Reproduce “Contrastive Learning on Multimodal Analysis of Electronic Health Records” on OpenResearch

About this paper

Tianxi Cai, Feiqing Huang, Ryumei Nakada, Linjun Zhang, Doudou Zhou

Researchers from Harvard, Rutgers, and the National University of Singapore introduce CLAIME, a privacy-preserving contrastive learning algorithm for multimodal Electronic Health Records (EHR) that statistically integrates structured and unstructured data. This framework provides theoretical guarantees for learning consistent feature embeddings, demonstrating superior performance over traditional methods in capturing complex clinical relationships, with up to an 8% AUC gain for 'causative' links in real-world EHR data.

Key insights

• Minimizing CLAIME's contrastive loss function is theoretically equivalent to a rank-constrained approximation of an empirical cross-modal PMI matrix, establishing a clear statistical basis.
• CLAIME provides asymptotically unbiased and consistent estimators for true feature embeddings, effectively separating clinical signal from modality-specific noise, unlike traditional biased methods.
• The reliance on aggregated co-occurrence counts enables privacy-preserving learning, which is critical for secure cross-institutional EHR data analysis and overcomes significant practical barriers.

Summary by alphaXiv.