Abstract: The main goal of oncology phase I clinical trials is to identify a maximum tolerated dose (MTD). Although many authors have offered a variety of methods for the MTD, conventional methods have potential shortcomings. Parametric model-based methods may provide incorrect results due to model misspecification because little is known about dose-toxicity relationships and sample sizes are small. Furthermore, most methods ignore uncertainty in the estimated relationship during dose selection. To address these concerns, we propose a Bayesian optimization (BO) design for estimating the MTD. Our method utilizes a BO framework where the dose-toxicity relationship is nonparametrically estimated via Gaussian Processes, and dose selection balances exploration and exploitation. Simulation results suggest that our proposed method successfully estimates the MTD, offering a robust alternative to existing methods like the Continual Reassessment Method (CRM) and the modified toxicity probability interval (mTPI).
Introduction
Identifying the Maximum Tolerated Dose (MTD) is the primary objective of Phase I oncology trials. This task requires a delicate balance: a dose must be high enough to be therapeutic but low enough to avoid unacceptable toxicity. Traditionally, researchers have relied on the Continual Reassessment Method (CRM), which uses a parametric "skeleton" to predict the dose-toxicity curve. However, CRM faces two major challenges: model misspecification, where the true biological relationship does not fit the assumed mathematical shape, and the neglect of statistical uncertainty (variance) in small-scale clinical data.
While "curve-free" methods like mTPI exist, they often focus narrowly on the current dose being tested, failing to synthesize information across the entire dose range. This paper introduces a Bayesian Optimization framework to overcome these limitations by treating the dose-toxicity relationship as a nonparametric distribution.
Methods
The proposed Bayesian Optimization (BO) design replaces rigid parametric models with a flexible Gaussian Process (GP). This allows the dose-toxicity curve to adapt its shape based on incoming data without being constrained by a predefined functional form.
Key elements of the design include:
- Nonparametric Modeling: The GP uses a squared exponential kernel to define how toxicity at one dose level informs the probability at others.
- Acquisition Function: The model employs a Lower Confidence Bound (LCB) approach to manage the "exploration vs. exploitation" trade-off. This ensures the trial tests doses near the suspected MTD (exploitation) while also investigating doses with high uncertainty (exploration) to avoid local optima.
- Safety Constraints: To prioritize patient safety, the algorithm is restricted from skipping intermediate doses when escalating to higher levels.
Simulation Study and Results
The method was evaluated against industry standards—including CRM (power and logistic versions), mTPI, and the Product-of-Beta Prior (PBP) method—across two primary simulation settings:
- Simulation (a): A low-target toxicity trial ($\theta = 0.1$) with 6 dose levels and 30 patients.
- Simulation (b): A standard-target toxicity trial ($\theta = 0.3$) with 10 dose levels and 45 patients.
Performance in Low-Target Scenarios
In Simulation (a), the BO method demonstrated superior accuracy in identifying the correct MTD. For instance, in Scenario A1, the BO method achieved a correct selection rate of 75.7%, significantly outperforming CRM-p (64.6%) and mTPI (62.5%).
Performance in Standard-Target Scenarios
In Simulation (b), the BO method maintained high performance across a larger dose range. It consistently matched or exceeded the correct selection percentages of traditional methods while maintaining efficient patient allocation to the MTD. For example, in Scenario B4, the BO method reached an 87.4% correct selection rate.
Discussion and Conclusion
The Bayesian Optimization approach offers a robust, "curve-free" alternative for Phase I trials. By embracing uncertainty and updating the toxicity map for all doses simultaneously, the BO method avoids the pitfalls of model misspecification and local optima.
While the method showed slightly higher observed toxicity in some scenarios, this can be mitigated by adjusting the acquisition function parameters to be more conservative. Future research will focus on extending this framework to drug combination trials, where identifying MTD contours across multiple agents presents a more complex optimization challenge.
Summary updated