<ul><li>Offline reinforcement learning (RL) in healthcare faces challenges due to out-of-distribution (OOD) issues, which can lead to harmful recommendations beyond clinical expertise.</li><li>Existing methods like conservative Q-learning (CQL) have limitations in addressing OOD problems by only constraining action selection, imitating short-term reward-focused clinician actions.</li><li>A new model-based Offline Guarded Safe Reinforcement Learning (OGSRL) framework is proposed to enhance treatment optimization strategies by regulating both action selection and downstream state trajectories.</li><li>OGSRL introduces an OOD guardian for safe policy exploration and a safety cost constraint to ensure policies remain within validated regions and align with medical safety boundaries, offering theoretical guarantees on safety and near-optimality.</li></ul>

Offline Guarded Safe Reinforcement Learning for Medical Treatment Optimization Strategies

Discover more