Results obtained with a given regimen in relapsed or refractory acute myelogenous leukemia (AML) are variable. This often reflects variability in patient selection. We have developed a system to account for such variability that stratifies patients with refractory or relapsed AML into four groups: group 1, first complete response (CR) duration > or = 2 years and receiving first salvage treatment (S1); group 2, first CR duration 1-2 years and receiving S1; group 3, first CR duration 0-1 years and receiving S1; and group 4, first CR duration 0-1 years and receiving S2, S3, or S4 after failing S1. CR rates achieved in the four groups are 73%, 47%, 14%, and 0, respectively. This system is useful for comparing results obtained with different therapies and for assigning patients to treatment. At our institution, patients in group 4 are enrolled in phase I studies, and phase I1/2-II studies are carried out separately in patients in groups 2, 3, and 4. A phase I1/2 study refers to one in which the intent is to select for phase II therapies emerging from phase I trial results. The design is Bayesian, and although false-negative rates are relatively high, they are lower than those obtained if a drug for phase II testing is arbitrarily selected.