Benefit and harm of intensive blood pressure treatment: Derivation and validation of risk models using data from the SPRINT and ACCORD trials

Abstract

Background: Intensive blood pressure (BP) treatment can avert cardiovascular disease (CVD) events but can cause some serious adverse events. We sought to develop and validate risk models for predicting absolute risk difference (increased risk or decreased risk) for CVD events and serious adverse events from intensive BP therapy. A secondary aim was to test if the statistical method of elastic net regularization would improve the estimation of risk models for predicting absolute risk difference, as compared to a traditional backwards variable selection approach.

Methods and findings: Cox models were derived from SPRINT trial data and validated on ACCORD-BP trial data to estimate risk of CVD events and serious adverse events; the models included terms for intensive BP treatment and heterogeneous response to intensive treatment. The Cox models were then used to estimate the absolute reduction in probability of CVD events (benefit) and absolute increase in probability of serious adverse events (harm) for each individual from intensive treatment. We compared the method of elastic net regularization, which uses repeated internal cross-validation to select variables and estimate coefficients in the presence of collinearity, to a traditional backwards variable selection approach. Data from 9,069 SPRINT participants with complete data on covariates were utilized for model development, and data from 4,498 ACCORD-BP participants with complete data were utilized for model validation. Participants were exposed to intensive (goal systolic pressure < 120 mm Hg) versus standard (<140 mm Hg) treatment. Two composite primary outcome measures were evaluated: (i) CVD events/deaths (myocardial infarction, acute coronary syndrome, stroke, congestive heart failure, or CVD death), and (ii) serious adverse events (hypotension, syncope, electrolyte abnormalities, bradycardia, or acute kidney injury/failure). The model for CVD chosen through elastic net regularization included interaction terms suggesting that older age, black race, higher diastolic BP, and higher lipids were associated with greater CVD risk reduction benefits from intensive treatment, while current smoking was associated with fewer benefits. The model for serious adverse events chosen through elastic net regularization suggested that male sex, current smoking, statin use, elevated creatinine, and higher lipids were associated with greater risk of serious adverse events from intensive treatment. SPRINT participants in the highest predicted benefit subgroup had a number needed to treat (NNT) of 24 to prevent 1 CVD event/death over 5 years (absolute risk reduction [ARR] = 0.042, 95% CI: 0.018, 0.066; P = 0.001), those in the middle predicted benefit subgroup had a NNT of 76 (ARR = 0.013, 95% CI: -0.0001, 0.026; P = 0.053), and those in the lowest subgroup had no significant risk reduction (ARR = 0.006, 95% CI: -0.007, 0.018; P = 0.71). Those in the highest predicted harm subgroup had a number needed to harm (NNH) of 27 to induce 1 serious adverse event (absolute risk increase [ARI] = 0.038, 95% CI: 0.014, 0.061; P = 0.002), those in the middle predicted harm subgroup had a NNH of 41 (ARI = 0.025, 95% CI: 0.012, 0.038; P < 0.001), and those in the lowest subgroup had no significant risk increase (ARI = -0.007, 95% CI: -0.043, 0.030; P = 0.72). In ACCORD-BP, participants in the highest subgroup of predicted benefit had significant absolute CVD risk reduction, but the overall ACCORD-BP participant sample was skewed towards participants with less predicted benefit and more predicted risk than in SPRINT. The models chosen through traditional backwards selection had similar ability to identify absolute risk difference for CVD as the elastic net models, but poorer ability to correctly identify absolute risk difference for serious adverse events. A key limitation of the analysis is the limited sample size of the ACCORD-BP trial, which expanded confidence intervals for ARI among persons with type 2 diabetes. Additionally, it is not possible to mechanistically explain the physiological relationships explaining the heterogeneous treatment effects captured by the models, since the study was an observational secondary data analysis.

Conclusions: We found that predictive models could help identify subgroups of participants in both SPRINT and ACCORD-BP who had lower versus higher ARRs in CVD events/deaths with intensive BP treatment, and participants who had lower versus higher ARIs in serious adverse events.

Fig 1. Flow of SPRINT trial participants (derivation cohort) and ACCORD-BP participants (validation cohort) into the current study.

Note that a large number of ACCORD-BP participants were deemed ineligible for the blood pressure study because the ACCORD trial had a factorial design in which all participants were randomized to intensive versus standard glycemic treatment, and only a subset of participants was additionally randomized to intensive versus standard blood pressure treatment (the other subset was additionally randomized to intensive versus standard lipid treatment).

Fig 2. Calibration plots for models fit by elastic net regularization versus traditional backwards selection.

Calibration plots showing the relationship between Cox-model-predicted Kaplan–Meyer event probabilities for each of the outcomes versus average observed Kaplan–Meyer event probabilities for each decile of risk in SPRINT and in ACCORD-BP. All deciles had >5 events observed per group. Diagonal lines show the perfect expected versus observed slope of 1. Note that the models required recalibration of the baseline Cox model hazard rate to fit the ACCORD-BP data (see main text and Table 2), although model coefficients were not adjusted for assessments. (A) CVD events/deaths by elastic net regularization. (B) Serious adverse events by elastic net regularization. (C) CVD events/deaths by traditional backwards selection. (D) Serious adverse events by traditional backwards selection. CVD, cardiovascular disease.

Fig 3. Predicted benefit and predicted harm from intensive blood pressure therapy based on models fit by elastic net regularization.

Scatterplot of predictive benefit and predicted harm with intensive blood pressure therapy among SPRINT participants (blue) and ACCORD-BP participants (orange), based on the Cox hazards models. The figure reveals wide variation in predicted benefit and predicted harm within both participant samples, but overall centering at lower predicted benefit and higher predicted harm for the ACCORD-BP participant sample. CVD, cardiovascular disease; int Rx, intensive treatment.

Fig 4. Predicted versus observed absolute risk differences in benefit and harm among SPRINT and ACCORD-BP trial participant subgroups, using predictions from the elastic net regularization model.

Dotted lines show the perfect predicted versus observed slope of 1. Dark colored lines show the mean of observed absolute risk differences, while light colored lines show 95% confidence intervals. (A) SPRINT, benefit. (B) ACCORD, benefit. (C) SPRINT, harm. (D) ACCORD, harm. CVD, cardiovascular disease.

Fig 5. Predicted versus observed absolute risk differences in benefit and harm among SPRINT and ACCORD-BP trial participant subgroups, using predictions from the traditional backwards selection model.

Dotted lines show the perfect predicted versus observed slope of 1. Dark colored lines show the mean of observed absolute risk differences, while light colored lines show 95% confidence intervals. (A) SPRINT, benefit. (B) ACCORD, benefit. (C) SPRINT, harm. (D) ACCORD, harm. CVD, cardiovascular disease.