General Information
Conflict of Interest
Request Clinical Trials
Associated Trial(s):- NCT01615029 - An Open Label, International, Multicenter, Dose Escalating Phase I/II Trial Investigating the Safety of Daratumumab in Combination With Lenalidomide and Dexamethasone in Patients With Relapsed or Relapsed and Refractory Multiple Myeloma
- NCT00473512 - A Phase I/II Open Label Study of the 17α-Hydroxylase/ C17,20 Lyase Inhibitor, Abiraterone Acetate in Patients With Prostate Cancer Who Have Failed Hormone Therapy
- NCT00473746 - Phase I/II Open Label Dose Escalation Study of the 17α-Hydroxylase/ C17,20-Lyase Inhibitor, Abiraterone Acetate in Hormone Refractory Prostate Cancer
Request Clinical Trials
Data Request Status
Status: OngoingResearch Proposal
Project Title: Improving Phase II Performance by Borrowing Phase I Data
Scientific Abstract:
Background:
In oncology drug development, the evaluation of efficacy in Phase II trials has become more rigorous under the FDA's Project Optimus initiative. As a result, conventional sample sizes used in Phase II trials may be insufficient to achieve adequate statistical power. To address this, we propose borrowing information from Phase I data to enhance the efficiency and robustness of Phase II efficacy analyses.
Objective:
To develop and evaluate a method that adaptively borrows information from Phase I using covariate similarity between patient populations.
Study Design:
We consider a seamless Phase I/II design. Phase I employs the BF-BOIN-ET design to assess both safety and efficacy and identify optimal dose levels based on utility scores. In Phase II, two doses are selected and compared in a randomized setting. To integrate Phase I data, we use a power prior framework weighted by covariate similarity between phases.
Participants:
Adult cancer patients eligible for early-phase trials. A case study on metastatic pancreatic ductal adenocarcinoma is used to illustrate the method.
Outcome Measures:
Primary: Objective Response Rate (ORR). Secondary: posterior estimates under different borrowing schemes.
Statistical Analysis:
Covariate similarity is assessed using Maximum Mean Discrepancy (MMD) based on kernel functions. The borrowing weight is calculated as w = 1 - (MMD / max.MMD), where max.MMD is the theoretical maximum. This weight controls the degree of contribution from Phase I data in the posterior analysis of ORR in Phase II.
Brief Project Background and Statement of Project Significance:
The development of new oncology drugs involves a multi-phase clinical trial process, with Phase I trials primarily focused on evaluating safety and identifying dose-limiting toxicities, while Phase II trials aim to assess preliminary efficacy at selected dose levels. Traditionally, Phase II trials have proceeded with fixed sample sizes and without incorporating information from Phase I beyond dose selection. However, the regulatory landscape is evolving. Under the FDA's Project Optimus initiative, sponsors are now encouraged to adopt more rigorous and data-driven approaches to dose optimization, including evaluation of multiple dose levels and a stronger emphasis on efficacy data early in development.
This paradigm shift has introduced new methodological challenges. In particular, the demand for robust efficacy evaluation in early-phase trials often conflicts with the practical limitations of patient recruitment and resource allocation, especially in rare or aggressive cancers. Conventional sample sizes in Phase II may lack sufficient power to detect clinically meaningful effects, thereby increasing the risk of both false-negative and false-positive conclusions.
To address these challenges, our project proposes a novel statistical approach to seamlessly borrow information from Phase I into Phase II analysis. Rather than assuming complete homogeneity between the trial phases, our method introduces a kernel-based measure of covariate similarity that allows flexible, data-adaptive borrowing. This strategy aims to maximize the utility of early-phase patient data while preserving the integrity of the efficacy analysis.
The significance of this project lies in its potential to improve the efficiency and reliability of early-phase oncology trials. By reducing dependence on large sample sizes and enabling more informed decision-making, the proposed method aligns with the objectives of modern regulatory frameworks and supports accelerated access to promising treatments for patients. Furthermore, the method's nonparametric foundation makes it robust to model misspecification, which is especially important in complex and heterogeneous patient populations typical of oncology studies.
Specific Aims of the Project:
This project aims to develop and evaluate a flexible information-borrowing framework for seamless Phase I/II oncology trials that improves statistical power and reduces required sample sizes in Phase II. Specifically, we seek to:
1. Develop a kernel-based similarity metric (using Maximum Mean Discrepancy) to quantify covariate distribution similarity between patient populations in Phase I and Phase II. This metric will serve as the basis for determining the degree of statistical borrowing.
2. Incorporate Phase I data into Phase II analysis through a power prior framework, where the influence of historical (Phase I) data is adaptively controlled by the similarity metric. This allows for a tailored balance between efficiency and robustness, depending on the comparability of populations.
3. Evaluate the operating characteristics (bias, variance, type I error, power) of the proposed method through extensive simulation studies under various clinical scenarios, including cases with mild to strong covariate imbalance.
4. Apply the proposed method to a real-world case study using publicly available data from a Phase I/II trial in metastatic pancreatic cancer to demonstrate practical feasibility and performance.
We hypothesize that this approach will lead to increased probability of success in Phase II, reduced sample size requirements, and ultimately, shorter clinical trial durations--while maintaining the validity and interpretability of efficacy evaluations.
Study Design: Individual trial analysis
What is the purpose of the analysis being proposed? Please select all that apply.: Develop or refine statistical methods
Software Used: RStudio
Data Source and Inclusion/Exclusion Criteria to be used to define the patient sample for your study: We request data on patient populations who were enrolled according to the actual clinical trial inclusion criteria and excluded based on the exclusion criteria. We do not have any specific requests regarding the detailed inclusion or exclusion criteria.
Primary and Secondary Outcome Measure(s) and how they will be categorized/defined for your study:
The primary outcome measure of this analysis is the Objective Response Rate (ORR), defined as the proportion of patients achieving a predefined tumor response according to standard criteria (e.g., RECIST). In the context of our seamless Phase I/II trial design, ORR is used to assess the preliminary efficacy of selected dose levels in Phase II.
Specifically, for each treatment arm in Phase II, we evaluate whether the observed ORR exceeds a prespecified efficacy threshold. The posterior distribution of the ORR is estimated using a Bayesian framework, where Phase I data are incorporated into Phase II analysis via a power prior. The degree of borrowing is determined by the similarity of covariate distributions between Phase I and II patients, quantified using a kernel-based Maximum Mean Discrepancy (MMD) metric.
Under this framework, the ORR posterior is calculated as Beta(a, b), where the shape parameters are informed by both current (Phase II) and discounted historical (Phase I) data. By dynamically adjusting the borrowing weight based on patient similarity, we aim to maintain validity while improving statistical precision.
Although we do not explicitly designate a secondary outcome, we also report posterior mean estimates and 95% credible intervals for ORR under various borrowing scenarios (e.g., no borrowing, full borrowing, dynamic borrowing). These results serve to evaluate the robustness and operating characteristics of the proposed method.
Our primary focus is not on the clinical effect per se, but on the performance and behavior of ORR estimation under adaptive borrowing. Therefore, ORR functions both as a clinical measure and as a metric for statistical method evaluation.
Main Predictor/Independent Variable and how it will be categorized/defined for your study:
The main independent variable in this study is the borrowing weight (denoted as w), which determines the extent to which information from Phase I is incorporated into the Phase II efficacy analysis. This weight is derived from a kernel-based similarity metric--specifically, the Maximum Mean Discrepancy (MMD)--which quantifies the difference between covariate distributions of patients in Phase I and Phase II.
The borrowing weight is defined as:
w = 1 - (MMD / max.MMD),
where max.MMD is the theoretical maximum value of MMD under the chosen kernel (e.g., 2 for Gaussian kernels). As w approaches 1, more information is borrowed from Phase I; as w approaches 0, Phase I data are effectively ignored.
This variable is not manipulated in a traditional experimental sense but is treated as the predictor in simulations and applied case analyses. It determines how much Phase I data influence the posterior estimation of the Objective Response Rate (ORR), which is the primary outcome.
In addition, dose level is also treated as a predictor in the comparison of efficacy between two arms in Phase II. The two dose levels are selected based on utility scores from Phase I. The relationship between dose level and ORR is used to assess comparative efficacy.
Together, borrowing weight and dose level serve as the key independent variables tested for their effect on ORR, under varying assumptions of patient similarity and treatment effect. These variables will be explicitly included in simulation settings and case study analysis, and clearly defined in the final publication.
Other Variables of Interest that will be used in your analysis and how they will be categorized/defined for your study:
The following demographic and baseline variables will be used as covariates in this study to assess the similarity between Phase I and Phase II patient populations:
Age: Continuous variable (years), summarized by mean and standard deviation.
Sex: Categorical variable (male/female).
ECOG Performance Status: Categorical variable with levels 0 and 1, used to assess baseline functional status.
Tumor Stage at Diagnosis: Categorical variable, including stages IIA, III, and IV, representing the extent of disease progression at initial diagnosis.
Other clinical baseline factors: Any additional variables available in the real-world or simulated dataset (e.g., prior treatment status, biomarkers) will be incorporated if deemed relevant.
These variables are selected because they are likely to influence treatment response and are commonly collected in oncology clinical trials. In our analysis, they will be used to calculate the covariate similarity metric (Maximum Mean Discrepancy) between Phase I and II groups. This similarity will inform the degree of information borrowing in the power prior framework.
All variables will be included in the kernel-based similarity computation as either continuous or categorical inputs, with appropriate preprocessing (e.g., standardization or one-hot encoding if necessary). Their distributions will be compared across Phase I and II to quantify population similarity and to ensure appropriate use of historical data.
Statistical Analysis Plan:
We propose a statistical analysis plan that focuses on evaluating the performance of a covariate-adaptive information borrowing approach for Bayesian estimation of the Objective Response Rate (ORR) in a seamless Phase I/II oncology trial design. The proposed method is motivated by the need to make efficient use of early-phase clinical trial data, while preserving statistical validity when baseline covariate distributions differ across phases.
1. Descriptive Statistics
We will first summarize patient characteristics separately for Phase I and Phase II populations using descriptive statistics. Continuous variables such as age will be summarized using means and standard deviations. Categorical variables such as sex, ECOG performance status, and tumor stage will be summarized using frequencies and percentages. These summaries will help describe population differences and potential covariate imbalance.
2. Covariate Similarity Assessment (Bivariate Analysis)
To quantify the similarity of baseline covariate distributions between Phase I and Phase II, we will use the Maximum Mean Discrepancy (MMD), a kernel-based nonparametric distance metric. MMD will be calculated based on standardized covariates using a Gaussian kernel. The bandwidth of the kernel will be selected using the median heuristic or cross-validation. The MMD measures distributional differences between groups in a way that captures both linear and nonlinear relationships.
3. Borrowing Weight Calculation
The similarity measure (MMD) will be transformed into a borrowing weight "w" using the following formula:
w = 1 - (MMD / max_MMD)
Here, max_MMD is the maximum possible value of MMD under the given kernel (e.g., 2.0 for a Gaussian kernel). The weight "w" ranges from 0 (no borrowing) to 1 (full borrowing), and determines how much Phase I data will be used in the Phase II analysis.
4. Bayesian Estimation of ORR (Multivariable Analysis)
The primary endpoint, ORR, will be modeled using a Bayesian binomial likelihood with a Beta prior. For each treatment arm (each dose level in Phase II), the posterior distribution of ORR will be computed using the power prior framework. The posterior will follow a Beta distribution with shape parameters defined as:
Posterior alpha = number of responders in Phase II + w * number of responders in Phase I + 1
Posterior beta = number of non-responders in Phase II + w * number of non-responders in Phase I + 1
We will estimate the posterior mean and 95% credible intervals of ORR for each scenario. Three borrowing scenarios will be compared: (a) no borrowing (w = 0), (b) full borrowing (w = 1), and (c) dynamic borrowing (w determined by covariate similarity).
5. Simulation Study
We will conduct simulation studies to evaluate the operating characteristics of the proposed approach under various conditions. Parameters to be varied include:
Covariate imbalance (none, mild, strong)
Sample sizes (e.g., Phase I: n = 7; Phase II: n = 25)
True ORR values
Performance metrics will include:
Posterior bias
Mean squared error (MSE)
Coverage probability of the credible interval
Statistical power
6. Case Study Analysis
We will apply the proposed method to a real-world Phase I/II clinical trial dataset in metastatic pancreatic cancer (Wainberg et al., 2021). This dataset includes detailed covariate and ORR data, allowing us to simulate Phase I and II scenarios. We will demonstrate how borrowing weights are computed and assess their impact on posterior estimates.
7. Sensitivity Analyses
We will conduct sensitivity analyses by changing the kernel function (e.g., Laplacian, linear), adjusting Beta prior hyperparameters, and using alternative definitions of max_MMD. These analyses will help assess the robustness of the results.
Narrative Summary: In cancer Phase I/II clinical trials, we consider improving the estimation accuracy in Phase II by borrowing information from Phase I data. Since the sample size per dose level in Phase I is typically small--around 10 patients or fewer--we propose a method that utilizes kernel-based approaches instead of model-based borrowing. The performance of the proposed method will be evaluated using real-world data. A data request is planned in due course.
Project Timeline:
The proposed project is expected to begin in July 2025, immediately upon receipt of the clinical trial data. The initial data cleaning and preparation phase will be conducted in late July to early August 2025, followed by exploratory analyses and implementation of the kernel-based borrowing method from August to mid-September 2025.
Simulation studies and sensitivity analyses will be conducted in parallel and are expected to be completed by the end of September 2025. The full statistical analysis, including case study application, will be finalized by October 15, 2025.
A draft of the manuscript will be prepared between October and November 2025, with the goal of submitting the manuscript to a peer-reviewed journal by November 30, 2025.
Results will be summarized and reported back to the YODA Project by December 15, 2025.
The anticipated 12-month data use period will begin in July 2025 and end in June 2026, with a possibility of extension if needed for revisions or secondary analyses.
Dissemination Plan: The research output will provide a novel clinical trial design that enables the borrowing of information from Phase I data. The manuscript is intended to be submitted to Statistics in Medicine.
Bibliography:
References
U.S. Food and Drug Administration. Project Optimus: FDA's initiative to reform dose optimization and selection in oncology drug development. https://www.fda.gov/media/164555/download, 2023. Accessed: 2025-06-05.
Aylin Sertkaya, Trinidad Beleche, Amber Jessup, and Benjamin D. Sommers. Costs of drug development and research and development intensity in the US, 2000--2018. JAMA Network Open, 7(6):e2415445--e2415445, 2024.
Peter F. Thall and John D. Cook. Dose-finding based on efficacy--toxicity trade-offs. Biometrics, 60(3):684--693, 2004.
Ick Hoon Jin, Suyu Liu, Peter F. Thall, and Ying Yuan. Using data augmentation to facilitate conduct of phase I--II clinical trials with delayed outcomes. Journal of the American Statistical Association, 109(506):525--536, 2014.
Ruitao Lin, Yanhong Zhou, Fangrong Yan, Daniel Li, and Ying Yuan. Boin12: Bayesian optimal interval phase I/II trial design for utility-based dose finding in immunotherapy and targeted therapies. JCO Precision Oncology, 4:1393--1402, 2020.
Kentaro Takeda, Masataka Taguri, and Satoshi Morita. Boin-et: Bayesian optimal interval design for dose finding based on both efficacy and toxicity outcomes. Pharmaceutical Statistics, 17(4):383--395, 2018.
Haolun Shi, Jiguo Cao, Ying Yuan, and Ruitao Lin. utpi: A utility-based toxicity probability interval design for phase I/II dose-finding trials. Statistics in Medicine, 40(11):2626--2649, 2021.
Mingyue Li, Zhonglong Guo, and Yingjie Qiu. United: A unified transparent and efficient phase I/II trial design for dose optimization accounting for ordinal graded, continuous and mixed toxicity and efficacy endpoints. Statistics in Medicine, 44(10--12):e70098, 2025.
Antje Hoering, Mike LeBlanc, and John Crowley. Seamless phase I--II trial design for assessing toxicity and efficacy for targeted agents. Clinical Cancer Research, 17(4):640--646, 2011.
Beibei Guo and Ying Yuan. Droid: Dose-ranging approach to optimizing dose in oncology drug development. Biometrics, 79(4):2907--2919, 2023.
Yanhong Zhou, J. Jack Lee, and Ying Yuan. A utility-based Bayesian optimal interval (u-boin) phase I/II design to identify the optimal biological dose for targeted and immune therapies. Statistics in Medicine, 38(28):S5299--S5316, 2019.
Yixuan Zhao, Rachael Liu, Jianchang Lin, and Ying Yuan. Bard: A seamless two-stage dose optimization design integrating backfill and adaptive randomization. arXiv preprint, arXiv:2409.15663, 2024.
Zev A. Wainberg, Tanios Bekaii-Saab, Patrick M. Boland, Farshid Dayyani, Teresa Macarulla, Kabir Mody, Bruce Belanger, Fiona Maxwell, Yan Moore, Arunthathi Thiagalingam, et al. First-line liposomal irinotecan with oxaliplatin, 5-fluorouracil and leucovorin (nalirifox) in pancreatic ductal adenocarcinoma: A phase I/II study. European Journal of Cancer, 151:14--24, 2021.