#### Event Title

### Session 3F: The Development of Bayesian Randomized Rules in Phase I Dose-Finding Clinical Trials

#### Session Number

Session 3F: 4th Presentation

#### Advisor(s)

Yuan Ji, University of Chicago

#### Location

Room A115

#### Start Date

28-4-2017 1:15 PM

#### End Date

28-4-2017 2:30 PM

#### Abstract

Phase I dose-finding clinical trials are utilized to optimize patient benefits at a specific dosage level of a drug. In the Toxicity Probability Interval (TPI) design, if the central tendency of a clinical trial’s beta-distribution falls in a safe interval, physicians will Escalate (“E”) their dosage. If the central tendency falls in an equivalence interval, physicians will Stay (“S) at the same dosage. If the central tendency falls in a dangerous interval, physicians will De-Escalate (“D”) their dosage. To improve upon TPI and put fewer patients on toxic dosages, a calibration-free modified TPI design (mTPI) was developed based on the Unit Probability Mass (UPM), the ratio of probability of falling within a specific TPI to the length of the TPI. However, using the Bayes Factor in the mTPI-2 design, some uncertainty behind each dose decision was observed. In this study, we created a formula and simulation to calculate the Bayes Factor of each decision. If the Bayes Factor is close to 1, then the evidence supporting one course of action over another is weak, which means that we can turn towards randomized rules to mitigate the noisy and uncertain evidence. This method will likely improve the mTPI-2 design as it takes into account the uncertainty in decisions by allowing randomization instead of optimization. Ultimately, this methodology demonstrates the potential utility of randomization in a traditionally non-randomized Phase I dose-finding clinical trial.

Session 3F: The Development of Bayesian Randomized Rules in Phase I Dose-Finding Clinical Trials

Room A115

Phase I dose-finding clinical trials are utilized to optimize patient benefits at a specific dosage level of a drug. In the Toxicity Probability Interval (TPI) design, if the central tendency of a clinical trial’s beta-distribution falls in a safe interval, physicians will Escalate (“E”) their dosage. If the central tendency falls in an equivalence interval, physicians will Stay (“S) at the same dosage. If the central tendency falls in a dangerous interval, physicians will De-Escalate (“D”) their dosage. To improve upon TPI and put fewer patients on toxic dosages, a calibration-free modified TPI design (mTPI) was developed based on the Unit Probability Mass (UPM), the ratio of probability of falling within a specific TPI to the length of the TPI. However, using the Bayes Factor in the mTPI-2 design, some uncertainty behind each dose decision was observed. In this study, we created a formula and simulation to calculate the Bayes Factor of each decision. If the Bayes Factor is close to 1, then the evidence supporting one course of action over another is weak, which means that we can turn towards randomized rules to mitigate the noisy and uncertain evidence. This method will likely improve the mTPI-2 design as it takes into account the uncertainty in decisions by allowing randomization instead of optimization. Ultimately, this methodology demonstrates the potential utility of randomization in a traditionally non-randomized Phase I dose-finding clinical trial.