An efficient single-arm Bayesian adaptive trial algorithm to evaluate de-intensified oncologic treatment Authors: Yuan Zhong, Zeynep Baskurt, Mahmood Aminilari, Jennifer Seelisch, Lindsay A Renfro, Sharon M Castellino, Wei Xu, David Hodgson Published: 2025 ABSTRACT AbstractBackgroundIn clinical trials, evaluating de-intensified oncologic treatment strategies can help reduce treatment-related toxicities while preserving patients’ quality of life. However, de-intensification is typically evaluated in cancers with a low relapse rate, and if the cancer type is uncommon, a randomized trial may require an impractically extended period to accumulate sufficient events for reliable inferential conclusions.MethodThis paper introduces a Bayesian adaptive method for the single-arm trial design that provides efficient analysis of survival data under these constraints. By incorporating data from previous studies to establish prior knowledge and a historical control arm, this approach enables robust and accurate estimations and predictions for trial design, sample size determination, and inferential decision-making. To support the implementation of this method, we developed an R package called “BayesAT,” which offers significant flexibility in modelling and supports multi-stage interim analyses, particularly for evaluating de-intensified oncologic treatments.ResultOur approach is validated through comprehensive simulation studies and sensitivity analyses. Additionally, this algorithm has been applied to a pediatric Hodgkin lymphoma trial, showcasing its capability to effectively leverage information from previous studies and conduct interim analyses that expedite conclusions regarding treatment efficacy.Keywords: Bayesian adaptive trial, Single-arm trial, Multi-stage interim analysis, Bayesian statistics Introduction Oncology research frequently evaluates approaches that compare de-escalation treatment to the standard therapy, attempting to reduce treatment-related toxicities and enhance patient quality of life. For example, in the Children’s Oncology Group AHOD 0031 trial, there was no significant reduction in the event-free survival (EFS) or overall survival (OS) with the omission of radiation therapy for children with Hodgkin lymphoma who achieved a rapid and complete response to dose-dense chemotherapy [1]. Similarly, a non-inferiority trial involving 6711 breast cancer patients compared adjuvant chemoendocrine therapy with endocrine therapy alone and found that the omission of chemotherapy did not significantly compromise outcomes [2]. Furthermore, a randomized phase II trial in oropharyngeal cancer assessed reduced- or standard-dose postoperative radiation therapy in intermediate-risk patients [3], demonstrating that a de-intensification strategy resulted in equally favourable outcomes. These examples underscore the potential benefits of reducing treatment intensity in specific clinical settings. While randomized control trials (RCTs) are the gold standard for assessing the safety and effectiveness of medical interventions, they have certain limitations. Reduction of treatment intensity is typically of clinical relevance when relapse risk (and hence event rate) is low, and only a small subset of patients is eligible for de-intensified treatment each year. As a result, the duration required to enroll sufficient patients and the subsequent follow-up can be prolonged to monitor study outcomes. Consequently, RCTs usually require extensive resources and can be infeasible to execute due to the complexity of design and rigorous regulatory requirements. Recently, there has been increased interest in utilizing non-randomized trial designs in the context of conditional drug approvals [4, 5]. Compared to traditional designs, various data-driven approaches like Bayesian statistical methods, single-arm studies, and adaptive designs are increasingly used in clinical trials [6]. These alternatives follow specific guidelines and regulations, offering faster patient access to innovative treatments than traditional RCTs. Bayesian methods integrate prior research and expert opinion to enhance accuracy and efficiency in trial design and analysis [7, 8]. Single-arm designs are preferable for conditions with limited patient populations and well-documented histories, although they require meticulous historical control comparison [9–13]. Adaptive trials allow for pre-planned modifications based on interim data, enhancing flexibility, and efficiency while maintaining scientific validity [14–16]. In practical applications, utilizing the Bayesian method for adaptive trials can significantly boost decision-making efficiency compared to traditional frequentist methods [17, 18]. Several studies indicate that this design needs smaller sample sizes and fewer events, leading to quicker conclusions in trials [19, 20]. For example, Bayesian adaptive trials were extensively used during the COVID-19 pandemic to evaluate effective treatments, leveraging the ongoing accumulation of data to improve treatment strategies [21]. In addition, the Bayesian method allows for the analysis of historical controls through extensive simulation studies and the formulation of prior knowledge via sensitivity analysis [22–24]. The Bayesian method provides great flexibility as an alternative approach for clinical trial design, which can be applied to both adaptive trials and single-arm designs. Despite its advantages, the effectiveness of this approach depends heavily on the selection of priors and likelihood models. In addition, the Bayesian algorithm can be computationally intensive, and existing algorithms may not cover the full spectrum of clinical research needs [25]. Therefore, there is a significant need to develop Bayesian adaptive trial algorithms tailored to specific research requirements. In this paper, we propose to use a Bayesian survival model (see Chapter 2 of [26]) in a single-arm adaptive trial. This method is particularly well-suited for the de-intensified designs in the context of infrequent outcome events, such as small patient populations with high cure rates and low event rates. Based on prior knowledge, the Bayesian survival model can effectively estimate hazard rates and predict survival probability for studies with low event rates. Moreover, this algorithm can be applied to generate pilot data and evaluate the potential efficacy or futility of a de-intensified treatment strategy before launching a randomized trial. The probabilistic results derived from Bayesian analysis provide robust, evidence-based scientific insights. To execute timely clinical trials, we have developed a Bayesian adaptive trial algorithm based on this survival model, which is capable of performing multiple-stage interim analyses. Unlike the two- or three-stage approach, we can implement Bayesian inference at each stage with updated interim data during both patient enrollment and early follow-up phases. Interim Bayesian inference can effectively utilize the posterior estimation to provide more flexible and prompt decisions for the sample size determination. With evidence of high efficacy during the interim analysis, we can conclude the effectiveness of the new treatment in a shorter time frame, allowing patients to reduce intensive treatment safely. In addition, we develop a computationally efficient algorithm for the Bayesian adaptive trial design, which is available as an R package called “BayesAT.” It offers various functions for users to conduct simulation studies, generate pilot data, and analyze real clinical data. The outline of the paper is as follows. The “Methodology” section introduces the Bayesian survival model and sample size determination methodologies, including the Bayesian interim analysis and the proposed algorithm. In the “Simulation” section, we implement the proposed algorithm across various simulation studies. The “Data applications” section discusses the application of the Bayesian adaptive trial to a Hodgkin Lymphoma trial, demonstrating how this approach can potentially lead to early conclusions about treatment efficacy. Methodology Suppose a clinical trial collects data Dn from n patients. The data Dn includes the follow-up time denoted by Tn=(t1,⋯,ti,⋯,tn) and the censoring times Cn=(c1,⋯,ci,⋯,cn) that is independent of Tn. In addition, the event of interest can be defined as the occurrence of death, disease progression, and relapse, which are binary outcomes Δn=(δ1,⋯,δi,⋯,δn). If the event was observed before the trial concluded (ti