Скачать или смотреть KDD 2023 - Deep Offline Reinforcement Learning for Real-world Treatment Optimization Applications

Mila Nambiar,Institute for Infocomm Research (I2R), A*STAR

There is growing interest in applying deep reinforcement learning to recommend optimal medical treatments in critical care and chronic disease management settings. But this is challenging as treatment optimization applications don?t permit learning through direct exploration of an environment. For safety reasons, recommendations must instead be learned from retrospective data, where suboptimal treatments can be overrepresented. To address these challenges, we introduce a practical and theoretically grounded transition sampling approach for deep offline reinforcement learning. We give a preview of our main findings on both diabetes and sepsis treatment optimization tasks. Namely, our proposed solution outperforms baselines in terms of expected health outcomes and consistency with clinical safety guidelines.