Biomedical Knowledge Graph Alignment with GPT-Augmented Similarity Feature Construction via Tree-based Particle Swarm...

Area of Science:

• Biomedical Informatics
• Artificial Intelligence
• Knowledge Representation

Background:

• Biomedical Knowledge Graph (KG) alignment is crucial for integrating diverse biomedical data but faces challenges due to heterogeneity in terminologies, structures, and semantic granularity.
• Existing methods struggle to effectively combine multiple Generative Pre-trained Transformer (GPT)-augmented similarity features (SFs) due to expanded search spaces and potential feature conflicts.

Purpose of the Study:

• To propose a novel framework, Tree-based Particle Swarm Optimization with Adaptive Fitness Optimization (T-PSO-AFO), for constructing robust GPT-augmented SFs for biomedical KG alignment.
• To enhance the accuracy and efficiency of integrating heterogeneous biomedical information sources.

Main Methods:

• Introduced a GPT-based method for constructing expressive and context-aware SFs to capture biomedical entity equivalence.
• Developed a tree-based symbolic representation within a Particle Swarm Optimization (PSO) framework to effectively explore complex SF combinations.
• Implemented an adaptive fitness landscape optimization mechanism to dynamically adjust the search space for improved convergence and alignment.

Main Results:

• The T-PSO-AFO framework significantly outperforms existing state-of-the-art biomedical entity matching approaches on benchmark datasets (OAEI's LargeBio and Disease and Phenotype).
• Demonstrated improved robustness, effectiveness, and scalability in aligning heterogeneous biomedical KGs.

Conclusions:

• The T-PSO-AFO framework offers a significant advancement in biomedical KG alignment by effectively integrating diverse semantic features.
• The proposed approach enhances the reliability of downstream applications such as clinical decision support and personalized healthcare.