Neural MPC for Safety-Critical Biological Systems: an Application to Tumor-Immune Cancer Dynamics

Model Predictive Control (MPC) enforces state and input constraints but depends on accurate models, which are often unavailable or too complex in biomedical settings. A neural MPC framework is proposed in which a feed-forward neural network, trained from input–state data, serves as the prediction model. The approach is validated in silico on a tumor–immune system coordinating chemotherapy and immunotherapy under a safety constraint. Neural MPC achieves regulation comparable to model-based MPC and optimal control, while remaining robust to parameter uncertainty, and outperforms a baseline ARX-MPC which converges to a displaced equilibrium with steady-state tracking error.