Reinforcement Learning-Optimized Personalized Cancer Treatment Strategies: A Case Study of Lung Cancer

Personalized cancer treatment strategies (PCTS) tailor treatments on the basis of a patient’s health status, cancer type, and stage. By considering the evolving interactions of treatment options over time, PCTS seeks to balance cancer suppression with minimizing harm and maximizing therapeutic benefits. However, limited clinical trial resources limit the ability to explore optimal PCTSs fully through experimentation, presenting a significant challenge to their development. In this study, we introduce a "digital twin" model that integrates comprehensive patient health data, cancer characteristics, and individual treatment responses and employs reinforcement learning (RL) to identify the optimal PCTS. Using lung cancer as a case study, we calibrated model parameters for various demographic groups, cancer stages, and treatment options, utilizing real clinical data from the SEER dataset. The RL-optimized PCTS significantly outperformed traditional clinician decisions, leading to notable improvements in patient survival. For example, among women aged 45--64 years with stage IIIA, IIIB, IVA, and IVB lung cancer, survival increased by 46%, 59%, 23%, and 149%, respectively. Similarly, for men aged 45--64 years, survival improved by 108%, 97%, 40%, and 62%, respectively, across the same stages. This study lays a critical foundation for the use of AI in optimizing PCTS and paves the way for further research and clinical applications.