Bridging Stepwise Lab-Informed Pretraining and Knowledge-Guided Learning for Diagnostic Reasoning.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Biomedical Data Science

Background:

• Electronic Health Records (EHR) are increasingly used for AI-assisted diagnosis prediction.
• Current data-driven models often lack comprehensive medical knowledge and structured reasoning capabilities.
• Integrating clinical expertise into AI models remains a significant challenge.

Purpose of the Study:

• To investigate whether incorporating medical knowledge can enhance AI predictive models for stepwise clinical reasoning.
• To develop a novel framework that combines external medical knowledge with patient-specific EHR data.
• To improve the accuracy and interpretability of AI-driven diagnostic predictions.

Main Methods:

• Proposed DuaLK, a dual-expertise framework combining a Diagnosis Knowledge Graph (KG) with patient EHR data.
• Constructed a KG encoding hierarchical and semantic relations, enriched by large language models (LLM).
• Introduced a lab-informed proxy task to guide stepwise clinical reasoning based on lab test signals.

Main Results:

• DuaLK consistently outperformed existing baseline models across four clinical prediction tasks on two public EHR datasets.
• The framework demonstrated improved accuracy and interpretability in diagnostic predictions.
• The lab-informed proxy task effectively guided the model towards clinically consistent reasoning.

Conclusions:

• Combining structured medical knowledge (KG) with individual clinical signals (EHR data) significantly enhances AI diagnostic prediction.
• The DuaLK framework offers a promising approach for developing more reliable and interpretable AI tools for clinical decision support.
• Future work can explore further integration of diverse medical knowledge sources and advanced reasoning mechanisms.