Several Bayesian Adaptive Designs In Cancer Clinical Trials

The research and development of new anti-tumor drugs has always been the focus of market demand,and there are many statistical problems in early clinical trials that need to be solved urgently.In the thesis,we focused on the problem of dose-finding in phase I and phase I/II clinical trials,respectively,according to the characteristics of oncology drugs.Although traditional cell-based drugs can directly kill or inhibit tumor cells,they also bring serious adverse reactions.With the deepening of the research on the molecular targeting mechanism,the strategy of tumor treatment has also begun to shift from traditional non-specific chemotherapy to new specific therapeutic regimens represented by molecular targeted drugs.Such drugs usually have different clinical manifestations from cytotoxic drugs,so there may be more possibilities for their actual response curves.A common difficulty in these early clinical trials is the scarcity of sample sizes,typically only 30 to 60 samples.Therefore,Bayesian method is a common approach to solve such problems.The research in this thesis aims to propose innovative methods based on previous work to improve the estimation efficiency under limited samples.This thesis first proposes a novel oncology phase I clinical trial design.The design is based on Bayesian stochastic approximation method,using local linear modeling and a Bayesian approach to designing a sequential trial to estimate the maximum tolerated dose for a dichotomous variable toxicity endpoint.The design requires neither the prior specification of the toxicity probability skeleton nor the subjective priors of the model parameters,and thus has good robustness,making it different from several existing class meters.At the same time,the design can also borrow historical information through a prior effective sample size to enhance estimation efficiency.Numerical simulation study has shown that the estimation accuracy and safety(overdose control)performance of this design are superior to existing methods.Compared with the classic ’3+3’ design,this design still has good operability while the estimated efficiency is greatly improved.Second,this thesis proposes an adaptive design to explore optimal biological doses for mixed data types,of which the toxicity endpoint is a dichotomous variable and the efficacy endpoint is a continuous variable.The design uses two different model-based approaches(Bayesian stochastic approximation and MCPMod)to separately consider optimal doses based on the two response types,and to obtain estimates of optimal biological doses in a simple functional form.Unlike the conventional way of determining the optimal biological dose based on the utility function,this design does not need to provide the utility function in advance,but at the same time can obtain an estimate based on the optimal biological dose in common use.Simulation shows that the efficiency of estimation and overdose control of this design are significantly superior to existing methods.At the same time,the design can better adapt to different dose-response curves,since MCPMod method is added.Sensitivity simulation study also shows that the design also has good and stable practical performance for correlated mixed data.Third,this thesis transforms the traditional optimal dose exploration problem into an optimization problem with black-box constraints,and proposes a novel optimal biological dose finding design,for the data with both toxicity and efficacy endpoints as dichotomous variables,based on the Bayesian optimization algorithm.The design innovatively applies the Bayesian optimization algorithm to the fied of early clinical trials.First,model the unknown dose-response curve by using a Gaussian process,and then estimate the optimal biological dose by maximizing the corresponding acquisition function,proposed for the constraint optimization problem.Simulation shows that the design is more accurate in estimating optimal biological dose than existing nonparametric model-based or model-free-based methods,while also enabling patients to achieve higher benefits in clinical trials.This thesis proposes three innovative Bayesian adaptive designs for three different scenarios in early phase oncology clinical trial.The effectiveness of the method is verified by the comparison in numerical study.The thesis also provides the corresponding open-source software for the proposed method,which is freely released on sharing platforms(shinyapps,etc.)for the convenience of professionals.