Final published version, 1.47 MB, PDF document
Available under license: CC BY-ND: Creative Commons Attribution-NoDerivatives 4.0 International License
Research output: Thesis › Doctoral Thesis
Research output: Thesis › Doctoral Thesis
}
TY - BOOK
T1 - Bayesian adaptive methods to incorporate preclinical data into phase I clinical trials
AU - Zheng, Haiyan
PY - 2019
Y1 - 2019
N2 - Basing informed decisions on available, relevant information is essential in all phases of drug development. This is particularly true in early phase clinical trials, when our knowledge about toxicity of a new medicine remains limited. Thus, borrowing of information across seemingly disparate sources is appealing. Statistical literature has been written about augmenting a new clinical trial with data from historical studies designed for similar investigational purpose. But very few has looked into leveraging preclinical data into phase I first-in-man trials.The work presented in this thesis attempts to fill the gap by providing solutions in the Bayesian paradigm, with purposes of improving the design and analysis of adaptive phase I dose-escalation trials. Specifically, our focus is on the transition step of early drug development, where phase I clinical trials are preceded only with some preclinical information. We see preclinical data as a special type of historical data, say, historical animal data. This is not an obvious application of the existing approaches for data augmentation, since information collected from preclinical studies first needs to be translated to account for potential physiological differences between animals and humans. Furthermore, due to their intrinsic variabilities in drug metabolism, inconsistency between the translated preclinical and clinical data may still emerge however careful and correct the interspecies translation would be completed. We note this thesis will exclusively consider toxicity data, assuming that relationship between dose and risk of toxicity can be adequately described using a two-parameter logistic regression model.Grounded in Bayesian statistics, our idea is to represent preclinical data into a prior distribution for the dose-toxicity model parameters that underpin the human trial(s). Our aim is to propose robust Bayesian approaches, keeping in mind the possibility that toxicity in humans could be very different from what we have learnt in one or multiple animal species even after appropriate translation. The main challenge in statistical inferences is essentially to address issues of prior-data conflict emerging in a small trial.This thesis consists of two perspectives on the robust use of preclinical animal data. A “sensible” amount of animal data to be leveraged into the phase I human trial(s) is determined by either (i) assessing the commensurability of the prior predictions of human toxicity, which are obtained using animal data alone, with the observed toxicity outcomes from the ongoing trial(s), or (ii) fitting a hierarchical model with weakly informative priors placed on the variance parameters. Correspondingly, we have proposed a Bayesian decision-theoretic approach in Chapter 2 and a robust Bayesian hierarchical model in Chapter 3, which build the core of this thesis. We have also extended the Bayesian hierarchical model to address potential heterogeneity between patient groups in Chapter 4, where the methodology has been illustrated in the context of bridging strategies considered in phase I clinical trials planned in various geographic regions.Throughout, the proposed Bayesian adaptive methods have been elucidated with representative data examples and extensive simulations. Particular attention has been paid to balancing the information from different sources to draw robust inferences. Numerical results show that our proposals have desired properties. More specifically, preclinical data can be essentially discounted when they are in fact inconsistent with the toxicity in humans. In cases of consistency, benefits are seen as increased precision of estimate of the probability of toxicity at a range of doses, and higher proportion of patients allocated to the target dose(s).
AB - Basing informed decisions on available, relevant information is essential in all phases of drug development. This is particularly true in early phase clinical trials, when our knowledge about toxicity of a new medicine remains limited. Thus, borrowing of information across seemingly disparate sources is appealing. Statistical literature has been written about augmenting a new clinical trial with data from historical studies designed for similar investigational purpose. But very few has looked into leveraging preclinical data into phase I first-in-man trials.The work presented in this thesis attempts to fill the gap by providing solutions in the Bayesian paradigm, with purposes of improving the design and analysis of adaptive phase I dose-escalation trials. Specifically, our focus is on the transition step of early drug development, where phase I clinical trials are preceded only with some preclinical information. We see preclinical data as a special type of historical data, say, historical animal data. This is not an obvious application of the existing approaches for data augmentation, since information collected from preclinical studies first needs to be translated to account for potential physiological differences between animals and humans. Furthermore, due to their intrinsic variabilities in drug metabolism, inconsistency between the translated preclinical and clinical data may still emerge however careful and correct the interspecies translation would be completed. We note this thesis will exclusively consider toxicity data, assuming that relationship between dose and risk of toxicity can be adequately described using a two-parameter logistic regression model.Grounded in Bayesian statistics, our idea is to represent preclinical data into a prior distribution for the dose-toxicity model parameters that underpin the human trial(s). Our aim is to propose robust Bayesian approaches, keeping in mind the possibility that toxicity in humans could be very different from what we have learnt in one or multiple animal species even after appropriate translation. The main challenge in statistical inferences is essentially to address issues of prior-data conflict emerging in a small trial.This thesis consists of two perspectives on the robust use of preclinical animal data. A “sensible” amount of animal data to be leveraged into the phase I human trial(s) is determined by either (i) assessing the commensurability of the prior predictions of human toxicity, which are obtained using animal data alone, with the observed toxicity outcomes from the ongoing trial(s), or (ii) fitting a hierarchical model with weakly informative priors placed on the variance parameters. Correspondingly, we have proposed a Bayesian decision-theoretic approach in Chapter 2 and a robust Bayesian hierarchical model in Chapter 3, which build the core of this thesis. We have also extended the Bayesian hierarchical model to address potential heterogeneity between patient groups in Chapter 4, where the methodology has been illustrated in the context of bridging strategies considered in phase I clinical trials planned in various geographic regions.Throughout, the proposed Bayesian adaptive methods have been elucidated with representative data examples and extensive simulations. Particular attention has been paid to balancing the information from different sources to draw robust inferences. Numerical results show that our proposals have desired properties. More specifically, preclinical data can be essentially discounted when they are in fact inconsistent with the toxicity in humans. In cases of consistency, benefits are seen as increased precision of estimate of the probability of toxicity at a range of doses, and higher proportion of patients allocated to the target dose(s).
U2 - 10.17635/lancaster/thesis/573
DO - 10.17635/lancaster/thesis/573
M3 - Doctoral Thesis
PB - Lancaster University
ER -