Creating Personalized Medicine ML Models Using Distributed Learning via Ray

Sepsis is a life-threatening condition that requires timely and personalized treatment in intensive care units. Recent studies have applied reinforcement learning, or RL, to model optimal treatment policies, but traditional single-threaded training approaches are computationally expensive, requiring hundreds of thousands of episodes to converge. In this paper, we propose a distributed RL framework for sepsis treatment that parallelizes agent training using Ray, enabling scalable and efficient learning. We implement three RL algorithms: DQN, SARSA, and SAC, and evaluate them on the ICU-Sepsis environment. Our method significantly reduces training time and sample complexity while maintaining policy accuracy. Specifically, our Ray-based approach achieves up to $40 \times$ fewer total episodes and up to $125 \times$ faster convergence, without compromising average reward. These results demonstrate that distributed RL can accelerate the development of clinically meaningful policies in personalized medicine.