Adaptive designs for dose escalation early phase studies

A key objective of Phase I dose escalation trials is to establish a recommended dose and/or schedule for an experimental drug or combination for further testing in Phase II.  They balance avoiding unnecessary exposure of patients to subtherapeutic doses with maintaining safety and rapid enrolment. 

Traditional “3+3” designs and their variations are simple to implement, based upon pre-specified rules to guide how to react to accumulating data but lack firm statistical foundation.  These types of designs only “memorise” the toxicities in the previous step of dose-escalation and do not take into consideration cumulative toxicities. As a result, a large proportion of patients may receive low, non-effective doses and the estimated maximum tolerated dose (MTD), the highest dose at which a pre-specified proportion of patients experience a dose-limiting toxicity (DLT), may be inaccurate.  To overcome the uncertainty about a recommended dose for future studies and slow dose escalation, model-based Bayesian adaptive designs have been developed:  

  • The target toxicity at the recommended dose for future studies would be explicitly defined by seeking a suitable quantile of the dose-response curve, such as the probability of toxicity at the target dose being 20%;
  • The prior guess of the target dose on the pre-assumed dose-response curve before the trial starts would form the starting dose;
  • The decision criteria, such as treating future patients at the current estimated MTD or minimising the variance of the dose-response model parameters, would then be used to identify the next dose after the model is updated to take consideration of all available data from the trial.
  • A set of safety constraints can be put in place to control the risk of overdosing the patients, for example pre-specifying a maximum dose increase.

An example of a model-based design in oncology trials is the Continuous Reassessment Method (CRM) (O’Quigley et al 1990). The CRM is based on a single parameter logistic regression dose-response model with an exponential prior.  The clinical team’s knowledge of the pre-clinical data for the drug, or clinical data with similar drugs informs the initial estimate of the slope of the dose-response curve and starting dose.  The decision for the next dose is based on a patient gain - the posterior distribution of the model parameters is updated after observing whether each patient has a DLT.  The dose for the next patient, or next cohort of patients, is chosen as the one with probability of toxicity closest to the target response.  In clinical practice the CRM is often modified to address escalating doses too rapidly: initial patients may be treated at the lowest dose level based upon traditional criteria, only increasing the dose by one pre-specified level at a time and treating cohorts of patients at the same dose.

Extensions to the CRM model have been used, for example two parameters instead of the CRM’s one parameter model, so that the model will be more flexible in capturing the real dose-response curve.    A modified version of the CRM is the estimation with overdose control (EWOC) method.  This restricts the exposure of patients to likely toxic doses if the probability from the model of overdosing exceeds a pre-specified value e.g. 25%.  In this way EWOC identifies the MTD by the end of the study while formally limiting the number of patients exposed to toxic doses during the study. A further extension is using a mixed model rather than logistic regression so that dose-escalation can be implemented based on continuous responses, for example PK parameters.

The time-to-event CRM (TITECRM) incorporates time-to-event pattern of DLTs into the CRM model.  It weights each patient’s contribution to the model by the proportion of the time-period for observation of a DLT which each patient has completed.  For example, a patient who has completed half the follow-up period for DLT would contribute half the weight compared to a patient who has completed follow-up.  These weights can be varied, for example if DLTs are more likely to occur soon after the start of dosing then more weight can be given to the first weeks of observation.

There are now many model-based designs for dose finding which balance good statistical models and making use of all available data with flexibility of to handle practical considerations.  Setting up these designs does require time and expertise, but software to simulate trials and make decisions on an ongoing basis during a trial are readily available. Statisticians must partner with clinical teams at the design stage to choose the most appropriate design from the many designs on offer. Tailoring the trial design to the unique characteristic of the drug, combination, prior knowledge, and therapeutic indication should involve assessing whether the advantages of a model-based adaptive design may be beneficial for a trial.  In summary model-based designs:

  • Provide a better estimate of the MTD.
  • Allow more rapid progression through early dosing levels.
  • Offer flexibility – different cohort sizes can be used, different levels of toxicity targeted, and different stopping rules applied.
  • Overcome practical hurdles, for example limiting dose escalation steps.
  • Enable allocation to doses based both upon tolerability and efficacy and can be based upon binary or continuous responses.