Deep reinforcement learning for personalized insulin dosing and glucose control of hospitalized in ICU patients

Panagiotis Symeonidis, Evangelos Rizos, Christos Andras, Stergios Hairistanidis, Yannis Manolopoulos, Markus Zanker

Abstract: Patients with diabetes struggle to control blood glucose with appropriate insulin doses. Classic prediction methods such as multi-linear regression often lead to ineffective blood’s glucose control, because they are not capable to identify the correct neighborhood (other similar patients) of the target patient, since they do not take under consideration demographic (age, gender, etc.) or other clinical similarities of patients’ therapeutics (blood glucose levels, etc.). In this paper, we use deep reinforcement learning (DRL), and define a transition probability function to intelligently select the right neighbors of the target patient. We also define a reward and policy function for our DRL method, which can predict the optimal insulin dose for a patient and achieves optimal glucose control, while decreasing hypoglycemia risk. Our experimental results show that our Deep Q Network (DQN) algorithm can assist doctors to administer optimal insulin dose for diabetes treatment and keep the glucose levels in blood for longer time periods in accepted values range.

External IDs:dblp:journals/ijdsa/SymeonidisRAHMZ25