Maja Skataric: No financial relationships to disclose
Objectives: Titrated drug regimens are increasingly used to mitigate the risk of adverse events (AEs). For example, GLP1 agonists and antipsychotics are titrated to mitigate common AEs such as nausea/vomiting, and akathisia/dystonia, respectively. During the clinical development of such drugs, multiple titration regimens are tested to help identify optimal regimens. However, analysis of titrated data including dose-exposure-response modeling is challenging due to confounding associated with “drug titration paradox” [1]. In this methodological work, we performed Monte Carlo simulations to gain insights into this confounding. Additionally, we present a fit-for-purpose framework for analyzing titrated data.
Methods: AE data were generated using a multivariate logistic regression (MLR) model for 2-step titration regimens with a starting, an intermediate and a target dose. A total of 64 combinations of doses were simulated with or without the assumption of development of tolerance in AEs associated with titration. Dose-response analyses of simulated data were conducted using MLR and the results were compared to the “ground truth” used in data generation. In addition, univariate regression and random forests were explored to identify starting doses and titration steps, respectively.
Results: In the scenario without tolerance, MLR resulted in reliable estimation of the regression coefficients for the target dose (Root Mean Square Error (RMSE) < 10%), while the coefficients for the starting dose were poorly estimated (RMSE>50%). In some cases, estimated coefficients had opposite signs compared to the ground truth, indicating confounding. In the scenario with tolerance, this was more pronounced, with estimated coefficients further diverging from the ground truth and impacting dose selection. Although the target dose range with acceptable AE incidence rates was determined accurately ( < 20% deviation from the ground truth), starting doses and step sizes (ratios of intermediate to the starting dose, and the target to the intermediate dose) were poorly determined (>50% deviation from the ground truth). Additionally, factors such as decreased sample size, or reduced design space (a subset of doses used) further amplified the impact of confounding. In an attempt to remove confounding among starting doses and step sizes, a univariate model was developed for AEs associated with the starting dose only, i.e., AEs that occurred after the starting dose and before the intermediate dose. Model results were adequate (~30-40% deviation from the ground truth) to inform acceptable starting doses. Furthermore, random forests suggested that the step size between the intermediate and the first dose should be smaller compared to the step size between the target and the intermediate dose to ensure acceptable AE rates.
Conclusions: In the presence of confounding, optimization of titration regimen is challenging, especially the choice of starting doses and step sizes. Via Monte Carlo simulations, this work shed light on potential impact of confounding. Fit for purpose approaches including univariate regression as well as Machine Learning tools such as random forests can help navigate the challenges associated with titrated data.
Citations: [1] T. Schnider, CF. Minto, Martin Luginbuhl and T.D. Egan. The drug titration paradox: more drug does not correlate with more effect in individual clinical data. British Journal of Anaesthesia, 129 (6): 861e867 (2022)
Keywords: drug titration, Monte Carlo simulations, Machine Learning
