Associate Director Fractal Therapeutics, Arizona, United States
Objectives: In this study, we have examined the accuracy of a range of estimation methods for finding the maximum tolerated dose (MTD) in a clinical trial. A popular approach is to use 3+3 dose escalation, where three patients at a time are dosed in each cohort and the MTD occurs when at least two patients experience dose limiting toxicity (DLT). In practice, toxicities are often recorded on a graded scale, MTD estimation often requires dichotomizing this graded data to a binary measure of dose limiting toxicity (DLT). Estimation algorithms that retain graded data and estimate MTD by modeling patient response heterogeneity may reduce information loss and improve MTD estimation accuracy, but these strategies have yet to be comprehensively tested. Here we put forward a novel approach for MTD estimation, the Population Response Estimate (PRE), and compare its performance to other approaches such as ordinal logistic regression, logistic regression, and 3+3.
Methods: Our approach, PRE, applies Maximum Likelihood Theory to graded toxicity data to estimate the toxicity central tendency curve and population heterogeneity, which in turn is used to construct a probability model for patient DLT. We evaluated the performance of MTD estimation algorithms by simulating synthetic clinical toxicity data. This was done by randomly generating patient toxicity samples from an assumed-true underlying toxicity relationship relating drug dose to exposure, and drug exposure to patient toxicity (based on a first-order Hill function describing the central tendency combined with a population response heterogeneity/probability density modeled as a Gaussian distribution). The true MTD (MTDTRUE) of the is defined at the drug exposure that causes 25% of the patient population to experience DLT (toxicity of grade 3 or higher). We evaluated each MTD estimation approach for its ability to estimate MTDTRUE on the same dataset, containing 5000 simulated curves for each toxicity function examined. We used both 3+3 dose escalation as well as an alternative paradigm (adaptive dose escalation) to evaluate each method.
Results: In this dataset, we found that PRE performs best in estimating MTDTRUE in both dose-ranging paradigms and all true toxicity relationships examined, providing the best estimate for MTDTRUE in 41% of the cases (vs 22% for 3+3, 22% for ordinal logistic regression and 18% for binary logistic regression) . We found that all tested MTE estimation approaches, including PRE, tend to strongly underestimate MTE for asymptotic exposure-toxicity relationships when 3+3 dose escalation is used. The use of the alternative dose-escalation paradigm (adaptive dose escalation) improves the accuracy of PRE as well as the other methods.
Conclusions: The traditional 3+3 dose escalation scheme possesses weaknesses in both the escalation and estimation algorithms. The use of adaptive dose escalation coupled with better MTD estimation algorithms (such as PRE) can improve MTD estimation.
