Investigation of phase I dose finding methods: Bayesian methodology for phase I studies

Abstract

Clinical trial is a scientific study designed to examine whether new potential treatments are safe and effective. Despite its name, a clinical trial is not conducted in a laboratory, in fact, the clinical trial experience is very similar to a regular doctor visits. Healthy volunteers or patients with the illnesses that are being studied participate in the trial. The aim is to gather enough evidence to understand if a medicine works and it is safe. Participating in a clinical trial means being part of the advancement of health care and science. During the clinical trial, the participants often have the opportunity to access promising new treatments that may be more effective than the current standard of care. Safety of the potential treatment is the first determined treatment during phase I trials, but it is continuously monitored through all phases. This thesis reviews the most popular and used phase I dose response methods, and explores key limitations of these methods, and introduces a comparative simulation study that has different model structures and prior distributions in the Continual Reassessment Method (CRM). The common phase I methods are the 3+3 design, A+B design, Continual Reassessment Method (CRM), Bayesian Model Averaging Continual Reassessment Method (BMA-CRM), Bayesian Optimal Interval Designs (BOIN), Modified Toxicity Probability Interval Method (mTPI ) and a Bayesian Interval Dose-Finding Design Addressing Ockham's Razor (mTPI -2). These methods are used in the clinical trials to select a true maximum tolerated dose (MTD). In the first part of the thesis, these methods were compared to two different stories and twelve different scenarios. After examining and comparison of each scenario, which methods were more effective and efficient in selecting the correct MTD was concluded. According to the results, CRM, BMA-CRM, mTPI and mTPI-2 were the best performing methods in our simulation runs. In the second part of the thesis, a different story and eight scenarios are implemented. The prominent methods in the first part were compared with the CRM method where the model structure and prior distribution were different. Overall, in designs where the model structure is hyperbolic tangent and prior distribution is uniform, the CRM calculated the selection probability of the correct MTD higher than the others. On the other hand, in designs where the model structure is logit and prior distribution is lognormal, the CRM calculated the selection probability of the correct MTD lower than other CRM designs. In addition, the BMA-CRM produced very effective results if the difference between the correct MTD dose and the previous-subsequent dose is greater. Moreover, the mTPI and mTPI-2 designs can produce better results in the case of where the target toxicity of the trial is not included in the study. In conclusion, more reliable and applicable results for phase I dose finding trials are produced by the BMA-CRM and CRM, when the model structure and prior distributions are different, in our study. As a result, the model-based designs performed much better than the rule-based designs.