Machine learning-enhanced Sensitivity Analysis for Complex Pharmaceutical Systems

Pharmacokinetic and pharmacodynamic (PK/PD) models are used to predict drug transport in the body and to assess treatment efficacy and optimal dosage. The kinetic parameters embedded in the models, which define transport across body compartments or drug efficacy, can be linked to patient-specific characteristics; understanding the parameter space-model output relationship is critical towards linking patient population heterogeneity to the therapeutic outcome variability. Global Sensitivity Analysis (GSA) is a well-established tool used to examine parameter-to-parameter interactions, shedding light on underlying interactions towards enhanced system understanding. Despite its potential and usefulness, GSA performance is dependent to the model complexity; large-scale and nonlinear PK/PD models, which often have large sets of parameters, can render GSA challenging to perform, requiring excessive computational effort. Proposed approaches to reduce GSA complexity, such as segmentation in parameter subsets or the introduction of surrogate metamodels, become less effective as the number of kinetic parameters grows. In this work, we investigate the potential of Machine Learning (ML) to reduce the complexity of PK/PD models by exploring how the level of hybridisation can impact the GSA performance and, critically, whether the use of surrogates affects the resulting model sensitivity to parametric uncertainty. We show that ML-based surrogates can reliably identify parameter interactions and sensitivities while requiring only a limited number of simulations of the reference mechanistic model. Further, surrogates models effectively reduce the computational expenditure of GSA of multi-dimensional nonlinear PK/PD models. The accelerated execution of GSA enables performing patient cohort-specific analysis, with potential applications for optimal study design and for precision medicine.