荆炳义，南方科技大学

Electronic health records (EHR) have provided a great opportunity to exploit personalized health data to optimize clinical decision making and achieve personalized treatment recommendation. In this talk, we explore how AI could help physicians in prescribing medicines for patients with multi-morbidity. Both Supervised Learning (SL) and Reinforcement Learning (RL) have been employed for this purpose, but with their own drawbacks. For instance, SL relies highly on the clinical guideline and doctors personal experience while RL may produce unacceptable medications due to lack of the supervision from doctors. In this talk, we propose a novel SAVER framework by fusing RL and SL, where RL learns the optimal policy and SL gives a regularization to avoid unacceptable risks. Our experiments show that our SAVER framework can provider more accuracy treatment recommendation than the existing methods.