Explainable differential diagnosis with dual-inference large language models.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Automatic differential diagnosis (DDx) is crucial for patient care.
• Large language models (LLMs) show diagnostic potential but struggle with high-quality DDx explanations.
• Lack of specialized datasets and LLM reasoning challenges hinder DDx explanation development.

Purpose of the Study:

• To develop the first publicly available DDx dataset with expert-derived explanations.
• To propose and evaluate a novel framework (Dual-Inf) for generating precise LLM-based DDx explanations.
• To enhance the explainability of LLMs in clinical differential diagnosis.

Main Methods:

• Creation of a novel DDx dataset with 570 expert-annotated clinical notes.
• Development of the Dual-Inf framework to elicit high-quality DDx explanations from LLMs.
• Comprehensive evaluation of LLM explainability for differential diagnosis.

Main Results:

• The first specialized DDx explanation dataset is now publicly available.
• The Dual-Inf framework demonstrates effectiveness in generating precise DDx explanations.
• This work represents a significant advancement in LLM-based clinical explainability.

Conclusions:

• Bridged a critical gap in differential diagnosis explanation generation.
• Enhanced the capability of LLMs to provide clear and accurate DDx explanations.
• Aimed at improving clinical decision-making through better AI-driven diagnostic reasoning.