Medical Knowledge-Driven Contrastive Learning for Similar Patient Retrieval.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Machine Learning

Background:

• Similar patient retrieval is crucial for diagnosis and treatment recommendations.
• Traditional methods struggle with implicit semantic relationships in clinical data.
• Deep learning retrieval methods show promise but require task-specific adaptation.

Purpose of the Study:

• To enhance general-purpose embedding models for medical text using a knowledge-driven contrastive learning approach.
• To improve the accuracy and robustness of similar patient retrieval.
• To address limitations in existing dense retrieval methods for medical informatics.

Main Methods:

• Developed a medical knowledge-driven contrastive learning framework.
• Introduced a novel negative sampling strategy using International Classification of Diseases (ICD) codes.
• Implemented an external knowledge-based negative sampling method incorporating statistical and ambiguous knowledge to address data imbalance and improve differentiation of medical conditions.

Main Results:

• The proposed method significantly improved patient representation capacity.
• Achieved substantial performance gains over state-of-the-art baseline models on real-world medical datasets.
• Demonstrated enhanced ability to differentiate fine-grained medical conditions and complex clinical scenarios.

Conclusions:

• The proposed medical knowledge-driven contrastive learning approach effectively enhances similar patient retrieval.
• The novel negative sampling strategies overcome data imbalance and improve model robustness.
• This method offers a promising advancement for medical informatics and clinical decision support systems.