Introduction: Constructing an accurate and comprehensive knowledge graph of specific diseases is critical for practical clinical disease diagnosis and treatment, reasoning and decision support, rehabilitation, and health management. For knowledge graph construction tasks (such as named entity recognition, relation extraction), classical BERT-based methods require a large amount of training data to ensure model performance. However, real-world medical annotation data, especially disease-specific annotation samples, are very limited. In addition, existing models do not perform well in recognizing out-of-distribution entities and relations that are not seen in the training phase.Method: In this study, we present a novel and practical pipeline for constructing a heart failure knowledge graph using large language models and medical expert refinement. We apply prompt engineering to the three phases of schema design: schema design, information extraction, and knowledge completion. The best performance is achieved by designing task-specific prompt templates combined with the TwoStepChat approach.Results: Experiments on two datasets show that the TwoStepChat method outperforms the Vanillia prompt and outperforms the fine-tuned BERT-based baselines. Moreover, our method saves 65% of the time compared to manual annotation and is better suited to extract the out-of-distribution information in the real world.

The proliferation of electronic health records (EHRs) and advances in deep learning have enabled personalized drug combination recommendations. However, traditional deep learning models often lack the contextual understanding and medical knowledge integration necessary for accurate predictions. While large language model (LLM)-based approaches address some of these challenges, they still fall short in incorporating critical medical knowledge, addressing comprehensive safety constraints such as multi-disease drug contraindications (MDCs), and providing sufficient interpretability of the causal mechanisms behind their outputs. To overcome these limitations, we propose KELLM, a knowledge-enhanced LLM framework for drug recommendations. By linking medical entities in EHRs to an external medical knowledge graph, inputs are enriched with causal chains, enhancing both prediction accuracy and interpretability. Additionally, we introduce a fine-tuned label-wise LLaMA model designed for multi-label classification, which incorporates safety considerations such as drug-drug interactions (DDIs) and MDCs to ensure clinically accurate and safe recommendations. Experimental results show that KELLM achieves state-of-the-art performance in effectiveness and safety metrics, while also providing evidence-based insights through causal chains that clarify its reasoning process. This establishes a new benchmark for trustworthy, interpretable drug combination recommendations.

The objective of entity alignment (EA) is to automatically discover entities across different knowledge graphs (KGs) that refer to the same real-world object, which is a crucial step for knowledge fusion. Graph neural network based methods currently dominate the EA landscape. However, the entity neighborhood often contains information about various aspects. Most of these methods treat the neighborhood of an entity as a whole, aggregating all its neighbors into a single, static representation. This can introduce noise and lead to incorrect alignment results due to the heterogeneous neighborhood structures of aligned entities. In this paper, we propose a novel relational reflective disentangled graph neural network (RRDGNN) for EA, which learns disentangled representations for different aspects of entities, thereby alleviating the prevalent issue of neighborhood heterogeneity in EA. Specifically, RRDGNN designs a relational reflective neighborhood routing mechanism for learning disentangled, dynamic entity representations at both micro and macro levels. An alignment-aware self-attention strategy is then employed to assign adaptive weights to different entity aspects. Finally, multi-aspect seed generators are co-trained in a semi-supervised learning framework. They favor entities with specific sparsity, which helps to refine the seed selection process and generate high-quality alignment seeds. Extensive experiments on five real-world datasets, along with detailed ablation studies and analyses, demonstrate that RRDGNN consistently outperforms existing EA methods.