Reconstructing Sepsis Trajectories from Clinical Case Reports using LLMs: the Textual Time Series Corpus for Sepsis.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Clinical Data Science

Background:

• Clinical case reports offer comprehensive patient data but are finalized post-encounter.
• Structured data streams are timely but often incomplete.
• Accurate temporal data is crucial for training robust clinical models.

Purpose of the Study:

• To develop and validate a pipeline for extracting and annotating time-localized clinical findings from case reports using large language models (LLMs).
• To create an open-access textual time series corpus for Sepsis-3 research.
• To assess the efficacy of LLMs in time-localizing clinical events.

Main Methods:

• Constructed a pipeline to phenotype, extract, and annotate time-localized findings within case reports using LLMs.
• Applied the pipeline to generate a Sepsis-3 corpus from 2,139 PubMed-Open Access (PMOA) Subset case reports.
• Validated the system on PMOA and I2B2/MIMIC-IV data, comparing results to physician-expert annotations.

Main Results:

• Achieved high recovery rates for clinical findings (event match rates ~0.75).
• Demonstrated strong temporal ordering accuracy (concordance ~0.93).
• Validated LLM performance against physician-expert annotations.

Conclusions:

• LLMs possess significant capabilities for time-localizing clinical findings within textual data.
• Limitations exist in LLM-based temporal reconstruction, suggesting avenues for improvement.
• Multimodal integration presents a potential strategy for enhancing LLM performance in clinical temporal analysis.