ABSTRACT
Phase I designs for finding optimal doses within ordinal prognostic subgroups have been proposed by J. O'Quigley and X. Paoletti; Y. Yuan and R. Chappell; A. Ivanova and K. Wang. The efficacy-toxicity design with covariates (EffToxCovs) design is useful for trials with binary Toxicity and Efficacy and a small number of well-established patient baseline prognostic covariates. The biomarker-based personalized dose finding (PDF) design is compared to modified version of the EffToxCovs design that includes dose, biomarkers, and dose-biomarker interactions in the regression model. The PDF design outperformed the modified EffToxCovs design. To deal with the curse of dimensionality and potential high correlations among biomarkers, the design employs the canonical partial least squares (CPLS) method to extract a small number of components from the covariate matrix consisting of dose, biomarker profile, and dose by biomarker interactions. These CPLS components capture most of covariate information that is predictive of response and will be used to model relationships between dose/biomarker and Toxicity/Efficacy outcomes.
