Combining clinical notes and records for more accurate patient timelines

Reconstructing when clinical events happened—not just what happened—is critical for predicting outcomes in conditions like sepsis, but neither clinical notes nor structured EHR tables alone are sufficient. This framework treats timeline reconstruction as a graph-based process: it extracts anchor events from narratives to build a temporal scaffold, positions other events relative to it, then calibrates timestamps using retrieved EHR rows as external evidence. Evaluated on the i2b2 i2m4 benchmark (MIMIC-III and MIMIC-IV), the multimodal pipeline consistently improves absolute timestamp accuracy (AULTC) without hurting event detection rates. A key finding: 34.8% of text-derived events are entirely absent from structured records, underscoring why neither modality alone suffices.