Personalized Causal Graph Reasoning for LLMs: An Implementation for Dietary Recommendations.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Personalized Medicine

Background:

• Large Language Models (LLMs) possess general reasoning capabilities but struggle with personalized decision-making using multifactorial individual data.
• This limitation hinders LLM application in personalized healthcare, necessitating context-specific adaptations.
• Existing methods lack the ability to integrate individual-specific causal factors for tailored recommendations.

Purpose of the Study:

• To introduce a novel framework, Personalized Causal Graph Reasoning, for LLMs to perform personalized reasoning over individual-specific causal graphs.
• To enable LLMs to construct and utilize longitudinal data for creating user-specific causal models.
• To enhance LLM applicability in data-driven domains like healthcare by enabling personalized decision-making.

Main Methods:

• Constructing individual-specific causal graphs from longitudinal data, encoding user-specific factors and their influence on outcomes.
• Developing an LLM-based framework to traverse these causal graphs, identify relevant pathways, and simulate outcomes.
• Implementing the framework for nutrient-oriented dietary recommendations, focusing on personalized glucose control strategies.
• Employing counterfactual evaluation to assess the effectiveness of LLM-generated food suggestions.

Main Results:

• The Personalized Causal Graph Reasoning framework demonstrated a reduction in postprandial glucose Incremental Area Under the Curve (iAUC) across three time windows compared to previous methods.
• LLM-generated dietary suggestions showed improved effectiveness in managing glucose levels.
• LLM-as-a-judge evaluations confirmed enhanced personalization quality of the generated recommendations.

Conclusions:

• Personalized Causal Graph Reasoning empowers LLMs to provide highly tailored health recommendations by reasoning over individual causal graphs.
• This framework significantly improves personalized dietary strategies for glucose control, outperforming prior approaches.
• The study highlights the potential of integrating causal inference with LLMs for advancing personalized medicine and healthcare applications.