Optimal medication dosing from suboptimal clinical examples: a deep reinforcement learning approach.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Pharmacometrics

Background:

• Medication misdosing, particularly with drugs like heparin, poses significant risks, increases hospital stays, and strains resources.
• Current dosing guidelines may not adequately account for individual patient variability and evolving clinical status.

Purpose of the Study:

• To develop and evaluate a clinician-in-the-loop sequential decision-making framework for individualized heparin dosing.
• To leverage deep reinforcement learning to create an adaptive dosing policy based on patient data.

Main Methods:

• Utilized retrospective data from the MIMIC II intensive care unit database.
• Developed a deep reinforcement learning algorithm to learn optimal heparin dosing strategies from electronic medical records.
• Employed separate training and testing datasets to validate the model's performance.

Main Results:

• The developed deep reinforcement learning model proposed heparin doses that led to improved expected patient outcomes compared to standard clinical guidelines.
• The clinician-in-the-loop framework demonstrated effectiveness in adapting dosing policies to individual patient phenotypes.
• The model achieved better outcomes than existing clinical guidelines in retrospective analysis.

Conclusions:

• A sequential modeling approach using retrospective data can inform bedside clinical decisions for personalized medication dosing.
• Deep reinforcement learning offers a promising avenue for optimizing the administration of medications with sensitive therapeutic windows.
• Individualized dosing policies derived from AI can enhance patient safety and resource utilization in intensive care settings.