Research On Seamless Phase Ⅱ/Ⅲ Design For Trials With Multiple Co-primary Endpoints Based On Frequency And Bayesian Approaches

Research background and purpose:Clinical trials’ efficiency is becoming more and more important for both the pharmaceutical industry and the public sector.In general,the phase Ⅱ and phase Ⅲ trials are carried out separately,requiring different protocols to re-recruit subjects and restart the project between the end of phase Ⅱ and the beginning of phase Ⅲ.In fact,phase Ⅱ and phase Ⅲ trials may have similar research objectives,primary endpoints,and trial duration.If phase Ⅱ and phase Ⅲ are merged into a single trial,we could use one master protocol to design these two phases of clinical trials.By doing so,the time needed to collect the amount of information is reduced and sample sizes are saved;thus,the costs associated with their development are reduced.To achieve the above advantages,adaptive design came into being.The adaptive design allows early stopping and re-estimating sample size during interim analysis,thus allowing researchers to respond to interim data,thereby saving the whole study time and sample sizes.Adaptive design has been widely used in the development of oncology drugs.The seamless phase Ⅱ/Ⅲ design has been widely accepted for clinical trials with a single endpoint.However,for trials with multiple co-primary endpoints,seamless phase Ⅱ/Ⅲ design still faces some issues that need to be addressed.Our study focuses on the operating characteristics of seamless phase Ⅱ/Ⅲ designs,such as methods based on conditional power and methods based on Bayesian predictive power for trials with multiple co-primary endpoints.We explored the application of the Dirichlet-multinomial distribution model with the framework of Bayesian predictive power.Besides,we explored the performance of frequently used dynamic borrow methods based on one historical control group’s information.Methods:This study mainly focuses on the seamless phase Ⅱ/Ⅲ design for trials with multiple co-primary endpoints,primarily focus on the interim monitoring strategies and sample size re-estimation methods.Both frequency statistics approach(conditional power)and Bayesian approach(Bayesian predictive power)are used in our study.Moreover,historical control data were considered a supplementation of the concurrent control group with dynamic borrow methods.Specifically,in Chapter 2,we discuss using conditional power for futility monitoring and sample size re-estimation as Design 1 in seamless phase Ⅱ/Ⅲ vaccine trials with nine co-primary endpoints and further extend Design 1 to Design 2 with an interim analysis added for efficacy monitoring.We investigate the operating characteristics of the proposed designs in comparison with traditional trial design and group sequential design in terms of overall power,the Type I error rate,expected sample sizes,trial duration,stopping probabilities,and correct dose selection while varying the correlations among the endpoints,and the simulated effect size levels.In Chapter 3,we proposed a Bayesian predictive power approach for futility monitoring at interim analysis.The Dirichlet-multinomial model is employed to accommodate the outcomes representing the combination of seroresponses results for four binary endpoints in a non-inferior seamless phase Ⅱ/Ⅲ vaccine trial.Sample size re-estimation is conducted based on Bayesian predictive power.An alternative approach based on conditional power was also discussed for comparison purposes.Chapter 4 re-design a published seamless phase Ⅱ/Ⅲ clinical trial with two co-primary endpoints.One historical control group was used as a supplement of a control group.The Test then test,power prior,normalized power prior,and robust Meta-Analytic-Predictive prior was adopted to accommodate trials with multiple co-primary endpoints in comparison with a traditional trial conducted without a historical control group.Results:In Chapter 2,we found that Design 1 has better overall power than the other designs in most cases.Compared with group sequential design,design 2 performs better in terms of overall power when the nine endpoints are independent of each other(ρ = 0)or less relevant(ρ = 0.3).Concerning the efficacy stop percentage(ESP),when the simulated GMT response reaches 90% or more of that of the real GMT response,Design 2 tends to stop the trial early for efficacy with the probability of 98% and more.When the simulated GMT response is less than 90% of that of the real GMT response,there is 5%or more ESP for Design than for group sequential design.Regarding the stop for futility percentage,when ρ = 0 and the null hypothesis is true,FSP is 94.85% for both designs.In Chapter 3,we found out that the BPP approach performs better than or equivalent to the CP approach in terms of overall power in scenarios where the sample size used for phase Ⅱ stage is relatively small(say n1=50 or 100).The BPP method has a higher probability of stopping a futility trial correctly than the CP method with equivalent performance in overall power when phase Ⅱ stage enrolled more subjects(say n1=150 or 200).When the sample size of phase Ⅱ stage was 50,the futility stop percentage(FSP)of the BPP method is 10% or lower than that of the CP method.In general,the mean difference of FSP between BPP and CP-based methods was 14.17 ± 1.61(n1 = 50),3.28 ± 1.95(n1 = 100),0.27 ± 2.08(n1 = 150)and-1.34 ± 2.30(n1 = 200).Finally,in Chapter 4,we found that the Type I error rate of TTP,PP,NPP,and RMP were all less than 0.025.Besides,as the difference in the effect size between the historical control group and the concurrent control group increased from-0.3 to 0.3,the PP approach’s overall power gradually decreased.In terms of sample size,the NPP approach is the most sensitive,which first takes the historical control into consideration.In contrast,the TTP approach has the highest peak,meaning that the TTP approach includes the most sample sizes when the specified criteria are satisfied.Conclusion:First of all,Chapter 2 shows that the use of seamless phase Ⅱ/Ⅲ design for trials with multiple CPEs is desirable because of its flexibility and effectiveness in stopping the trial early for efficacy or futility,re-estimation the sample size based on information accumulated at interim analysis and shortening overall duration while maintaining a relatively small type 1 error rate.Design 1 performs well in most scenarios in terms of overall power while maintaining a rational sample size.Design 2 could save more sample size than Design 1 since it can stop the trial early for efficacy at the interim analysis when the differences in effect sizes between the experimental and control groups are detectable(i.e.,GMT ratio ≥ 0.8).In general,both methods can select the correct dosage and maintain a pre-specified type 1 error rate.Secondly,what stands out in Chapter 3 is that the BPP1(BPP2)approach can give the exact probability of success if the trial continues.Besides,the evaluation of BPP does not rely on a single point but calculates the likelihood of success based on the overall distribution of each endpoint’s effect size.Also,the use of the Dirichlet model in the BPP approach improves the performance of overall power by limiting the probability of stopping the study ’wrongly’ compared to the CP approach.Dirichlet conjugate distributions also help in the reduction of the computational burden of the BPP approach.Finally,in Chapter 4,the inflation of Type I error rates varies with scenarios.Generally,TTP leads to the most apparent inflation among the three dynamic information borrowing methods,while RMP has the best performance in terms of Type I error rates.Regarding the overall power,with the PP approach,the incorporation of historical information will significantly impact the final results.While the influence of the three dynamic information borrowing methods on the control group was relatively small.Among the three dynamic information borrowing methods,NNP is the most sensitive,and RMP is the most robust.