| New drug development is a series of complicated processes with characteristics oftime-consuming, high risks and necessary requirements of repeated interimdecision-makings. A new agent must proofed for safety and efficacy by phase IIIconfirmatory clinical trials before it goes to market. The function of phase III stage is tovalidate the agent’s safety and efficacy, while the experimented agent’s dose in phase III isrecommended by the previous phase I and II stage. Therefore, the good quality of earlyphases studies is crucial to the whole process of drug development, and length of earlyphases studies also exerts a vital impact to the total studies’duration. From the other side,in early phases studies, it is commonly that the participants are server patients. So a gooddesign must achieve at least two goals, one is to protect the patients from harmful toxicdoses, another is that the patients should get most of the trial by joining the studies. All ofthe above-mentioned requirements for clinical trials indicate that the clinical trial designshould be adapted to the updated information accumulated through the trial process.Bayesian adaptive design is the most heated research areas in clinical trials design due toit’s utilization of Bayesian statistical tools. Especially in early phases stage, Bayesianadaptive designs’ performances are superior than the other counterparts offrequentist-method-based designs because there are ordinarily small sample sizes herein.Bayesian adaptive seamless phase I/II design is an important research topic that has attracted many top-level researchers.The goal of seamless phase I/II design is that the whole developing process can beexpedited and the sample size of the trial can be reduced through saving the lead timeperiod between phase I and phase II stages and meanwhile the seamless design canguarantee the validity and integrity of the trial. The Bayesian statistics naturally providestools to make a design have the feasibility of updatedly learning the toxicity and efficacyinformation of the doses along the study process, which is helpful to design all kinds offlexible interim analyses to meet the requirements of any specific trials. Although therehave been several proposed Bayesian phase I/II adaptive seamless designs, most of themare based on the idea of multiple-stage frame, that is, their design still separate the phase Istage and phase II stage. The seamless property is only reflected by borrow strengthamong these two phases.Our opinion is that the clinical trial should be regarded as an integral part and anyseparation of trial processes is just betrayal to the essence of the real trials conduct.Nowadays, the different phases categorization is just a reluctance under theunderdeveloped monitoring and statistical tools. We think that a good trial design shouldreflect the true processes of clinical trials.The first project of this thesis focuses on the seamless phase I/II design by using theBayesian inference and computational algorithms. Specifically, we propose a new frameof designing seamless phase I/II clinical trials. The details are as following:1Proposing a new framework of Bayesian phase I/II–SEARS designThe main distinction between the SEARS design and other seamless phase I/IIdesigns is that the phase I and phase II studies can be conducted simultaneously withoutany separations of the two phases by the way of Graduation rule and other interimdecision procedures. The graduation rule makes it possible that the promising doses couldbe graduated to phase II from phase I, which ultimately let the information shared acrossthe phase I and phase II stages. Therefore, the SEARS design is able to reflect the true trialdeveloping programs. SEARS design uses a Bayesian design--mTPI approach for doseescalations in phase I and takes an adaptive randomization method proposed by Huang toallocate the participants to the desirable dose levels in phase II. There are safety rules andfutility rules in phase I and II studies of SEARS design. The safety rules can protect theparticipants from harmful doses by the way of removing these over-toxic doses and the futility rule has the function of excluding the inefficacious doses from the trial studies.Furthermore, the maximum sample size in phase I is30in SEARS design due to thecommon experiences of around30subjects in phase I clinical trials.2SEARS design with binary and continuous endpointsThe endpoints here refer to the efficacy responses’ endpoints. We use the robustBeta-Binomial model to describe the efficacy response rates when the endpoints are binary.Here there are conjugate priors for efficacy rates, therefore, the relevant computations arerelatively easy and corresponding inferences are fairly stable. When the efficacy responsesare continuous type, we use two alternatives to model the efficacy rates, quadratic logisticmodel and four-parameter Emax model, in which two models, the posterior distributionsof efficacy rates are complicated and we resort to the MCMC algorithms to approximatelyget the estimations of efficacy rates in various dose levels. Specifically, we develop arandom-walk Metropolis-Hastings algorithm for estimating the quadratic logistic modeland Metropolis-Hastings within Gibbs algorithm for the four-parameter Emax model.Based on the above computational ways, we compare the performances of SEARS with aconventional design and a competing Bayesian design–XJT design through extensivesimulations, the conclusions are:(i)SEARS design can considerably save total samplesizes while keeping the desirable trial power,(ii)In all kinds of scenarios, SEARS designcan allocate a large portion of participants to the optimal doses and select the optimal doselevel accurately,(iii)The SEARS design with the quadratic logistic model andfour-parameter Emax model, the performances of the design are related with the scenarios,that is, if the model can capture the scenario, the performances will be good, otherwise,the performances will be not ideal. Therefore, we recommend to use the SEARS designbased on the Beta-Binomial model if that transforming the data from continuous type tobinary type causes no clinical information loses; if in the situation of having to use thenon-linear model-based SEARS design, the clinicians should carefully study thenon-linear curve and specific characteristics of clinical trials.Another project of this thesis is the study of non-linear model’s optimal design,which is an independent study from the above SEARS project. It is ordinarily a difficultproblem to find design points and their weights for the non-linear models due to theunknown parameters within the Fisher information matrix. The optimal multi-stage designis a way to find the solutions. Though there have been some researches showing the efficiency of the optimal multi-stage design, there still lack of related theoreticalresearches in this area. Our work focuses on the existence of optimal design for non-linearmodels. Based on the three Sigmoid Emax models proposed by Dragalin and a theorem byYang, we derive the existence theorem of optimal design, and we give the rigorousmathematical proofs. The contents of the existence theorem are: there exist optimaldesigns in the above-mentioned three Sigmoid Emax models; the optimal design points donot exceed4. This theorem facilitates the process of finding the optimal design becausethe range of searching of the optimal design points has been limited to a small interval.The contributions of this thesis are as following:(i)proposing a new phase I/II designframe—SEARS design;(ii)developing the MCMC algorithms for the quadratic logisticand four-parameter Emax models;(iii)developing not only the simulation R code ofSEARS design, but also the monitoring code;(iv)giving the existence theorem of optimaldesign for the non-linear models.The proposed SEARS design could apply straightly to the drug developing programs;the existence of the optimal design theorem is a powerful theoretical tool and reducesconsiderably the range of searching for the optimal design of the non-linear models. |
PDF Full Text Request