A time-dependent model to estimate clinical benefit with lipid-lowering therapies

Introduction and methods

Large randomized clinical trials (RCTs) typically have a limited duration of several years, and therefore evidence demonstrating long term benefit of LDL-c lowering is limited. Furthermore, not all populations are included in RCTs, such as diabetes patients in primary setting. For these purposes, it can be useful to construct a model derived from past trial evidence to evaluate different scenarios not tested in RCTs.

A model was developed from data of past RCTs of lipid-lowering trials (LLTs, including statin and nonstatins trials) for a generalizable assessment of impact of LDL-c lowering on CV risk reduction. Time-dependent clinical benefit in RCTs with LLTs was the focus of the study. The model was validated by comparing predictions of the model with trial-reported outcomes. Implications of LDL-c lowering over extended (5-15 years) time in both primary and secondary prevention settings were examined. In addition, baseline risk estimates were calculated using this model.

Published data from RCTs of LLTs (statins, ezetimibe, PCSK9 inhibitors and anacetrapib) with ≥1000 individuals and end points defined as MACE or mortality [1-4] were used to construct the model. Exclusion criteria were open-label design, reported data not suitable for model estimation, trials involving special populations and trials with bococizumab. Also, data from the Mendelian randomization meta-analysis from Ference et al. [5] were used as a single study and informed the model over a time span of 40 years.

Main results

• 22 RCTs were included and enrolled primary and secondary prevention populations with follow-ups ranging from 0.3 to 6.7 years. Indicator variables in the final model were: individual end point types (nonfatal MI, ischemic stroke, CHD death, unstable angina requiring hospitalization, and coronary revascularization), LLT type, and high baseline hsCRP level.
• The model confidence intervals for outcomes were narrower than trial-reported confidence intervals because the prediction was based on information from all 22 RCTs. For example, the reported HR for non-fatal MI from the TNT trial was 0.78 (0.66-0.93) and the model-estimated HR based on TNT-specific LDL-c reduction and follow-up was 0.79 (0.75-0.83).
• There was also concordance between reported and predicted KM curves for composite event rates over trial-specific follow-up times. Also when long-term data at 40 years from the Mendelian analysis of Ference et al. was used.
• 15 out of 22 RCTs had closer predictions for a trial-reported 3-part composite of nonfatal MI, ischemic stroke and CHD death with the time-dependent model than with 7 RCTs with CTT (Cholesterol Treatment Trialists’) estimates.
• There was a slight decrease in the model performance in the recursive 1-trial holdout validation analysis, measured by mean absolute difference of 6.4% vs. 4.7% with the final model.
• Using a commonly accepted threshold of NNT ≤50 (ARR ≥2%), this threshold would be met with intensive treatment of high-intensity statin, ezetimibe, and high-dose PCSK9 inhibitor in a standard recent ACS population with ≥ 2.0 years of treatment; in the stable ASCVD population with ≥4.7 years; in the diabetes primary prevention population with ≥8.3 years; and in the primary prevention population with ≥11.0 years.

Conclusion

References

Find this article online at J Am Heart Assoc