An interpretable RL framework for pre-deployment modeling in ICU hypotension management

Area of Science:

• Computational medicine
• Clinical decision support systems
• Artificial intelligence in healthcare

Background:

• Reinforcement learning models offer potential for clinical decision-making, such as hypotension management.
• A key challenge is the lack of interpretability in these models, hindering clinical validation and trust.
• Existing data-driven strategies often fail to provide individualized treatment recommendations.

Purpose of the Study:

• To develop a general framework for creating interpretable computational treatment strategies.
• To identify specific clinical contexts where treatment choices differ significantly.
• To facilitate clinical validation and adoption of AI-driven recommendations.

Main Methods:

• Developed a framework to identify critical clinical decision points and their associated treatment choices.
• Focused on creating succinct sets of recommendations for specific contexts.
• Applied the framework to hypotension management in the intensive care unit (ICU).

Main Results:

• Generated interpretable treatment strategies that are easily visualized and verified by clinicians.
• Enabled clinicians to integrate their expertise with historical data for validation.
• Demonstrated the framework's utility in a critical care setting.

Conclusions:

• The developed framework enhances the interpretability and clinical utility of AI-driven treatment strategies.
• This approach supports data-driven, individualized decision-making in complex medical domains like ICU hypotension management.
• The framework has broad applicability for AI-assisted clinical decision-making across various specialties.