Logo image
Efficient Bayesian adaptive designs for oncology clinical trials with multiple biomarker subgroups
Dissertation   Open access

Efficient Bayesian adaptive designs for oncology clinical trials with multiple biomarker subgroups

Daniel Hee Jin Kang
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2023
DOI: 10.25820/etd.007110
pdf
DHK_Thesis_12_20_2022_Final_Corrected4.33 MBDownloadView
Open Access Free to read and download

Abstract

Regulatory affirmation of master protocol trials and new understanding of genetic mechanisms for cancer have renewed efforts to develop genetic screening trials which can identify promising new targeted therapies tested in smaller genetic subgroups. Hierarchical Clustering of Multiple Biomarker Subgroups (HCOMBS) is an adaptive multi-stage phase II umbrella design which utilizes simultaneous clustering and interim hypothesis testing to reduce the heterogeneity of treatment effects across arms. This allows borrowing within clusters of arms with similar effect sizes and simultaneously removing futile arms from reaching final effect size classification. Motivated by the Genomically-guided Treatment trial in Brain Metastases (A071701) supported by the Alliance for Clinical Trials in Oncology, HCOMBS was compared to the optimal Simon’s 2-stage design and shown to decrease sample size needed per treatment arm (19 versus 25), maintain individual power for active arms near or above 80%, and control the family-wise error rate while having comparable early stopping probabilities under the null. We further developed the HCOMBS algorithm to accommodate unequal accrual rates using multiple interim clusterings to favor early stopping in individual arms over waiting for each arm to accrue to at least the interim sample size required for interim futility testing in said arm before making an interim assessment with more power. Comparing the original HCOMBS to the modified version under more realistic equal accrual settings, we found with an increasing number of inactive arms present that the modified HCOMBS required ~20 to 30 fewer total samples across all arms, with greater reductions in sample size as accrual rates became more unequal. Furthermore, the modified HCOMBS maintained family-wise error rate under 10%, while the original HCOMBS did not, for the tradeoff of a slightly reduced ability to detect an effective intervention effect across clinical scenarios in the modified HCOMBS, around 2% -11% reduced first quartile marginal power. This provides evidence that the decision to favor early stopping of multiple arms through multiple interim futility clusterings over waiting until all arms reached the interim minimum sample size maybe justified when looked at through the lens of total sample size used in the trial. Furthermore, in simulations looking at the modified HCOMBS under varying accrual scenarios, we found that the modified HCOMBS is able to meet the criteria for first quartile marginal power being greater than 80%, FWER being under 10%, and false positive rates in arms simulated at the null response rate being under 10% in all scenarios, except in one of the scenarios with an extremely fast accruing arm simulated at the null response rate. In this scenario, the fast accruing arm decreased the marginal power of the other arms and led to a situation where both the modified and original HCOMBS had first quartile marginal powers lower than 80%.Therefore, the results seem to suggest that the modified HCOMBS is more adapted to the realistic situation where there is unequal accrual and unequal timing of futility analyses across arms. The modified HCOMBS shows a lot of promise as a developing phase 2 genetic screening design. It is a multi-arm hierarchical Bayesian design which can improve marginal power and maintain control over false positive rates, can handle unequal accrual, and can reduce the overall number of individuals enrolled on ineffective treatment arms.
public abstract

Details

Metrics

5 File views/ downloads
38 Record Views
Logo image