In randomized controlled trials, the homogeneity of treatment effect estimates in pre-defined subgroups based on clinical, laboratory, genetic, or other baseline variables is frequently investigated using forest plots. Estimation of subgroup-specific treatment effects is typically based on the data from the respective subgroup only. However, the interpretation of these naive subgroup-specific treatment effect estimates requires great care because of the smaller sample size of the subgroups (implying large variability of the estimated effects) and the frequently large number of investigated subgroups. Bayesian analyses for treatment effect estimation have been discussed in the literature. These methods frequently focus on disjoint subgroups, whereas subgroups for different variables in forest plots are typically overlapping. We propose a general strategy for treatment effect estimation in subgroups for survival outcomes. We first build a flexible model based on all available observations including categorical covariates identifying the relevant subgroups and their interactions with the treatment group variable. Interaction terms are penalized using lasso or ridge regression to shrink subgroup-specific estimates towards the population treatment effect. Alternatively, we put a Bayesian shrinkage prior (the horseshoe prior) on the interaction terms. An advantage of the Bayesian approach is that it is straightforward to derive credible intervals for subgroup-specific estimates. In a second step, this model is marginalized to obtain treatment effect estimates (hazard ratios) for all subgroups. The non-collapsibility of the hazard ratio complicates marginalization and leads to marginalized survival curves for the treatment groups that are not proportional. To deal with these non-proportional survival curves we use the average hazard ratio corresponding to the odds-of-concordance to quantify the treatment effect.

The methods are illustrated with data from a large randomized clinical trial in follicular lymphoma and compared in an extensive simulation study. For all simulation scenarios, the overall mean-squared error (MSE) of all methods is drastically improved compared to naive subgroup-specific treatment effect estimates. The method based on the horseshoe prior performs slightly better in terms of bias, MSE and frequentist coverage of 95% credible intervals, compared to the other methods, in scenarios where only one of the subgrouping variables is associated with treatment effect heterogeneity. We are currently implementing all these methods in an R package which we plan to upload to CRAN.

Oncology clinical development benefits from well-defined and widely accepted standards to evaluate anti-tumor drug activity. Indeed, the RECIST1.1 guideline offers a framework to harmonize the interpretation of changes in tumor burden for patients with solid tumors. To this aim, it brings various information together, namely radio imaging data on target and non-target lesions, as well as data on the emergence of new lesions, or clinical judgment. For target lesions, the sum of diameters (SLD) is calculated in every patient at regular time intervals during the trial. For active drugs and responder patients, the expectation is for SLD to decrease over time (tumor shrinkage); after a while, SLD is typically growing due to the loss of treatment effect. For inactive drugs or non-responder patients, SLD is monotonically increasing over time (tumor growth). The variability in individual longitudinal profiles can be studied using tumor growth inhibition (TGI) non-linear mixed-effect models.

Progression-free survival (PFS) is commonly used as a primary or secondary endpoint in oncology clinical trials. This right-censored variable is defined as the time to the earliest of either death (from any cause, OS) or disease progression (PD, per RECIST1.1). As such PFS is mixing events of different natures, occurring in sequence with PD preceding death. Instead of mingling them, we consider a joint model including (i) a TGI sub-model, (ii) a time-to-event OS sub-model, and (iii) a term measuring the strength and nature of the TGI-OS association. With this model, it becomes possible to investigate the effect of an intervention on each component of the bivariate process and to quantify the proportion of treatment effect on overall survival mediated by changes in SLD. We present an example showing the proportion of treatment effect mediated by SLD in urothelial cancer patients treated with atezolizumab. The model can also be used to predict OS based on early measures of SLD. Based on simulations, we illustrate how SLD can be used as a proxy for OS to support early decision-making. Extensions of this work are then discussed, including the potential role of non-target and new lesions in predicting OS.

The development of new drugs is often time-consuming and expensive – especially in oncology. Hence, there is a need to speed up the development process. Programs like the FDA’s Critical Path Initiative in general and the Oncology Center of Excellence Project Optimus specifically focus on optimizing the dose selection process in oncology.

Unlike in other therapeutic areas, there is no formal dose finding in the sense of establishing a dose-response relationship and selecting an optimal dose. Typically, a dose (MTD /RP2D) is identified in Phase 1 and then carried forward to Phase 2 in which data on a binary clinical outcome like RECIST-based objective response is collected in a single-arm trial. If the response rate achieves some pre-defined criteria, the drug enters a Phase 3. One main question is whether other tumor measurements can be used in Phase 2 trials to better characterize the dose-response relationship. Tumor growth models which describe the change of the tumor burden over time in response to treatment using exponential models could provide alternative measures, as for example the g(rowth)-rate or the d(ecline)-parameter.

We investigate the mathematical properties of exponential tumor growth models and derive several equations and algorithms linking the g- and d-parameters to other tumor measures like response and progression as well as time-to-response, time-to-progression, and duration of response. The mathematical framework allows us to specify constrains like desired response rate, follow-up-time, and median time-to-response yielding unique solutions for the mean of the logarithm of the g- and d-parameter. Based on this, the framework can be used to jointly simulate response and time-to-event endpoints in oncology. In addition, it can be used to investigate the advantages and disadvantages of using the g- and d-parameter instead of the response rate for establishing a dose-response relationship in Phase 2.

Dose finding trials in Oncology have traditionally been based on sequential dosing cohorts where escalation or de-escalation is guided by the incidence short term dose limiting toxicities. However in the recent past there has been greater push towards supplementing such sequential dose finding designs by randomized dose cohorts evaluating toxicity, tolerability and key efficacy endpoints in order to better optimize the dose prior to starting the Ph 3 pivotal study

Progression free survival (PFS) is a common clinical endpoint in oncology studies to assess the clinical benefit . When used in such randomized dose optimization studies, PFS data can guide the selection of the dose It is typically analyzed per the intention to treat principle. However, a challenge here would be to compare different doses because of the intercurrent event of dose reduction. As the patients in the higher dose groups reduce their doses during the course of the study it makes the cohorts less distinct. Any comparison between the doses, the treatment effect would be underestimated per ITT principle.

In this talk we present two case studies one in pre-market and other in post market dose finding to explore the implications of such dose reductions. We propose to use the estimand framework to assess the role of such dose reductions as an intercurrent event when assessing the PFS difference and present different strategies to handle it. We have considered three estimators multistate survival model, time dependent survival, and G-estimation to estimate the treatment benefit between different dose groups. The performance of these methods will be evaluated by different simulation scenarios and a real clinical study.

CONsolidated Standards Of Reporting Trials (CONSORT) 2010 provides guidance on reporting completed parallel group randomised trials. Using the Enhancing the QUAlity and Transparency Of health Research (EQUATOR) methodological framework for guideline development, CONSORT 2010 has recently been extended to address the special features of early phase dose-finding trials [1]. Such trials have been found to be poorly reported with a detrimental impact on their informativeness and the possibility to make evidence-informed decisions [2]. The resulting consensus-driven CONSORT-DEFINE (CONSORT – DosE-FIndiNg Extension) statement recommends essential items to be included in completed early phase dose-finding trial reports to promote greater transparency, utility of results, and reproducibility. These highly adaptive trials are usually employed as phase I or seamless phase I/II trials and, using dose escalation/de-escalation strategies, they aim to recommend a dosing regimen or a range of dosing regimens for subsequent later phase trials based on safety and other information such as pharmacokinetics, pharmacodynamics, biomarker activity, and clinical activity. They may study any intervention that can be given in different dosages (doses and/or schedules), e.g., drugs, vaccines, cell therapies, gene therapies, digital therapeutics, rehabilitation, or radiotherapy, and may either involve healthy volunteers or people with a condition of interest.

In this presentation, we will emphasize the importance of a widespread implementation of this CONSORT extension, and we will focus on the statistical aspects of new items in CONSORT-DEFINE and items modified from CONSORT 2010 that cover trial design, statistical methods, and analysis. Such items comprise, among others, reporting details on underlying statistical methods for dose escalation/de-escalation strategies and decision-making criteria as well as reporting key outcomes by dosing regimen. We will highlight the importance of including specific items, discuss good examples, and provide practical advice on clear reporting. We hope that CONSORT-DEFINE will ultimately improve participant safety and benefits in early phase dose-finding trials and contribute to transparent reporting while sparing research resources.

Acknowledgement: CONSORT-DEFINE Group and CONSORT-DEFINE Example Guidance Working Group.

[1] Yap, C., et al., The need for reporting guidelines for early phase dose-finding trials: Dose-Finding CONSORT Extension. Nature Medicine, 2022. 28(1): p. 6-7.

[2] Yap, C., et al., Assessing the reporting quality of early phase dose-finding trials. Annals of Oncology, 2022. 33: p. S24-S24.