New Statistical Methods for Phase I Clinical Trials of a Single Agent

Abstract

The primary goal of phase I clinical trials in oncology is to determine a safe and possibly effective dose of a treatment, and to recommend this dose for further testing in larger trials. With chemotherapeutic treatments, the risk of severe dose-limiting toxicity (DLT) is the primary concern, assuming that the probability of efficacy will necessarily increase with dose. A phase I trial then seeks the highest dose with acceptable risk of DLT, called the maximum tolerated dose (MTD). Increasing the dose beyond the MTD would lead to unacceptable risk, while decreasing the dose would decrease the probability of benefit. In contrast, many newer therapies are molecularly targeted, where the probabilities of DLT and efficacy may plateau or even rise and then fall after some threshold. In this setting, a phase I trial must account for both toxicity and efficacy in identifying an optimal dose.

In this dissertation, we present three new approaches to modeling data in phase I trials of a single agent. Our methods improve on current practice by making more use of commonly available data. First, for chemotherapies, we investigate the utility of counting multiple DLTs per patient, in addition to counting lower-level toxicities (LLT). Typically, methods for phase I trials model only binary DLT responses, and ignore LLTs. We find that using event counts and including LLTs increases the probability of correctly identifying the MTD, particularly when the MTD is not among the highest dose levels being considered.

Second, we consider chemotherapies that are administered over multiple cycles, where dosage may vary across cycles. Multi-cycle treatments are often analyzed using only DLTs observed in the first cycle, ignoring DLTs from later cycles and thus potentially underestimating the DLT rates. We develop a latent process model, representing a continuous level of toxicity over time, which rises and falls after each administration, but which is only observed discretely in each cycle as either no toxicity, LLT, or DLT. The process is inspired by the pharmacokinetics of drug absorption and clearance. We use our model to re-estimate the MTD at the end of adaptive trials that originally used only first-cycle data, and we find that our model typically increases the probability of correctly identifying the MTD. Our method can also recommend how to adjust a patient's dose mid-treatment, to attain a target DLT rate.

Third, we develop a method for molecularly targeted therapies, incorporating both DLTs and efficacy responses and allowing the rates of both responses to vary flexibly with dose. In particular, we adopt the conditional autoregressive model, which allows us to share information between dose levels without imposing any functional form on the dose-response curves. We find that our method can adapt to a variety of dose-toxicity and dose-efficacy patterns, and often performs at least as well as competing methods.