Speaker
Description
Mathematical modeling in oncology frequently encounters a trade-off between the structural interpretability of mechanistic equations and the flexible predictive power of data-driven algorithms. While mechanistic models provide essential biological constraints, they are often subject to structural misspecification when faced with the high dimensionality of clinical heterogeneity. Conversely, purely statistical machine learning (ML) approaches lack the inductive biases necessary to generalize from the sparse and noisy datasets typical of clinical practice. This work presents a methodological framework that treats mechanistic models and ML as complementary components of a unified architecture. The use of ML to learn mechanistic model residuals is discussed, thereby augmenting the predictive accuracy of first-order mechanistic approximations. Hierarchical Bayesian modeling is examined as a robust approach for parameter estimation across patient cohorts, facilitating the personalization of mechanistic dynamics. Operator learning is investigated to map high-dimensional clinical data to patient-specific dynamics. Finally, the integration of reinforcement learning (RL) with mechanistic simulations is demonstrated, where the latter provides a biologically grounded environment for optimizing sequential treatment policies. By formalizing these hybrid methodologies, this work aims to contribute to the ongoing development of decision support tools in oncology.
