Comparative Analysis of Methods for Transforming EHR Data to Canonical Graph Elements

Sammanfattning

Electronic Health Record (EHR) data is frequently fragmented across siloed systems with varying structures and semantics, which introduces significant challenges for data integration and secondary use. This thesis develops a proof-of-concept pipeline that transforms EHR tables into canonical graph elements, enabling the linking of heterogeneous clinical sources. The research focuses on three primary stages: structural profiling, classifying tables into functional categories such as entities and events, and decomposing complex mixed tables into meaningful substructures. A systematic comparison was conducted between traditional automated methods, including rule-based methods, supervised ML classifiers, and dependency analysis, and Large Language Model (LLM) approaches. Results indicate that while traditional automated approaches achieve strong performance and separability in entity-event classification, LLMs are beneficial for identifying ambiguous mixed tables. Furthermore, for the more complex task of table decomposition, LLM-based approaches significantly outperform purely structural methods by utilising semantic understanding to identify entity boundaries and identifiers that traditional methods often miss. The findings show that while structural signals are largely sufficient for profiling and classification, incorporating semantic reasoning via LLMs is essential for effective identification and decomposition of complex clinical tables.