Abstract
The design of adaptive chemotherapy is complicated by nonlinear tumor-immune interactions, delayed drug transport and uncertainty in state evolution. This study develops a delay-aware computational framework that couples a normalized normal-tumor-immune model with central and peripheral pharmacokinetic compartments. A Proximal Policy Optimization (PPO) agent observes the population states, their rates of change and both drug concentrations, and selects a bounded continuous infusion action. Pharmacodynamic killing is driven by peripheral exposure, thereby distinguishing systemic administration from the concentration acting at the tumor site. Positivity and boundedness of the deterministic biological-pharmacokinetic system are established for nonnegative initial states and bounded infusion. The learning objective jointly penalizes tumor burden, positive tumor growth, cumulative exposure, abrupt normalized-action variation and depletion of normal and immune populations. Numerical experiments compare untreated and periodic regimens, one- and two-compartment environments, nominal and stochastic transitions, and PPO with two actor-critic baselines. Explicit distribution dynamics change the learned policy from a sustained high-action plateau to an early loading phase followed by gradual tapering. Moderate transition randomization yields the smallest reported tumor-tracking errors and reduced trajectory dispersion, whereas stronger perturbations substantially widen the uncertainty bands. These results identify pharmacokinetic representation as a consequential component of learning-based treatment design. The framework is an in-silico mechanism study rather than a clinical dosing recommendation and requires drug-specific calibration, hard toxicity constraints, complete statistical evaluation and biological validation.
Funding Information
  • National Natural Science Foundation of China (12301486)

This publication has 25 references indexed in Scilit: