Adaptive Phase I and II Clinical Trial Designs in Oncology with Repeated Measures using Markov Models for the Conditional Probability of Toxicity.

Abstract

We consider models for the dose toxicity relationship in early clinical trials in oncology where different dose levels of a study drug are being tested over multiple cycles in the same patient and an assessment of toxicity is made for each cycle. We propose three models using conditional probability of toxicity in specifying the dose-toxicity relationship in patients receiving repeated doses assuming that they did not have any dose limiting toxicities (DLTs) on past cycles. We first develop the conditional Markov model in a phase I settings where the patients are allowed to escalate/de-escalate dose levels, from a choice of five possibilities, over six cycles. In the second setting the conditional Markov model is applied to a completed phase II clinical trial in sarcoma patients from the paper by Worden et al. (2005) where two dose levels of the study drug, ifosamide, were tested over four cycles. The model adequately fits the dose-toxicity relationship at each of the cycles and demonstrates flexibility offered in including additional covariate terms to describe the relationship. Finally the conditional Markov model is extended to the ordinal case where patient responses are classified as severe, mild or none and might prove beneficial in assigning future doses closer to the patient's actual frailty. Bayesian estimation of the parameters is formulated and evaluated through simulations in all the three methods. Methods for utilizing the dichotomous and ordinal outcome method to conduct a phase I study, including choices for selecting doses for the next cycle for each patient, are developed and designs of clinical trials using the models in simulation settings are presented. Comparison of the dichotomous and ordinal outcome Markov models are also presented exploring the potential benefits of using ordinal outcomes in conducting a trial.