Evaluation of a model averaging algorithm for model-informed precision dosing in the context of parameter misspecifications

A primary challenge for integrating model-informed precision dosing (MIPD) in clinical settings is selecting the best-fit model for newly presented patients. A model averaging algorithm (MAA) proposed by Uster et al., addresses this by leveraging multiple models simultaneously. As the ‘library’ of population pharmacokinetic models for a given therapeutic expands, an understanding of how model misspecifications influence MAA can provide insight to guide the optimal selection of model combinations. The present simulation study evaluated the impact of parameter model misspecifications on MAA performance across three scenarios: (1) misspecified structural parameter (θ CL ), (2) misspecified residual error (σ²), and (3) misspecified interindividual variability on clearance (ω 2 CL ). Predictive performance for Bayesian forecasting was assessed within each misspecification scenario for single candidate models, a re-estimated model and various combinations of models in MAA. Performance was evaluated for precision and accuracy. The results showed greatest MAA benefit under structural model misspecification, where combining models with opposing bias showed improved predictive performance. In the case of misspecified residual error and interindividual variability (IIV), MAA tended to down-weight models with high residual error or favor those with higher IIV, respectively. Overall, these findings can be used to inform the selection of model combinations to improve MAA performance, supporting more robust MIPD implementation.