Randomized controlled dose-escalation design to evaluate the safety of a novel pharmacological cardiopulmonary resuscitation strategy

Abstract

Background/aims The motivating randomized controlled phase I trial evaluates three sodium nitroprusside doses in a novel sodium nitroprusside-enhanced cardiopulmonary resuscitation strategy for improved end-organ perfusion relative to local standard of care. Sodium nitroprusside is a vasodilator with an established safety profile in other indications, whereas the local standard of care uses vasoconstrictors, typically epinephrine. The purpose of the proposed trial is to identify the highest safe dose of sodium nitroprusside in this new context as excessive doses may cause severe hypotension with compromised end-organ perfusion. Methods The proposed phase I trial design expands upon traditional dose-finding designs to include a randomized control arm, which is needed to assess safety through the relative increase in serum lactate on hospital admission. For guiding dose escalation, we propose and compare six Bayesian models which characterize expected serum lactate as a function of sodium nitroprusside dose and randomization group. Each model makes a different assumption about the expected change in serum lactate across control cohorts concurrently randomized with each dose. Model selection aims to minimize the expected number of times that a dose is incorrectly classified as safe or unsafe while sample size selection targets an expected number of incorrectly classified doses. Randomization is 1:1 for the initial cohort, and for subsequent cohorts is chosen to maximize the lower confidence bound. Results The spike-and-slab model minimizes the expected number of times that a dose is incorrectly classified as safe or unsafe under the most scenarios in the motivating three-dose trial, but all six models exhibit relatively similar performance. A 2:1 randomization ratio for the second and third cohorts maximizes the lower confidence bound when using the spike-and-slab model. With the optimal design, on average, 70 individuals will ensure 1 incorrectly classified dose in 6 opportunities. Conclusion We recommend that the motivating trial use the spike-and-slab model with a 1:1 randomization ratio for the initial cohort and 2:1 randomization ratio for subsequent cohorts; however, the simpler fixed effects approaches performed similarly well.