Optimizing sepsis treatment strategies via a reinforcement learning model

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Sepsis Management

Background:

• Current sepsis treatment relies heavily on clinician experience, lacking standardized references.
• Effective medication strategies are crucial for improving patient outcomes in sepsis.

Purpose of the Study:

• To develop a reinforcement learning model for assisted sepsis medication treatment.
• To optimize vasopressor and intravenous infusion dosages for sepsis patients.

Main Methods:

• Utilized Sepsis 3.0 criteria and 19,582 patient records from MIMIC-III database.
• Employed Dueling Double Deep Q-Network (DDQN) for predicting medication strategies.
• Developed a hybrid model (Safe-Dueling DDQN + expert strategies) for enhanced safety and optimization.

Main Results:

• Dueling DDQN model demonstrated superior performance over clinical strategies, reducing mortality from 16.8% to 13.8%.
• Safe-Dueling DDQN minimized high-risk actions, particularly large vasopressor dose fluctuations.
• The hybrid model effectively balanced AI-driven and expert-guided treatment decisions.

Conclusions:

• The proposed reinforcement learning model offers practical clinical value for sepsis medication treatment.
• The AI-assisted approach can enhance patient survival rates while ensuring medication safety and balance.