Large language model-augmented offline reinforcement learning framework for sepsis management in critical care

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Healthcare Data Analytics

Background:

• Sepsis management requires timely and accurate clinical decisions.
• Current Reinforcement Learning (RL) models for sepsis often lack comprehensive patient context due to reliance on structured data.
• Integrating unstructured clinical notes can provide crucial contextual information.

Purpose of the Study:

• To develop a Multimodal Offline Reinforcement Learning framework (MORE-CLEAR) for sepsis management.
• To leverage Large Language Models (LLMs) for extracting semantic representations from clinical notes.
• To enhance patient state representation by integrating multimodal data for improved RL-based treatment strategies.

Main Methods:

• Developed the MORE-CLEAR framework utilizing LLMs to process clinical notes and extract rich semantic features.
• Implemented gated fusion and cross-modal attention mechanisms for dynamic integration of multimodal data.
• Validated the framework using two public (MIMIC-III, MIMIC-IV) and one tertiary ICU dataset (SNUH).

Main Results:

• MORE-CLEAR significantly improved estimated survival rates compared to single-modal RL approaches.
• Enhanced policy performance was observed in RL models utilizing the MORE-CLEAR framework.
• The framework demonstrated effective integration of structured data and clinical note-derived information.

Conclusions:

• The MORE-CLEAR framework offers a novel approach to sepsis management by incorporating clinical narrative context.
• LLM-driven multimodal data integration in RL can lead to more robust and effective clinical decision-making.
• This approach has the potential to expedite sepsis management and improve patient outcomes.