Zhilong Yuan1, Rick Chappell. 1. Department of Statistics, University of Wisconsin, Madison, WI, USA. zhilong@stat.wisc.edu
Abstract
BACKGROUND: In phase I cancer clinical trials, adjustment for patient differences in toxicity susceptibility can be carried out with stratification into risk groups. Separate trials conducted for each risk group can lead to conflicting decisions, in which higher doses are recommended for higher risk groups. Designs which covariate adjust often require assumptions that clinicians may be uncomfortable with. METHODS: We extend up-and-down designs, isotonic designs and the continual reassessment method (CRM) to multiple risk groups with two-way isotonic regression. The only assumption about the groups is that they can be ordered according to their toxicity risk. The first two extensions, in particular, are nonparametric and easy for clinicians to understand. RESULTS: Simulations were based on an ongoing helical tomotherapy trial. Seven different toxicity scenarios were considered. The proposed methods compared favorably to a covariate adjusted CRM. The extended up-and-down designs inherited the conservativeness from the original designs. CONCLUSION: Our experience demonstrates that the escalation rules of multiple risk groups can be linked, without a parametric assumption about the group toxicity curve, to borrow strength and to ensure nonconflicting dosage recommendations.
BACKGROUND: In phase I cancer clinical trials, adjustment for patient differences in toxicity susceptibility can be carried out with stratification into risk groups. Separate trials conducted for each risk group can lead to conflicting decisions, in which higher doses are recommended for higher risk groups. Designs which covariate adjust often require assumptions that clinicians may be uncomfortable with. METHODS: We extend up-and-down designs, isotonic designs and the continual reassessment method (CRM) to multiple risk groups with two-way isotonic regression. The only assumption about the groups is that they can be ordered according to their toxicity risk. The first two extensions, in particular, are nonparametric and easy for clinicians to understand. RESULTS: Simulations were based on an ongoing helical tomotherapy trial. Seven different toxicity scenarios were considered. The proposed methods compared favorably to a covariate adjusted CRM. The extended up-and-down designs inherited the conservativeness from the original designs. CONCLUSION: Our experience demonstrates that the escalation rules of multiple risk groups can be linked, without a parametric assumption about the group toxicity curve, to borrow strength and to ensure nonconflicting dosage recommendations.