C. Anthony Hunt, Serge Guzy and Daniel L. Weiner
The University of California, San Francisco, and Quintiles Inc.,Research Triangle Park.
There is increasing interest in methodologies that can simulate clinical trial results. We present a paradigm for validating such methodologies, and we apply it to the Forecaster approach discussed at http://www.mis.ucsf.edu/decision/pub/forecast/. The final, pivotal task is to use existing early clinical trial data to predict later existing data for a specific drug. First, however, it is necessary to validate that the system and its algorithms functions as designed. The evaluation process requires a structured, known environment where options can be controlled, the robustness of the framework can be tested, and assumptions evaluated (e.g., see Stat. Med. 15: 361-387, 1996). Such a validation provides a context for judgment when one moves to the clinical trial setting. First, we generated a true or reference population density of PK/PD parameter sets based on the reported results for single dose Medazolam (Clin. Pharmacol. Ther. 45: 356-64, 1989). A two-step process gave 1500 sets of simulated individual experimental data where the individual drug level and response data exhibited realistic variability. The PK and then the PD data were each fit to a model (WinNonlin). A random sample of N = 30 fitted parameter sets (eight parameters per individual) was used as input for the Forecaster. We aimed to predict the fraction of a population that is expected to have an effect within the therapeutic range at specified times during a fixed dose regimen. To do so, the Forecaster generates a Surrogate Population. We bracketed each prediction with a “confidence interval” such that, under ideal conditions, the probability is (say) approximately 85% that the actual population target value is within the “confidence interval.” To confirm and validate the result we repeated the above process 300 times. We thus document that the Forecaster can be used to make useful PD predictions, and that the approximate prediction interval actually captures the population target value approximately 85% of the time. The results clearly support the feasibility of generating and then using a Surrogate Population as the core database for a drug development decision support system.
Reference: PAGE 6 () Abstr 596 [www.page-meeting.org/?abstract=596]
Poster: oral presentation