Approximate Knowledge Graphs: Privacy-Preserving Healthcare Data Synthesis via LLM-Driven approximation

This research presents a novel framework for generating privacy-preserving synthetic healthcare data by integrating Large Language Models (LLMs) and Knowledge Graphs (KGs). The method transforms raw Electronic Health Records (EHRs) into a structured KG using LLM-based entity extraction, then applies a calibrated approximation parameter ranging from 1 (highest approximation) to 10 (None approximation). Experimental results on a 20,000-patient dataset demonstrate that more aggressive approximation drastically reduces re-identification risk while retaining sufficient fidelity for downstream tasks, as evidenced by a linear classification model predicting tumor type (Benign vs. Malignant). Even moderate parameter settings (5–7) preserve much of the model’s predictive accuracy while substantially mitigating privacy risks. This balance between privacy and data utility underscores the potential of a semi-automated, parameterized approach for generating synthetic EHR data that safeguards sensitive information yet remains clinically valuable. Future work will further refine the approximation strategy, incorporate more complex KG structures, and scale to diverse healthcare datasets.CCS Concepts• Computing methodologies Natural language processing; Knowledge representation and → reasoning; • Security and privacy → Privacy protections; Usability in security and privacy; Data anonymization and sanitization; • Applied computing → Health care information systems.