Abstract
The most commonly used method for dose-finding, the 3+3, has poor performances. New adaptive designs are more efficient. Nevertheless, they have reached a maximum performance level, and further improvement requires either larger sample sizes, or outcomes measures richer than the simplistic severe toxicity measured at cycle 1. Body In this issue of Clinical Cancer Research, Yan et al. (1) propose a new adaptive dose-finding method, termed Keyboard as an alternative to the 3+3 design. The performance of the most commonly used dose-finding method, the 3+3, is disappointing: The chance of finding the correct dose in a phase I trial is no more than 40% (1). How many active, potentially promising agents are dropped due to incorrect dose selection during development? Few researches address this question (2), but stakeholders acknowledge the limitations of early phase trials (3). The statistical community unanimously agrees that the 3+3 is not an efficient method, raising ethical concerns according to NIH guidelines (4) and alternatives should be implemented. Keyboard combines simplicity, performance and flexibility. Simplicity for two reasons: (i) it is based on a natural definition of the maximum tolerated dose (MTD), i.e., the dose at which the risk of dose-limiting toxicity (DLT) is within a predefined range (typically 20–35%), and (ii) the decision rules to (de)escalate are driven by the accumulated observations, and the observed proportion of DLT at a given dose level. The more patients enrolled at a specific dose, the more confident we can be that this dose is (close to) the MTD, and the more likely the same dose is recommended for the next patients.