A benchmark for dose-finding studies with unknown ordering

A comparison of model-free phase I dose escalation designs for dual-agent combination therapies

It is increasingly common for therapies in oncology to be given in combination. In some cases, patients can benefit from the interaction between two drugs, although often at the risk of higher toxicity. A large number of designs to conduct phase I trials in this setting are available, where the objective is to select the maximum tolerated dose combination. Recently, a number of model-free (also called model-assisted) designs have provoked interest, providing several practical advantages over the more conventional approaches of rule-based or model-based designs. In this paper, we demonstrate a novel calibration procedure for model-free designs to determine their most desirable parameters. Under the calibration procedure, we compare the behaviour of model-free designs to model-based designs in a comprehensive simulation study, covering a number of clinically plausible scenarios. It is found that model-free designs are competitive with the model-based designs in terms of the proportion of correct selections of the maximum tolerated dose combination. However, there are a number of scenarios in which model-free designs offer a safer alternative. This is also illustrated in the application of the designs to a case study using data from a phase I oncology trial.

On the role of benchmarking data sets and simulations in method comparison studies

Method comparisons are essential to provide recommendations and guidance for applied researchers, who often have to choose from a plethora of available approaches. While many comparisons exist in the literature, these are often not neutral but favor a novel method. Apart from the choice of design and a proper reporting of the findings, there are different approaches concerning the underlying data for such method comparison studies. Most manuscripts on statistical methodology rely on simulation studies and provide a single real-world data set as an example to motivate and illustrate the methodology investigated. In the context of supervised learning, in contrast, methods are often evaluated using so-called benchmarking data sets, that is, real-world data that serve as gold standard in the community. Simulation studies, on the other hand, are much less common in this context. The aim of this paper is to investigate differences and similarities between these approaches, to discuss their advantages and disadvantages, and ultimately to develop new approaches to the evaluation of methods picking the best of both worlds. To this aim, we borrow ideas from different contexts such as mixed methods research and Clinical Scenario Evaluation.

Quasi-partial order continual reassessment method: Applying toxicity scores to cancer dose-finding drug combination trials

The primary endpoint of most dose-finding cancer trials is patient toxicity, and the primary goal is to identify the maximum tolerated dose (MTD), that is, the highest dose that falls below or within a pre-specified toxicity tolerability threshold. Conventionally, dose-finding methods have utilized a binary toxicity endpoint based on whether or not a patient experiences a dose limiting-toxicity (DLT). Improving upon this, in recent years several methods have been developed for modeling toxicity scores, a novel continuous endpoint designed to more precisely estimate patient toxicity burden. Separately, drug-combination trials have become increasingly prevalent, and due to added complexities regarding estimating 'true' dose ordering and potential for more complex patient toxicity profiles, provide an ideal setting which may benefit from the improved precision of toxicity scores. In this paper, we merge two frameworks based on the Continual Reassessment Method (CRM) - the Quasi-CRM and the Partial Order CRM (POCRM) - to propose a novel approach for modeling toxicity scores in a combination-trial setting. We demonstrate that utilizing toxicity scores has the potential to greatly improve correct dose-selection over a variety of trial scenarios. We further present a simple adaptation to the toxicity-score model to control for potential over-dosing issues such that it adheres to the conventional DLT definition and will, at worst, perform equivalently to that of the traditional binary DLT framework. We demonstrate that extending toxicity scores to the combination-trial setting offers potential for improvement over the conventional binary endpoint models.

Practical implementation of the partial ordering continual reassessment method in a Phase I combination-schedule dose-finding trial

There is a growing medical interest in combining several agents and optimizing their dosing schedules in a single trial in order to optimize the treatment for patients. Evaluating at doses of several drugs and their scheduling in a single Phase I trial simultaneously possess a number of statistical challenges, and specialized methods to tackle these have been proposed in the literature. However, the uptake of these methods is slow and implementation examples of such advanced methods are still sparse to date. In this work, we share our experience of proposing a model-based partial ordering continual reassessment method (POCRM) design for three-dimensional dose-finding in an oncology trial. In the trial, doses of two agents and the dosing schedule of one of them can be escalated/de-escalated. We provide a step-by-step summary on how the POCRM design was implemented and communicated to the trial team. We proposed an approach to specify toxicity orderings and their a-priori probabilities, and developed a number of visualization tools to communicate the statistical properties of the design. The design evaluation included both a comprehensive simulation study and considerations of the individual trial behavior. The study is now enrolling patients. We hope that sharing our experience of the successful implementation of an advanced design in practice that went through evaluations of several health authorities will facilitate a better uptake of more efficient methods in practice.

Flexible use of copula-type model for dose-finding in drug combination clinical trials

Identification of the maximum tolerated dose combination (MTDC) of cancer drugs is an important objective in phase I oncology trials. Numerous dose-finding designs for drug combination have been proposed over the years. Copula-type models exhibit distinctive advantages in this task over other models used in existing competitive designs. For example, their application enables the consideration of dose-limiting toxicities attributable to one of two agents. However, if a particular combination therapy demonstrates extremely synergistic toxicity, copula-type models are liable to induce biases in toxicity probability estimators due to the associated Fréchet-Hoeffding bounds. Consequently, the dose-finding performance may be worse than those of other competitive designs. The objective of this study is to improve the performance of dose-finding designs based on copula-type models while maintaining their advantageous properties. We propose an extension of the parameter space of the interaction term in copula-type models. This releases the Fréchet-Hoeffding bounds, making the estimation of toxicity probabilities more flexible. Numerical examples in various scenarios demonstrate that the performance (e.g., the percentage of correct MTDC selection) of the proposed method is better than those exhibited by existing copula-type models and comparable with those of other competitive designs, irrespective of the existence of extreme synergistic toxicity. The results obtained in this study could motivate the real-world application of the proposed method in cases requiring the utilization of the properties of copula-type models.

Practical recommendations for implementing a Bayesian adaptive phase I design during a pandemic

BACKGROUND

Modern designs for dose-finding studies (e.g., model-based designs such as continual reassessment method) have been shown to substantially improve the ability to determine a suitable dose for efficacy testing when compared to traditional designs such as the 3 + 3 design. However, implementing such designs requires time and specialist knowledge.

METHODS

We present a practical approach to developing a model-based design to help support uptake of these methods; in particular, we lay out how to derive the necessary parameters and who should input, and when, to these decisions. Designing a model-based, dose-finding trial is demonstrated using a treatment within the AGILE platform trial, a phase I/II adaptive design for novel COVID-19 treatments.

RESULTS

We present discussion of the practical delivery of AGILE, covering what information was found to support principled decision making by the Safety Review Committee, and what could be contained within a statistical analysis plan. We also discuss additional challenges we encountered in the study and discuss more generally what (unplanned) adaptations may be acceptable (or not) in studies using model-based designs.

CONCLUSIONS

This example demonstrates both how to design and deliver an adaptive dose-finding trial in order to support uptake of these methods.

Operating characteristics are needed to properly evaluate the scientific validity of phase I protocols

PURPOSE

Operating characteristics for proposed clinical trial designs provide insight into performance regarding safety and accuracy, allowing the study team and review entities to determine the design's suitability to achieve the study's proposed objectives. Advances in cancer therapeutics have augmented the needs of early phase clinical trial design. Additionally, advances in research on early-phase trial design have led to the availability of a wide range of methods that show vast improvement over outdated approaches.

METHODS

Three trials utilizing variations of the 3 + 3 decision rule are discussed. The protocols lacked detail, including operating characteristics and guidance for decision-making that deviated from the 3 + 3 decision rule and MTD determination. We provide a discussion of the statistical issues associated with each design and operating characteristics for the proposed design compared to alternatives better suited to achieve the aims of each trial.

RESULTS

Our results illustrate how operating characteristics inform a design's safety and accuracy. Operating characteristics can unmask poor behavior, such as a high percentage of particiapnts exposed to overly toxic doses, a low probability of correctly identifying the MTD, and inappropriate early study termination.

CONCLUSION

Selection of early-phase trial design has significant implications on a trial's ability to meet its objectives. Operating characteristics are a necessary component in the design and review of a protocol, determining if the study's objectives can be achieved and documenting the study's scientific validity. Continued use of outdated approaches due to historical acceptance hinders scientific rigor and the effort to move effective agents through the drug development process.

Using an Interaction Parameter in Model-Based Phase I Trials for Combination Treatments? A Simulation Study

There is growing interest in Phase I dose-finding studies studying several doses of more than one agent simultaneously. A number of combination dose-finding designs were recently proposed to guide escalation/de-escalation decisions during the trials. The majority of these proposals are model-based: a parametric combination-toxicity relationship is fitted as data accumulates. Various parameter shapes were considered but the unifying theme for many of these is that typically between 4 and 6 parameters are to be estimated. While more parameters allow for more flexible modelling of the combination-toxicity relationship, this is a challenging estimation problem given the typically small sample size in Phase I trials of between 20 and 60 patients. These concerns gave raise to an ongoing debate whether including more parameters into combination-toxicity model leads to more accurate combination selection. In this work, we extensively study two variants of a 4-parameter logistic model with reduced number of parameters to investigate the effect of modelling assumptions. A framework to calibrate the prior distributions for a given parametric model is proposed to allow for fair comparisons. Via a comprehensive simulation study, we have found that the inclusion of the interaction parameter between two compounds does not provide any benefit in terms of the accuracy of selection, on average, but is found to result in fewer patients allocated to the target combination during the trial.