Spectral Community Detection in Clinical Knowledge Graphs

An end-to-end clinical graph pipeline is demonstrated using a synthetic corpus of clinical notes to identify latent patient groups. The process involves generating synthetic patient data with an LLM, extracting disease entities using an NER model, and constructing a patient-disease knowledge graph. The Leiden algorithm is then applied to detect communities of patients who share related conditions. Finally, spectral graph theory methods, specifically algebraic connectivity and the Fiedler vector, are used to analyze the internal structure and homogeneity of these detected communities.