Title: Bayesian Tools for Clinical Trials
Host: Department of Biomedical Informatics, University of Pittsburgh
Speaker: Daniel Normolle
Summary and thoughts:
Dr. Daniel Normolle from the Department of Biostatistics presented at the DBMI colloquium this week on the topic “Bayesian Tools for Clinical Trials”. Although I have worked with bayesian networks and models in other projects, I was completely unfamiliar with their application in clinical trials and this talk was a good way to learn about the field. Historically, clinical trials have been performed in 4 phases (as most of us are now aware due to the ongoing pandemic and search for a COVID vaccine). These phases assess the safety, efficacy, confirm safety and efficacy, and expanded labeling respectively. Each phase has straightforward guidelines to be followed and required sample sizes of population for testing. In some cases, such as rare diseases and life-threatening disorders, some of the guidelines are removed to give way to accelerated approval or approval after limited testing. However, most of the time, phase 1 of clinical trials must assess the safety of the treatment based on the number of individuals tested and generalize this result to a larger population.
Bayesian networks are probabilistic models that can be used to predict the outcome of clinical trials based on a number of parameters. Taking example of cancer therapy clinical trials, Phase 1 clinical trials generally need to evaluate the dose-level toxicity of the treatment as well as any adverse events that may occur (for example, radiation-induced liver disease). Methods to evaluate the dose-level toxicity in Phase 1 may include a common logistic model called continual reassessment method (CRM), TITE-CRM (i.e. time to event CRM), and hierarchical beta-gamma models. Bayesian methods (such as Bayesian CRM) can be used to extend these methods to predict the dose-level toxicity before the actual adverse events occur.