Text Knows What, Tables Know When: RAG Multimodal Alignment Fuses Clinical Narratives with Structured EHR Timelines

arXiv·medium signal

This paper addresses a fundamental tension in retrieval-augmented generation: unstructured text provides semantic richness but ambiguous timing, while structured data provides precise timestamps but misses clinical context. The proposed retrieval-augmented multimodal alignment framework fuses both modalities for clinical timeline reconstruction in conditions like sepsis, where temporal precision is critical. For practitioners building RAG systems over heterogeneous data, the core insight — aligning complementary modalities by their respective strengths rather than treating them as interchangeable — is broadly applicable beyond healthcare.