MedKPL: A heterogeneous knowledge enhanced prompt learning framework for transferable diagnosis.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Clinical diagnosis relies on clinician expertise and external medical knowledge.
• Current AI diagnosis systems struggle with extensibility and integrating diverse knowledge sources for clinical note classification.

Purpose of the Study:

• To propose a novel Medical Knowledge-enhanced Prompt Learning (MedKPL) framework.
• To enhance the performance and transferability of AI-based clinical note classification systems.

Main Methods:

• Developed MedKPL to unify heterogeneous knowledge sources (knowledge graphs, QA databases) into fixed-format text sequences.
• Integrated unified medical knowledge into prompts for improved context representation in AI models.
• Evaluated the framework on two medical datasets for classification and cross-departmental transfer tasks.

Main Results:

• MedKPL achieved superior medical text classification results compared to existing methods.
• Demonstrated enhanced performance in few-shot and zero-shot cross-departmental transfer tasks.
• The framework effectively integrates explicit and implicit medical knowledge.

Conclusions:

• MedKPL offers a robust solution for integrating diverse medical knowledge into AI diagnostic systems.
• The framework improves the interpretability and transferability of AI diagnostic systems, particularly in low-data scenarios.
• MedKPL has the potential to significantly advance AI-powered clinical decision support.