Graph-Augmented Retrieval Techniques for Medical QA Applications

Machine Learning (ML) has emerged as a powerful computational framework capable of uncovering complex patterns from large-scale datasets and generating valuable insights across diverse domains. At the core to enhancing ML capabilities, especially in applications demanding precise information retrieval of textual source data, are knowledge graphs— structured representations consisting of entities and the relationships connecting them. Knowledge graphs explicitly encode semantic relationships and structural interdependencies, enabling ML models to leverage rich contextual information. This thesis investigates the integration of Machine Learning (ML) techniques with structured knowledge graphs to enhance retrieval capabilities within Retrieval-Augmented Generation (RAG) frameworks with a particular focus on biomedical question-answering applications. Despite the significant progress achieved by traditional RAG methods, which primarily employ semantic embeddings derived from pre-trained transformers like BERT, these techniques inherently suffer from limitations. Embedding-based similarity searches typically treat textual segments independently, neglecting the rich structural and semantic interconnections inherently present among biomedical documents. This often results in incomplete or fragmented retrieval. To systematically address these limitations, this thesis constructs a comprehensive biomedical knowledge graph by integrating structured data extracted from PubMed articles along with standardized Medical Subject Headings (MeSH) terms, capturing explicit semantic relationships and ensuring structural coherence across biomedical concepts and textual entities. Subsequently, the research explores innovative retrieval methods designed to exploit the structured information and semantic richness embedded within the constructed knowledge graph. Two distinct categories of retrieval methodologies are thoroughly investigated: nonmachine learning (non-ML) methods and machine learning (ML) methods. The non-ML techniques integrate conventional embedding-based similarity searches with graph traversal and ontology-based filtering via MeSH terms, thereby leveraging explicit structural and ontological knowledge to improve retrieval precision and contextual relevance. On the other hand, ML-augmented approaches incorporate advanced Graph Neural Networks (GNNs) to construct enhanced graph embeddings that capture both local semantic content and global graph structure, along with Graph Attention Networks (GATs) to project query embeddings effectively into this enriched semantic-structural embedding space. Through extensive experimentation on the BioASQ Biomedical Question-Answering corpus, modeled and queried using a Neo4j knowledge graph, this thesis provides detailed evaluations of retrieval performance achieved through graph-aware methodologies. The evaluation metrics—including precision, recall, F1-score, Mean Reciprocal Rank (MRR), and Normalized Discounted Cumulative Gain (NDCG)—demonstrate that although the baseline semantic embedding retrieval is robust, the integration of structural graph vi information can potentially contribute meaningful enhancements, particularly evident in scenarios requiring cross-document reasoning or complex semantic summarization. This research highlights the substantial potential of advanced graph-augmented retrieval strategies to significantly improve biomedical question-answering systems, offering clear pathways for future research into more generalized embedding strategies and deeper integration of semantic and structural information.