Representation Learning for Integrative Analysis of Multi-institutional EHR Data

Abstract

The widespread adoption of electronic health records (EHR) presents unprecedented opportunities for advancing biomedical research and improving patient care. However, integrating data across diverse healthcare systems is challenging due to differences in EHR coding systems, patient demographics, and care models. To overcome these challenges, this talk introduces novel methods for co-training multi-source feature embeddings using (1) block-wise overlapping noisy matrix completion and (2) graph neural networks. These methods address privacy concerns by relying on summary-level EHR data, enabling secure collaboration among institutions. The resulting harmonized embeddings support a range of clinical applications, including cross-institutional code mapping, feature selection, knowledge graph construction, and patient profiling, demonstrating their potential to enhance precision in healthcare decision-making.