Towards Superhuman Imitation Learning for Sequential Head-and-Neck Cancer Treatment Decisions

This work presents the design of a simulator-driven imitation learning approach for sequential treatment decisions in head and neck cancer, built around Superhuman Policy Gradient Optimization (SPGO). Rather than simply replicating physicians' actions, the method leverages a clinical simulator to generate complete patient trajectories and incorporates an inverse-reinforcement-learning-inspired loss that rewards policies for outperforming experts on key clinical outcomes, including relapse rates and long-term toxicities.