Rule-augmented constraint learning for semantic error detection in MIMIC-III knowledge graph.

Area of Science:

• Clinical Informatics
• Knowledge Representation and Reasoning
• Data Mining and Machine Learning

Background:

• High-quality data is crucial for reliable clinical decision support systems.
• Clinical knowledge graphs (KGs) offer structured data but face challenges in consistency and correctness.
• Existing rule mining methods often yield redundant or irrelevant constraints for clinical data.

Purpose of the Study:

• To propose a novel framework for constraint learning in clinical KGs.
• To transform high-confidence rules into clinically plausible constraints for semantic error detection.
• To enhance the trustworthiness and clinical usability of KGs.

Main Methods:

• Developed a framework for constraint learning in clinical KGs.
• Employed two approaches: class disjointness and literal clustering combined with rule mining.
• Validated clinical relevance using expert-curated constraints and large language models (LLMs).

Main Results:

• Rule filtering effectively preserved clinically meaningful rules aligned with medical knowledge on the MIMIC-III dataset.
• A clustering-based method achieved reliable value groupings for numeric data.
• LLM validation confirmed the clinical relevance of a portion of the discovered rules.

Conclusions:

• The proposed constraint learning framework offers interpretable and scalable solutions for semantic inconsistencies in clinical KGs.
• This approach enhances KG trustworthiness and clinical utility for data-driven applications.
• The methods contribute to improving the reliability of clinical decision support systems.