Temporal Relation Extraction in Clinical Texts: A Span-based Graph Transformer Approach introduces GRAPHTREX for temporal relation extraction in clinical notes.

The architecture combines span-based extraction, clinical pretrained encoders, and heterogeneous graph transformers. The important design choice is the use of graph structure to capture both local and global dependencies in long clinical documents.

Why we're excited

Clinical timelines are not sentence-local. Events, tests, medications, and conditions may be scattered across a note, while the relation between them depends on temporal order and context.

GRAPHTREX is therefore a strong example of a domain-specific graph architecture matching the shape of the task.

Our community read

The transferable idea is not limited to healthcare: long documents often need graph landmarks and structured propagation, not only longer context windows.

The limitation is domain specificity. Clinical temporal extraction has specialized labels, data, and evaluation conventions. But as a design pattern, span extraction plus graph propagation is broadly relevant.

Source

• arXiv: 2503.18085