Arijit Chakravarty: No relevant disclosure to display
Objectives: Traditional Phase I dose-escalation strategies in Phase I oncology trials, such as the ‘3+3’ method, are widely used but estimate the MTD (maximum tolerated dose) poorly, being vulnerable to both bias (underestimation of MTD) and the risk of subtherapeutic dosing. Bayesian adaptive designs offer a model-based alternative that incorporates accumulating data to refine dose selection in real time. Here we present a novel approach for the assessment of Phase I MTD, based on the integration of a Bayesian decision metric with a dose escalation algorithm to estimate the MTD.
Methods: We have previously presented [1] a novel decision metric for estimating Phase I MTD based on a Bayesian algorithm, the Population Response Estimate (PRE). PRE applies Maximum Likelihood Theory to graded toxicity data to estimate both the toxicity central tendency curve and population heterogeneity, which are then used to estimate the likelihood of encountering dose-limiting toxicities at each dose. Here we extend our method to incorporate a novel dose escalation algorithm that leverages the PRE to calculate the highest safe dose (HSD) given the current certainty level, defined as the highest dose with ≤15% posterior probability of DLT > grade 4. Patients are then dosed at this HSD, and the posterior distribution is updated iteratively following each dosing cohort. Dose escalation continues until the standard deviation of the MTD estimate falls below a prespecified threshold, defined as <=25% of the MTD estimate. We then benchmarked the performance of this novel Bayesian method against the 3+3 design using simulation across 1,000 virtual trials with known dose-toxicity relationships.
Results: The Bayesian adaptive method identified the true MTD within ±20% in 73% of simulated trials, compared to only 38% using the 3+3 design. Notably, the probability of overestimating the MTD (i.e., selecting a dose above the true MTD) was only 1% with the Bayesian method, versus 12% for 3+3. The adaptive design required fewer patients on average and concentrated observations near the MTD, enhancing trial efficiency and ethical conduct.
Conclusions: The Bayesian adaptive design outperforms the conventional 3+3 method in both accuracy and efficiency of MTD estimation. By focusing data collection near the target toxicity threshold and continuously updating model estimates, this method supports faster, safer, and more informative Phase I trials. These findings align with FDA guidance encouraging the use of model-based designs and highlight the role of pharmacometric methods in optimizing early-phase clinical development.
Citations: [1] Yuan, L., Hooper, R., Stoddard, M., White, D., Bottino, D., Chakravarty, A. (2024, October). Design and evaluation of a novel decision metric for estimating Phase 1 maximum tolerated dose [Poster presentation]. American Conference on Pharmacometrics (ACoP2024), Phoenix, AZ.
