Dose escalations in phase I studies: Feasibility of interpreting blinded pharmacodynamic data

Abstract

Aims: During phase I study conduct, blinded data are reviewed to predict the safety of increasing the dose level. The aim of the present study was to describe the probability that effects are observed in blinded evaluations of data in a simulated phase I study design.

Methods: An application was created to simulate blinded pharmacological response curves over time for 6 common safety/efficacy measurements in phase I studies for 1 or 2 cohorts (6 active, 2 placebo per cohort). Effect sizes between 0 and 3 between-measurement standard deviations (SDs) were simulated. Each set of simulated graphs contained the individual response and mean ± SD over time. Reviewers (n = 34) reviewed a median of 100 simulated datasets and indicated whether an effect was present.

Results: Increasing effect sizes resulted in a higher chance of the effect being identified by the blinded reviewer. On average, 6% of effect sizes of 0.5 between-measurement SD were correctly identified, increasing to 72% in 3.0 between-measurement SD effect sizes. In contrast, on average 92-95% of simulations with no effect were correctly identified, with little effect of between-measurement variability in single cohort simulations. Adding a dataset of a second cohort at half the simulated dose did not appear to improve the interpretation.

Conclusion: Our analysis showed that effect sizes <2× the between-measurement SD of the investigated outcome frequently go unnoticed by blinded reviewers, indicating that the weight given to these blinded analyses in current phase I practice is inappropriate and should be re-evaluated.

Keywords: clinical trials; drug development; pharmacodynamics; phase I.

FIGURE 1

Stochastic simulation of blinded response data for 2 cohorts with 6 active and 2 placebo subjects, presented as individual measurements (A), change from baseline (B) and summarized as mean ± standard deviation (C, D) over time. Slope of 2× between‐measurement variability of the heart rate parameter was simulated in this scenario. Dashed horizontal lines in Figure A and C present normal range based on literature. Dashed horizontal lines in Figure B and D depict reference line for no change from baseline.

FIGURE 2

Heatmap with % correct decision of each parameter over effect size, based on blinded data of 1 or 2 cohorts (A), the calculated delta (B) and all data combined (C). The % correct and the simulated effect size of a parameter is reported in each cell (C). Parameters are sorted on the level of between measurement variability (high to low). ALAT = alanine aminotransferase; GammaGT = γ‐glutamyl transferase; DiastBP supine = diastolic blood pressure in supine position; SystBP supine = systolic blood pressure in supine position; QTcF = Fridericia corrected QT interval.

FIGURE 3

Barplot with % correct decision when no effect was present in different simulation scenarios. ALAT = alanine aminotransferase; GammaGT = γ‐glutamyl transferase; DiastBP supine = diastolic blood pressure in supine position; SystBP supine = systolic blood pressure in supine position; QTcF = Fridericia corrected QT interval.

FIGURE 4

Heatmap with % correct decision of each parameter over effect size, based on blinded or unblinded data (A), the calculated delta (B) and all data combined (C). Data are presented of 2 individuals who were shown both blinded or unblinded data. The % correct when no effect was simulated is presented in C. Parameters are sorted on the level of between measurement variability (high to low). ALAT = alanine aminotransferase; GammaGT = γ‐glutamyl transferase; DiastBP supine = diastolic blood pressure in supine position; SystBP supine = systolic blood pressure in supine position; QTcF = Fridericia corrected QT interval.

FIGURE 5

Statistical power for each parameter with an effect size between 2.0 and 4.5 standard deviation effects as calculated with a linear mixed effect model for both single or double drug cohort simulations. Dotted horizontal line highlights the statistical power to detect a minimal detectable effect size with 80% certainty. ALAT = alanine aminotransferase; GammaGT = γ‐glutamyl transferase; DiastBP supine = diastolic blood pressure in supine position; SystBP supine = systolic blood pressure in supine position; QTcF = Fridericia corrected QT interval.