Leveraging ChatGPT and explainable AI for enhancing clinical decision support

Area of Science:

• Artificial Intelligence
• Clinical Decision Support
• Machine Learning

Background:

• Large language models (LLMs) show promise in natural language processing but struggle with structured clinical data.
• Tabular clinical data requires specific reasoning mechanisms for accurate analysis and prediction.
• Existing methods lack effective integration of LLM reasoning with domain-specific tabular data.

Purpose of the Study:

• To introduce HealthAI-Prompt, a novel framework for adapting LLMs to tabular clinical data.
• To enhance LLM capabilities for clinical decision-making, specifically diabetes risk prediction.
• To improve the accuracy and transparency of AI in healthcare through domain knowledge integration.

Main Methods:

• Developed HealthAI-Prompt framework using contextual prompts with task descriptions and domain knowledge.
• Integrated insights from automated machine learning (AutoML) models and their local explanations.
• Evaluated explanation reliability using fidelity, stability, and monotonicity metrics.
• Embedded validated explanations into prompts for LLM interpretation of structured features without fine-tuning.

Main Results:

• HealthAI-Prompt enables LLMs to interpret structured clinical features meaningfully.
• The framework improves predictive accuracy for tasks like diabetes risk prediction.
• Comparative analysis demonstrated the impact of prompt engineering strategies on model performance.
• The approach offers enhanced transparency in healthcare AI compared to traditional methods.

Conclusions:

• HealthAI-Prompt effectively bridges AutoML and LLM reasoning for tabular clinical data.
• The method enhances LLM performance and interpretability in healthcare AI applications.
• This framework represents a significant advancement in applying LLMs to domain-specific clinical challenges.