Statistical Innovation and Opportunities for Oncology Clinical Trials

Chaoran Hu Chair
Eli Lilly and Company
 
Monday, Aug 4: 8:30 AM - 10:20 AM
4043 
Contributed Papers 
Music City Center 
Room: CC-209C 

Main Sponsor

Biopharmaceutical Section

Presentations

A Framework for Designing and Modeling Combination Dose Response Studies

Characterizing the dose-response relationship from early phase studies is important in making dose decisions for phase 3 studies. While there is rich research in the dose-response model for mono-therapies, the dose-response model has not been studied for the combination treatment. We propose a flexible and efficient estimation process by expressing the combination response as a function of the mean responses from individual mono-therapies. We present two primary families of combination dose response models on top of the additive and multiplicative models.
Closed-form solutions for parameter estimation are provided for certain models, while more complex models may require numerical methods. We illustrate that using a two-step procedure enables robust combination dose response estimation. The variance-covariance structure of these estimates can be derived via the Delta method or bootstrap to assess model variability. This framework is particularly valuable in scenarios where mono-therapy data is already available, allowing for a structured approach to modeling combination therapies. 

Keywords

Combination dose-response

Drug interaction models

Clinical statistics 

Co-Author

Yongming Qu, Eli Lilly and Company

First Author

Zhili Qiao, Eli Lilly and Company

Presenting Author

Zhili Qiao, Eli Lilly and Company

Choice of index date for an externally controlled arm in oncology trials for late line therapies

External control arms (ECAs) are increasingly used in clinical trials for rare diseases or when ethical concerns prevent the use of a placebo. A significant challenge in including an ECA in oncology trials for pre-treated patients is to determine the proper time zero (index date) such that they are comparable with the clinical trial patients. Because the number of prior lines of therapy and patient characteristics heavily influence the assessment of clinical benefits, a naïve comparison between external controls and the treated cohort in the current trial may lead to biased estimates of the causal treatment effect. We propose an Emax frailty model to simulate the correlated recurrent event times for multiple lines of therapy. We conduct extensive simulation studies to assess the impact on the bias of treatment effect, Type 1 error, and power of the Wald test from Cox models. Various methods of selecting time zero for a starting treatment line are considered, including random selection, all lines with and without variance adjustment, and matching treatment lines using propensity scores. We also assess the impact of adjusting for both measured and unmeasured confounders. 

Keywords

External control

Index date

Emax model



frailty

Propensity score 

Co-Author(s)

Dai Feng, AbbVie
Laura Fernandes, COTA Healthcare Inc.
Haijun Ma, Exelixis
Philip He, Daiichi Sankyo Inc.
Arunava Chakravartty, Novartis

First Author

Binbing Yu, AstraZeneca

Presenting Author

Binbing Yu, AstraZeneca

Dose Optimization Design using Patient-Reported Outcomes via Bayesian Additive Regression Trees

In drug development, dose optimization is crucial and challenging due to the inherent variability and exploratory nature in early phase trials. It requires careful evaluation of dose-response and toxicity to ensure that the treatment is effective and safe while maintaining an acceptable level of tolerability. The FDA's Project Optimus Initiative highlights the importance of refined dose optimization strategies. To address this challenge, we propose the Dose Optimization Design using Patient-Reported Outcomes via Bayesian Additive Regression Trees (DOD-PRO-BART). This innovative approach integrates patient-reported outcomes with clinician-reported toxicity and efficacy data, enabling a more personalized and patient-centered method in both dose escalation and dose randomization. DOD-PRO-BART provides a comprehensive assessment of treatment effects, improving our understanding of patient responses to different dosage levels. Our simulation study illustrates that the proposed method can substantially improve the optimal dose selection by integrating patient-reported data along with clinician-reported toxicity and efficacy data. 

Keywords

Dose optimization

Dose randomization

Patient-reported outcomes

Bayesian additive regression trees 

Co-Author(s)

Yunqi Zhao
Meizi Liu, Takeda
Jianchang Lin, Takeda
Rachael Liu, Takeda Pharmaceuticals

First Author

Yu-Che Chung, Takeda

Presenting Author

Yu-Che Chung, Takeda

Optimizing Combination Therapies Using a Bayesian Adaptive Design with a Two-dimensional NDLM

The 2023 American Heart Association Guidelines have identified combination therapies as an important knowledge gap and an area of future research likely to offer the best chance of success for delayed cerebral ischemia (DCI) in patients with an aneurysmal spontaneous subarachnoid hemorrhage (aSAH). In this talk, we present an optimized Bayesian adaptive design to identify the best combination of Cilostazol and Human Albumin using a two-dimensional normal dynamic linear model. This design is shown to be smaller, stronger, faster, and benefit more trial participants than fixed and adaptive designs that use an independent model. Further, the two-dimensional approach avoids the difficulty of prespecifying the order of combination therapies required in a one-dimensional normal dynamic linear model. 

Keywords

response adaptive randomization

smoothing ratio

phase II trial 

Co-Author(s)

Renee Martin, Medical University of South Carolina
Jonathan Beall, Medical University of South Carolina
Christy Cassarly, MUSC
ADNAN QURESHI, University of Missouri
Jose Suarez, Johns Hopkins University

First Author

Byron Gajewski, University of Kansas Medical Center

Presenting Author

Byron Gajewski, University of Kansas Medical Center

Quality and Interpretability of PFS and OS in Oncology Clinical - ICH E9(R1) Perspective

In oncology drug development, the quality and interpretability of time-to-event endpoints, progression-free survival (PFS) and overall survival (OS), have long been recognized as significant challenges. These challenges arise from multiple factors, including issues related to data collection, analysis methodology for handling intercurrent events (ICEs) and missing data, trial integrity concerns related to study conduct, and complexities in interpreting multiple endpoints. For example, data challenges include imbalanced schedules of radiographic assessments, informative censoring due to Blinded Independent Central Review (BICR), incomplete or absent survival sweeps, a lack of clarity in defining lost-to-follow-up and withdrawal of consent, and inadequate collection of intercurrent event data. Adopting a prospective estimand framework helps mitigating risks associated with data collection, analysis methodology and interpretability. In this work, the DahShu Innovative Design Scientific Working Group (IDSWG) Oncology team systematically investigates the complexities of PFS and OS and presents practical recommendations according to E9(R1). 

Keywords

Progression Free Survival

Overall Survival

Estimand

Oncology clinical trials

E9(R1) 

Co-Author(s)

Revathi Ananthakrishnan, Bristol-Myers Squibb
Gu Mi, Sanofi
Laura Fernandes, COTA Healthcare Inc.
David Leung, Daiichi Sankyo
Caleb Lee, Daiichi Sankyo
Natalie Ren, iTeos Therapeutics
Sohail Chaudhry, Nektar Therapeutics
Hui Yang, Takeda
Helen Zhou, GSK
Haijun Ma, Exelixis
Cassie Dong

First Author

Philip He, Daiichi Sankyo Inc.

Presenting Author

Philip He, Daiichi Sankyo Inc.

Quantifying Potential Bias Introduced by Imbalanced Disease Assessment Schedule in PFS Analysis

It is well known that imbalanced disease assessment schedule (DAS) between arms could introduce bias in Progression-free survival (PFS) analysis. Understanding the magnitude of such potential bias on key study parameters would be greatly helpful to trial design and analysis, although have not been well-studied.

In this paper, we simulate comprehensive settings that are commonly seen in oncology clinical trials. We consider the scenarios with median PFS (mPFS) on the control arm ranging from 1.5 months to 24 months, and treatment cycles occurring every 3 weeks, 4 weeks or 6 weeks, and evaluate the impact of various disease assessment schedules on the following parameters: mPFS estimate, hazard ratio (HR) estimate, type 1 error and power. We use a real study to further illustrate how the factors including patients' deviations from the scheduled visits can impact the estimates.

In general, type 1 error is inflated and hazard ratio (treatment / control) is under-estimated when subjects in treatment arm are assessed less frequently, compared to control arm. The magnitude of the impact depends on the true median PFS, the frequency of the assessments, and how far apart the assessm 

Keywords

Progression-free Survival,

Disease Assessment Schedule

Clinical Trial Design 

Co-Author(s)

Tian Chen, Takeda
Hui Yang, Takeda
Guohui Liu, Takeda Pharmaceuticals International Co.

First Author

Xiaowei Ren, Novo Nordisk US R&D

Presenting Author

Shuli Li, Takeda

Stratification with predictive factors for clinical trials with time-to-event endpoint

In clinical trials setting, stratified randomization is frequently employed to mitigate imbalances between treatment groups related to specific factors. Such stratification variables are applied in the analyses to enhance analytical efficiency. For time-to-event endpoints, stratified Cox model is often used for analysis, which assumes a uniform treatment effect across all strata. However, the presence of predictive factors in stratified randomization and analysis, a situation that is not uncommon in real-world settings, can violate this assumption. To evaluate the impact of incorporating predictive factors in stratified analysis, the power of stratified analyses and unstratified analyses are compared. Furthermore, different methods of strata combination for stratified analysis in the context of including predictive factors are also explored and compared to address scenarios where certain strata exhibit very small event counts. 

Keywords

Stratification

Stratified analysis

Predictive factors

Time-to-event 

Co-Author(s)

Xuan Zhou, Merck & Co., Inc.
Xiao Fang

First Author

Xuan Peng, Merck & Co., Inc.

Presenting Author

Xuan Peng, Merck & Co., Inc.