Optimal Fluid And Vasopressor Interventions In Septic ICU Patients Through Reinforcement Learning Model

Introduction: Fluids and vasopressors represent the cornerstone for hemodynamic instability management in the intensive care unit (ICU). However, optimal personalized treatments strategies are still missing. Goal: To evaluate the ability of a reduced set of cardiovascular features in determining optimal actions with a reinforcement learning approach. Methods: Data were extracted from the MIMIC-III database Patients' trajectories were modeled as a Markov decision process with a target reward based on 90-day mortality. Performances with a reduced set of cardiovascular features (CARDIO), including heart rate, systolic and diastolic blood pressure, shock index, and oxygen saturation were compared with a random policy model (RANDOM) and a model with a full set of 48 clinical variables including physiologic, laboratory measurement, and ventilation parameters (FULL). Results: The CARDIa model achieved the highest results with a 95% lower bound (LB) of estimated policy value equal to 96.17 compared with the 86.00 obtained from the FULL model and 82.62 from the RANDOM policy model. Conclusions: Results show that cardiovascular features and ongoing treatments have the potential to determine the optimal dosage of fluids and vasopressors for septic patients when using reinforcement learning tools for the development of medical decision support systems.