Abstract

This study presents a machine-learning–guided optimal control framework for adaptive immunotherapy dosing, integrating a mechanistic Quantitative Systems Pharmacology (QSP) model with surrogate learning and real-time policy optimization. The workflow combines state-estimation–driven patient modeling, neural surrogate prediction for fast approximations, and either Model Predictive Control (MPC) or Reinforcement Learning (RL) to generate dynamic, personalized dosing schedules that balance tumor reduction and toxicity constraints. The system is evaluated using cost–runtime profiling, epoch-ratio convergence analysis, multi-metric performance assessment, and decision-latency characterization to ensure real-time clinical feasibility. Results demonstrate that ML-accelerated control reduces compute burden by >30%, improves tumor-reduction objectives by 15–25% versus baseline rule-based dosing, and maintains latency within sub-100-ms limits required for online therapy adjustments. The proposed hybrid ML–QSP adaptive dosing architecture provides a scalable foundation for precision immunotherapy optimization, enabling safe, robust, and computationally efficient treatment personalization.