TRACE Pipeline Reduces Clinical Note Bloat for LLM Efficiency in Healthcare
TRACE, a newly developed preprocessing pipeline, aims to tackle the issue of text duplication in clinical documentation, which escalates computational expenses for large language models. Contemporary healthcare environments often depend on templates, copy-paste techniques, and auto-filled fields, leading to excessive note inflation that obscures clinically relevant information. By utilizing EHR attribution metadata, TRACE identifies templated and replicated content, implementing frequency-based deduplication when metadata is lacking. The system underwent evaluation across four clinical cohorts, including liver transplantation, obstetrics, and inpatient care, encompassing 5.3 million notes. Blinded physician assessments and subsequent modeling tasks revealed that TRACE eliminated 47.3% of chart text while maintaining effectiveness in information extraction and clinical decision support. The findings were published on arXiv under identifier 2604.16364v1, classified as a cross announcement.
Key facts
- TRACE is a scalable preprocessing pipeline that removes clinical note bloat
- Modern documentation practices rely on templates, copy-paste shortcuts, and auto-populated fields
- Note bloat dilutes clinically meaningful signal and increases computational costs for LLMs
- TRACE uses EHR attribution metadata to identify templated and copied content
- Frequency-based deduplication is applied when metadata are unavailable
- Evaluation covered four real-world clinical cohorts: liver transplantation, obstetrics, and inpatient care
- Analysis involved 5.3 million clinical notes
- TRACE removed 47.3% of chart text while preserving performance for information extraction and clinical applications
Entities
Institutions
- arXiv