CRAFT: LLM Framework for Temporal Reasoning in Clinical Narratives
A new framework called CRAFT has been developed by researchers to reconstruct structured timelines of symptoms from clinical narratives that do not have clear temporal markers. This system utilizes a generator and a constraint-based verifier that work together to iteratively enhance symptom timelines with focused feedback. CRAFT was tested on MedTempo, a benchmark containing 5,347 narratives of adverse events related to vaccines, including expert-validated temporal stage annotations for 3,166 cases. Tests using four different LLM backbones showed that CRAFT significantly enhanced performance in temporal reasoning. This research fills a significant void in existing methods, which mainly concentrate on pairwise relation classification in detailed records, while largely overlooking symptom trajectory reconstruction from reports lacking explicit anchors. The results are relevant for disease monitoring, safety surveillance, and assessing causality in clinical environments.
Key facts
- CRAFT is an LLM framework for temporal reasoning over clinical narratives.
- It pairs a generator with a constraint-based verifier for iterative refinement.
- MedTempo is a new benchmark of 5,347 vaccine adverse-event narratives.
- The narratives span three COVID-19 vaccine types.
- Expert-validated temporal stage annotations are provided for 3,166 reports.
- Experiments were conducted across four LLM backbones.
- CRAFT improved temporal reasoning performance in evaluations.
- The framework addresses anchor-sparse clinical reports lacking explicit temporal anchors.
Entities
—