Reproduce “Generating Clinically Realistic EHR Data via a Hierarchy- and Semantics-Guided Transformer” on OpenResearch

About this paper

Guanglin Zhou, Sebastiano Barbieri

Researchers at the University of Queensland developed HiSGT, a Transformer-based model that generates clinically realistic Electronic Health Records by integrating the hierarchical organization and semantic descriptions of medical codes. HiSGT consistently surpassed state-of-the-art models in fidelity benchmarks and achieved F1-scores of 0.901 and 0.907 on phenotype classification tasks for MIMIC-III and MIMIC-IV datasets, respectively, closely approaching the performance of models trained on real data.

Key insights

• Explicitly leveraging the hierarchical and semantic knowledge embedded in clinical coding systems significantly enhances the clinical fidelity and utility of synthetic EHR data.
• A multi-objective training approach with consistency regularization guides the generative model to produce outputs that are not only statistically probable but also structurally coherent and semantically faithful.
• Enriched input representations, fusing token, hierarchical, and semantic embeddings, lead to more interpretable embedding spaces where medical codes cluster meaningfully by clinical categories.

Summary by alphaXiv.