Scalable Personalised Treatment Recommendation in Critical Care

Sepsis is the leading cause of death in crit-ical care. Most sepsis patients did not receive effective treatment timely. Non-optimal medical treatments can lead to a lower survival rate. We aim to develop an effective treatment recommendation system for a sepsis-related patient in critical care based on reinforcement learning (RL). The main challenges are uncertainty in treatment effect, large model creation, and intensive computation. In this paper, we proposed our novel personalised treatment recommendation. We designed our concise patient representation based on the treatment effect. We clustered all patients and computed individual optimal policies for each cluster. We applied parallel and distributed computing to reduce model training time. We evaluated our proposed method using the MIMIC-III database. The result shows that our proposed method produced better policy values than baseline methods. Our scalable and efficient computation may lead to an effective and timely treatment recommendation system to help medical staff provide better healthcare delivery and increase survival rates.