Selecting therapeutic strategies based on efficacy and death in multicourse clinical trials

Therapy of rapidly fatal diseases often requires multiple courses of treatment. In each course, the treatment may achieve the desired clinical goal (''response''), the patient may survive without response (''failure''), or the patient may die. When treatment fails in a given course, it is common medical practice to switch to a different treatment for the next course. Most statistical approaches to such settings simply ignore the multicourse structure. They characterize patient outcome as a single binary variable, combine death and failure, and identify only one treatment for each patient. Such approaches waste important information. We provide a statistical framework, including a family of generalized logistic regression models and an approximate Bayesian method, that incorporates historical data while accommodating multiple treatment courses, a trinary outcome in each course, and patient prognostic covariates. The framework serves as a basis for data analysis, treatment evaluation, and clinical trial design. In contrast with the usual approach of evaluating individual treatments, our methodology evaluates outcome-adaptive, multicourse treatment strategies that specify, within prognostic subgroups, which treatment to give in each course, We describe a general approach for constructing clinical trial designs that may be tailored to different multicourse settings. For each prognostic subgroup, based on a real-valued function of the covariate-adjusted probabilities of response and death, the design drops inferior treatment strategies during the trial and selects the best strategy at the end. The methodology is illustrated in the context of designing a randomized two-course, three-treatment acute leukemia trial with two prognostic covariates. To validate the model and develop a prior, we first fit the model to a historical dataset, We describe a simulation study of the design under several clinical scenarios. The simulations show that the method can reliably identify treatment-subgroup interactions based on moderate sample sizes.