Retrospective analysis of a fatal dose-finding trial

TL;DR

This paper critically examines a phase 1 oncology trial with a fatal toxicity, developing a Bayesian dose-response model to show that the trial's design was inherently unsafe and could have predicted the fatal outcome.

Contribution

It introduces a mechanistic Bayesian modeling approach to evaluate safety in dose-escalation trials, highlighting design flaws that compromise patient safety.

Findings

The trial's design was incompatible with safety expectations.

The Bayesian model predicted the fatal toxicity.

Design flaws could have foreseen the adverse outcome.

Abstract

The commonplace description of phase 1 clinical trials in oncology as "primarily concerned with safety" is belied by their near universal adoption of dose-escalation practices which are inherently unsafe. In contrast with dose titration, cohort-wise dose escalation regards patients as exchangeable, an indefensible assumption in the face of widely appreciated inter-individual heterogeneity in pharmacokinetics and pharmacodynamics (PKPD). I have previously advanced this argument in terms of a precautionary coherence principle that brings the well-known coherence notion of Cheung (2005) into contact with modern imperatives of patient-centeredness and precision dosing. Here, however, I explore these matters in some mechanistic detail by analyzing a trial of the bispecific T cell engager AFM11, in which a fatal toxicity occurred. To this end, I develop a Bayesian dose-response model for a single ordinal toxicity. By constructing this model's priors to align with the AFM11 trial as designed and conducted, I demonstrate the incompatibility of that design with any reasonable expectation of safety. Indeed, the model readily yields prospective estimates of toxic response probabilities that suggest the fatality in this trial could have been foreseen as likely.