadherence being important clinical problems.^3-5

It is unknown whether the optimal choice of BP-lowering therapy varies from one person to another and whether individually targeted BP treatments can maximize clinical benefit. In clinical practice, clinicians and patients misinterpret variation in serial clinic and home measures of BP as indicating treatment effects. In fact, differences due to normal within-person variation in BP are generally much larger than the differences achieved by titrating a BP-lowering drug.^6,7

To quantify the potential for using personalized medicine strategies to maximize the BP-lowering effects of antihypertensive drugs, a trial design that can control for the large background variability in individuals’ BP levels is needed.⁸ Designs used hitherto^7,9-13 have not been able to account for this variability.

The Precision Hypertension Care (PHYSIC) Trial (ClinicalTrials.gov NCT02774460) hypothesized that there is the potential for targeting specific drugs to specific individuals to maximize BP effects.¹⁴

We used a repeated crossover design, in which some of the treatments were repeat tested within a participant, to enable us to quantify both within-patient and between-patient differences in BP response to different antihypertensive treatments.^8,15,16 In particular, the repeat testing of the same treatments within an individual separates the treatment effects from the period effects and make it possible to quantify robustly the constancy of the response to a treatment and the likely magnitude of the benefit achievable with personalization of therapy. The full protocol is in Supplement 1. The study site was the outpatient research clinic of the Department of Medicine at Uppsala University Hospital. The study was approved by the Uppsala ethics committee (2016-135) and all participants provided written informed consent.

Patients were eligible to register for the trial if they (1) were aged between 40 years and 75 years (male or female); (2) had been previously diagnosed with hypertension, with systolic BP (SBP) between 140 and 159 mm Hg within a 5-year period prior to the start of the trial; (3) were pharmacologically untreated or used BP-lowering monotherapy at the inclusion visit; (4) willing and able to discontinue current BP-lowering therapy for the trial duration; and (5) gave written informed consent to participate in the study. Subsequent randomization was done only if participants also (1) did not take any BP-lowering medication during the placebo run-in period and (2) had an office SBP between 140 and 179 mm Hg and diastolic BP at or below 109 mm Hg at the randomization visit. Exclusion criteria are listed in eTable 1 in Supplement 2 and involved conditions such as possible secondary hypertension, other serious disease, gout, cardiovascular diseases, kidney failure, diabetes, or contraindications to the trial drugs. Data on medical history were based on electronic medical records and patient self-reports.

Overall, the participants reflected a low-risk primary prevention sample with an indication for BP-lowering pharmaceutical monotherapy.^17,18

All registered participants started a run-in period of 2 weeks using opaque placebo capsules with no background BP-lowering drugs. Participants who completed the run-in period were then assigned to a sequence of 6 treatment periods administered in random order. Every participant had 1 treatment period with candesartan, 16 mg (angiotensin-receptor blocker); lisinopril, 20 mg (angiotensin-converting enzyme inhibitor); amlodipine, 10 mg (calcium channel blocker); and hydrochlorothiazide, 25 mg (thiazide); in addition, every participant repeated 2 of the treatment periods selected at random. Each treatment period was of 7 to 9 weeks’ duration, with half doses scheduled for weeks 1 and 2 and full doses for weeks 3 through 9. There were 1-week washout periods with placebo between each treatment period. Participants were provided with 1 opaque capsule per day throughout the study. The 7 to 9 weeks’ treatment duration, titration schedules, and selected doses were based on relevant guidelines^17,19 and evidence that carryover effects are negligible after 4 weeks of treatment.²⁰ Overencapsulation and drug packaging and numbering were performed by Apotek Produktion & Laboratorier AB.

All participants received all 4 drugs and were randomized equally to a second treatment period for 2 of the drugs, using a single permuted block of size 300. The order of the 6 treatment periods for each participant was then randomized without restrictions. A research nurse dispensed the investigational product as numbered blister packs of identical opaque capsules, according to a computer-generated list programmed by an independent study statistician.

Participants underwent 24-hour ambulatory BP monitoring during the last 24 hours of the run-in period and each treatment period. Measurements were sought every 20 minutes during the day and every hour during the night with monitors fitted during the morning and removed 24 hours later. Successful registrations were at least 22 hours in duration with at least 2 measurements per hour and 14 measurements in total between 10:00 and 20:00 hours. The primary outcome was daytime (10:00-20:00) ambulatory SBP.²¹

The statistical analyses were predefined in the protocol and in the statistical analysis plan, and finalized before unblinding (Supplement 1). The sample size was determined as described in the eMethods in Supplement 2. The targeted estimand was biologic efficacy variation among adherent trial participants, and the primary analysis population, determined before unblinding, comprised all treatment periods with at least 90% adherence and recorded SBP. Adherence was assessed by recording dispensed and returned capsules. Analysis was performed by allocated treatment, defined as randomized treatment except for 2 periods in 1 participant, for whom the order was accidentally switched without breaking the blinding. We used 2-sided tests with a .05 significance threshold. The analyses were performed using R version 4.1.2,²² and packages lme4,²³ lmerTest,²⁴ pbkrtest,²⁵ nloptr,²⁶ and MASS.²⁷

The primary hypothesis was tested by comparing models that did and did not allow 1 or more treatments to be more effective than other treatments on an individual basis. The null model without participant-specific benefits was a linear mixed model for SBP, with treatment period and 3 independent treatment contrasts as fixed factors, and the intercept as a random factor by participant. The null model was compared with the primary full model, which added random effects by participant for the 3 independent treatment contrasts, allowing unrestricted correlations between the 4 participant-level random effects. All models were fitted using the maximum likelihood approach. The P values were obtained by parametric bootstrap with 10 000 iterations, where the likelihood ratio between the fitted full and null models was compared with the empirical distribution of ratios for the 2 models fitted to simulated data sets from the fitted null model.

Heterogeneity for individual treatment contrasts was tested by comparing the full model vs a restricted model removing only 1 of the random effects. Confidence intervals for individual treatment contrast variance parameters were estimated from P value curves obtained using parametric bootstrap for selected parameter values, comparing the full model vs a restricted model with 1 variance parameter value fixed. This method was decided post hoc when the predefined method using the lme4²³ package was found not to work. Results are presented for all 6 pairwise treatment contrasts, which are correlated because they are determined by the contrast of 3 of the treatments to the fourth. The pairwise contrasts are considered separate research questions and no multiplicity adjustment was used. Average treatment contrasts were obtained from the primary null model using Satterthwaite degrees of freedom. Predicted mean SBPs for the participants were obtained as conditional means from the primary analysis model at the maximum likelihood fit.

An estimate of theoretical maximum mean gain from personalization in the trial population was calculated by parametric bootstrap from the primary model. Theoretical maximum mean gain from optimal choice between pairs of treatments were obtained by dividing the estimates of standard deviations with the square root of 2π, as follows from standard formulae for the half-normal distribution.

As a secondary analysis of participant-specific treatment contrasts, the data from each set of participants with 2 complete crossovers between 2 treatments were analyzed using linear regression of the treatment contrast in the second crossover on the first crossover. This directly indicated whether a participant’s individual treatment difference when switching treatment on one occasion could predict the individual difference if switching again, without which there would be no potential for person-level treatment adaptation. For this analysis, the 2 first periods on each treatment were regarded as the first crossover, and the 2 second treatment periods as the second crossover, although a participant could have both periods with the first treatment before the first period with the second treatment.

We screened 391 participants between February 20, 2017, and May 25, 2020. After placebo run-in, 280 participants were eventually randomized to a total of 1680 scheduled treatment periods. The last participant visit was on June 11, 2021. Participant flow and SBP trajectories are shown in Figure 1 and eFigure 1 in Supplement 2, respectively, and adverse events are listed in eTables 3 and 4 in Supplement 2. The primary analysis set comprised 1468 periods (median length, 56 days) in 270 participants.

The randomized participants had a mean age of 64 years, and half of them were men (54.3%). The participants had hypertension for a mean of 3 years, 62.1% had previously used BP-lowering monotherapy, and the mean office BP after placebo run-in was 154/89 mm Hg (Table 1; eTable 5 in Supplement 2).

The selected treatment doses were on average not equipotent, with participants having higher BP when taking hydrochlorothiazide than when taking other treatments, when taking amlodipine compared with lisinopril, and when taking candesartan compared with lisinopril (Table 2). This is showed graphically in Figure 2A, where the blue line illustrates the mean difference in achieved BP for each of the 6 comparisons. The black line is where the line would lie if the doses were equipotent.

Figure 2A also illustrates the large between-patient variability in mean BP, illustrated by the spread of the data points along the diagonals of the plots. Substantial within-patient variability in BP is also showed by the horizontal and/or vertical error bars plotted for the subset of data points that represent patients with 2 intervention periods taking the same treatment. Further, the panel also shows the between-treatment variability in the SBP response within individuals to one treatment vs another. Participants lying above the diagonal black line had higher BP values on the first listed treatment, and participants below the black line had higher BP values on the second listed treatment.

These data showed that variation in SBP was large between treatments on average, between participants on average, within participants taking the same treatment, and between treatments in the same participant.

The primary assessment of the potential for personalized treatment choices to maximize BP response showed a preference for the model that allowed 1 or more treatments to be more effective than others for an individual compared with a model that assumed no differences in treatment effects between individuals (P < .001; Table 2). Assuming the fitted model to be true, personalized treatment using single-drug therapy would on average lead to a 4.4 mm Hg–lower SBP in the trial population than a fixed choice (eTable 6 in Supplement 2). Taking into consideration that lisinopril was found to be on average most efficacious of the drugs at the selected doses (Table 2), personalized treatment compared with lisinopril would still lead to a 3.1 mm Hg SBP improvement (eTable 7 in Supplement 2). Figure 2B illustrates the findings graphically for each of the 6 treatment comparisons with tight grouping of the data points around the diagonals for the comparisons of candesartan vs lisinopril and amlodipine vs hydrochlorothiazide indicating the constancy of treatment responses to these 2 pairs of drugs. By contrast, the more distributed sets of data points for the other 4 comparisons illustrate the marked differences in responses to treatment between individuals and the corresponding potential for getting a greater treatment effect by selecting one drug instead of the other.

The assumptions of the primary model were checked by normal distribution plots (eFigures 2-4 in Supplement 2) and comparisons of predicted model values to observed data (eFigures 5-6 in Supplement 2). Sensitivity analyses investigating model specification, missing data, and targeted population are presented in eTable 9 and eFigures 7 through 12 in Supplement 2. The homoscedasticity assumption was seen to be violated, with lower within-participant SBP variation with amlodipine than with the other treatments. We performed a number of sensitivity analyses either excluding the amlodipine periods or using models allowing heteroscedastic residuals. The results were close to those from the primary model, with somewhat larger hydrochlorothiazide-amlodipine variation estimates from the heteroscedasticity model. In the primary analysis, 212 of 1680 potential SBP values were unobtained. We repeated the primary analysis only including participants with full adherence and no missing data in any period, and performed analyses including all valid SBP measurements regardless of adherence, by randomized treatment. Both these analyses agreed well with the primary analyses.

In a complementary analysis in only those with 2 crossovers for the treatment pair, we investigated how well a participant’s treatment difference at the first crossover between 2 treatments predicted the same difference at a repeated crossover (Figure 3 and Table 2). Once again, there was no evidence of personalized effects for the comparisons of candesartan vs lisinopril or amlodipine vs hydrochlorothiazide, but there were significant correlations across the first and second comparisons for all other treatment pairs. This analysis had lower power but was less model-dependent, and the similarity of the results showed the robustness of our findings. To further decrease model dependence, we also performed nonparametric tests, giving similar results (eTable 10 in Supplement 2).

This study provided evidence that widely used antihypertensive drugs vary in effectiveness between individuals, with potential for greater BP reductions with personalized targeting of therapy. The mean additional BP reduction achievable was substantial, of a magnitude twice that achieved by doubling the dose of a first BP-lowering drug, and more than half that of adding a second drug²⁸ on average.

Using the robust repeated crossover design that separates time period from treatment effects, this study was able to rule out large differences in response to some therapies—candesartan vs lisinopril and amlodipine vs hydrochlorothiazide—showing that within these pairs the choice of therapy was unimportant for most. However, for all other comparisons tested, the choice was important with particularly large gains to be made by personalizing the choice between candesartan vs amlodipine and for choosing between lisinopril vs amlodipine.

The potential for large BP-lowering gains from personalizing antihypertensive therapy highlights the need for a mechanism that can be used to identify which individuals will benefit most from which treatments. Broadly, personalizing therapy could be achieved either by identifying the phenotypic characteristics that are associated with enhanced response to one treatment vs another or by directly measuring the individual’s responses to a series of treatments to ascertain which is most effective. The first is a method widely used to tailor therapies to patients with cancer where treatment selection is targeted, for example, to the expression of specific receptors. An example of the latter is continuous blood glucose monitoring, which has transformed the capacity to define the effects of different glucose-lowering therapies and to tailor treatment to individuals. Considering noninvasive, wearable BP measurement devices under development, it is possible to imagine a future where continuous BP measurement could differentiate between the effectiveness of multiple drug therapies provided to patients in standardized n-of-1 testing protocols. Of note, this study does not propose the year-long process for each patient used in this trial to identify an individual’s optimal treatment.

A key strength of this study was that it was designed explicitly to assess the potential for personalized medicine in a complex multifactorial disease.⁸ The repeated crossover design is recognized as the gold standard approach^8,15,16 (limitations of other trial designs are shown in eFigure 13 in Supplement 2) and has for the first time been used with high fidelity in this study. Repeated crossover designs are underused, but they could be more challenging in other settings. Hypertension is well suited for the repeated crossover design, with 4 drug class choices in clinical equipoise, and BP is a well-behaved outcome variable because it is continuous and normally distributed on a clinically relevant scale. A specific benefit of a well-powered repeated crossover trial is the ability to not only detect the potential for benefits from personalized treatment, but also to exclude effects. For example, the current study found that little would be gained by personalizing the choices of lisinopril vs candesartan or hydrochlorothiazide vs amlodipine. The absence of any potential benefit from choosing between the 2 agents inhibiting the renin-angiotensin-aldosterone system provides some reassurance about the validity of the study—these 2 agents share multiple aspects of their mechanisms of action. In the same way, the benefits of personalization observed for 4 of the 5 other comparisons between drugs with quite different mechanisms of action aligns with expectations,¹⁰ though the reason for the absence of a potential for benefit from personalizing hydrochlorothiazide vs amlodipine therapy is unclear. The consistency of the findings across the primary analyses based on all participants, as well as the analyses restricted to the repeat comparisons, also provides support for the primary conclusions about the importance of heterogeneity in BP response to therapy.

The study also had some limitations. First, the study was done in a specific patient group and with a specific set of drugs. The run-in period and the single-center design could lead to a more homogenous sample than general grade 1 hypertension populations, which could lead to an underestimation of heterogeneity in treatment effects, although between-person BP variability in this study was very similar to that in a large population-based sample.²⁹ Whether the results are generalizable to other individuals and across the drug classes is uncertain.

Second, while this study tried to select equipotent doses of the drugs, in some comparisons this was not achieved. However, this does not invalidate the study of the research question because the analysis is focused on the constancy of within-person and between-person responses, and this evaluation does not depend on the drugs being equipotent.

Third, there was some nonadherence to scheduled treatment regimens, and this may have attenuated the statistical power of the study. On average, though, adherence to the trial protocol was very high.

Fourth, the study evaluated effects of monotherapy for practical reasons, but it is likely that there would also be benefits from personalization of the dual combination therapies recommended for initial treatment by most guidelines. Optimizing monotherapy also has significant potential value in its own right because many patients still use single-drug therapy because of nonadherence⁵ or adverse effects.⁴ Despite different names for the same BP strata, current European¹⁸ and American³⁰ guidelines both recommend initiating treatment at an SBP of 140 mm Hg for all with low risk of cardiovascular events; while the European guidelines have a place for monotherapy in these persons, the American guidelines recommend combination therapy for them. Calculation of risk was not possible in this study due to a lack of lipid assessments.

The data from this study provide evidence of a substantial heterogeneity in BP response to drug therapy for hypertension. Given the size of the likely benefits, additional studies to confirm these findings, to test for the potential of personalization of combination antihypertensive therapy, and to identify mechanisms to enable the personalization of antihypertensive therapy in routine clinical practice should be a priority.

Corresponding Author: Johan Sundström, MD, PhD, Department of Medical Sciences, Uppsala University Hospital, Entrance 40, Fifth Floor, 75185 Uppsala, Sweden (johan.sundstrom@uu.se).

Accepted for Publication: February 21, 2023.

Author Contributions: Dr Östlund had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Concept and design: Sundström, Lind, Nowrouzi, Lytsy, Marttala, Östlund.

Acquisition, analysis, or interpretation of data: Sundström, Nowrouzi, Hagström, Held, Neal, Östlund.

Drafting of the manuscript: Sundström, Nowrouzi, Östlund.

Critical revision of the manuscript for important intellectual content: Lind, Nowrouzi, Hagström, Held, Lytsy, Neal, Marttala, Östlund.

Statistical analysis: Östlund.

Obtained funding: Sundström.

Administrative, technical, or material support: Sundström, Lind, Nowrouzi, Neal, Marttala.

Supervision: Sundström, Nowrouzi.

Other - Data collection: Held.

Conflict of Interest Disclosures: Dr Sundström reported owning stock in Symptoms Europe AB and Anagram Kommunikation AB. Dr Hagström reported receiving grants from Pfizer and Amgen and personal fees from Amgen, Novo Nordisk, Bayer, AstraZeneca, Amarin, and Novartis. Dr Östlund reported fees from Uppsala University paid to his institution, Uppsala Clinical Research Center, for its participation in the PHYSIC trial during the conduct of the study. No other disclosures were reported.

Funding/Support: This study was supported by the Swedish Research Council (grant 921-2014-7126), Kjell and Märta Beijer Foundation, and Anders Wiklöf.

Role of the Funder/Sponsor: The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Data Sharing Statement: See Supplement 3.

Additional Contributions: We thank Lars Lindhagen, PhD, Uppsala Clinical Research Center, for reviewing the statistical analysis and suggesting valuable improvements for numerical optimization. He did not receive compensation.