Patient-Specific Sedation Management via Deep Reinforcement Learning.

Area of Science:

• Artificial Intelligence in Medicine
• Pharmacology and Therapeutics
• Critical Care Medicine

Background:

• Individualized medication dosing is complex due to patient-specific factors.
• Existing data-driven sedative dosing models struggle with interindividual variability.
• Developing personalized dosing strategies for intensive care unit (ICU) sedation is crucial.

Purpose of the Study:

• To create an individualized framework for sedative-hypnotic drug dosing.
• To leverage deep reinforcement learning for personalized sedation management.
• To optimize the dosing of propofol and fentanyl in ICU patients.

Main Methods:

• Utilized a deep reinforcement learning (RL) approach, modeling sedation dosing as a Markov Decision Process.
• Developed an RL agent using a deep deterministic policy gradient with prioritized experience replay.
• Trained and evaluated the model on 1,757 patients from the MIMIC IV ICU database, comparing against clinician performance.

Main Results:

• The RL agent demonstrated superior performance in managing sedation (8% improvement) and mean arterial pressure (26% improvement) compared to clinician data.
• Statistical analysis confirmed the significant improvement in performance (p < 0.05).
• The model effectively learned personalized policies for propofol and fentanyl administration.

Conclusions:

• The developed RL framework significantly enhances the ability to manage sedation and hemodynamic stability in ICU patients.
• The model provides a data-driven approach to assist clinicians in making optimal sedation dosing decisions.
• This individualized dosing strategy contributes to better patient health condition management in critical care settings.