Clinical Note Bloat Reduction for Efficient LLM Use

Health systems are rapidly deploying large language models (LLMs) that use clinical notes for clinical decision support applications, but modern documentation practices create extensive duplicated text known as “note bloat” that dilutes clinically meaningful signal and substantially increases computational costs. TRACE is introduced as a scalable preprocessing pipeline designed to remove this note bloat by leveraging EHR attribution metadata to identify templated and copied content, thereby producing condensed notes that preserve clinically salient information while significantly reducing the computational cost of LLM use.

The gist

TRACE reduces redundancy in clinical notes using two modules: a Reference Module that uses EHR attribution metadata to identify templated and copied content, and a Frequency Module that removes high-frequency text blocks when attribution data are unavailable. TRACE outputs a condensed note produced by removing templated spans and de-duplicating repeated copied spans.

How it works

1. The Reference Module utilizes Epic Clarity databases to store note-level attribution metadata linking clinical notes to source texts, including note templates and other existing notes from which text was copied and pasted. To translate this note-level attribution into span-level matches, TRACE aligns each attributed source (template text and any copied source notes) to the final note using the Ratcliff–Obershelp algorithm to identify matching text blocks by recursively finding the longest common substring.

2. For each attributed source, matching is performed independently, and the resulting spans are pooled; TRACE then merges spans separated only by whitespace and discards those shorter than a user-defined threshold of 50 characters to avoid removing short contextual fragments. This process removes all templated spans and de-duplicates copied spans, retaining the first instance.

3. The Frequency Module addresses incomplete or unavailable attributions by applying reference-free frequency-based de-duplication to all notes after the Reference Module. This module chunks notes into sentences (delimited by a period) and paragraphs (delimited by new lines), normalizes them, and labels chunks as templated if found in more than 5 patients or copied if found more than once in a single patient.

Evaluation of Template Identification

TRACE's ability to identify templated text was evaluated using two complementary character-level assessments: (1) a blinded manual review of template spans identified by TRACE and (2) a comparison against human-annotated template annotations derived from Epic Clarity metadata. The Reference Module achieved a character-level precision of 0.97 and recall of 0.84 when evaluated against gold standard template annotations, while incorporating both modules led to a total recall of 0.86, suggesting that EHR attribution metadata is a valuable resource for identifying templated and copied text, particularly when a template is underrepresented or when imported text is edited.

Impact on Downstream Tasks

TRACE was rigorously evaluated across four real-world cohorts spanning liver transplantation, obstetrics, and inpatient care (5.3 million notes), comparing downstream task performance to original unmodified notes. TRACE substantially reduces token count while preserving clinically meaningful content; for example, it reduced note length by a mean of 47% in the specialty-agnostic cohort and 39% across complete patient timelines. Crucially, TRACE maintained similar performance on both information extraction (IE) and outcome prediction tasks, showing no statistically significant differences between Original and TRACE-processed notes for IE precision, recall, and F1 scores. Furthermore, embedding-based classifiers trained on TRACE-processed notes achieved AUROC scores matching those of classifiers trained on the Original notes across PPH, 30-day readmission (Inpatient-SHC), and 30-day readmission (MIMIC-III) tasks.

Economic Analysis and Conclusion

The economic analysis demonstrates significant decreases in projected hospital-wide costs for LLM inference when using TRACE-processed notes compared to original notes, with estimated total yearly savings ranging from approximately none for the per-patient level to over forty million dollars annually under certain deployment scenarios. These results support the practical adoption of TRACE as a task-agnostic pre-processing tool for clinical notes to increase information density and decrease inference costs without compromising performance on downstream tasks. TRACE is presented as a lightweight, open-source method that leverages underused EHR metadata to improve the quality and efficiency of clinical text inputs for scalable deployment of clinical AI.

Limitations and Future Directions

The study acknowledges limitations, including the lack of character-level ground truth for copied text identification due to the nature of the Clarity database, reliance on attribution metadata specific to Epic systems, and potential removal of clinically relevant phrases by the Frequency Module. Future research is suggested to determine whether training on TRACE-processed notes improves sample efficiency and reduces reliance on duplicated documentation artifacts, and to expand TRACE's capabilities to target additional sources of redundancy embedded within narrative text.

Data Availability

Documentation for running TRACE is available at https://github.

Here are specific improvements that can be made to existing AI systems by leveraging the TRACE methodology described in this paper, along with what those improved systems can achieve:

1. The core improvement is a scalable, open-source pre-processing pipeline called TRACE for clinical notes. This system leverages EHR attribution metadata to identify and remove redundant text caused by templating and copy-pasting, rather than relying solely on traditional text similarity metrics (like Levenshtein distance).

2. The TRACE pipeline consists of two modules:

Sul 1) A Reference Module that uses EHR attribution metadata (from systems like Epic Clarity) to perform span-level alignment between a clinical note and its source templates/copied notes using the Ratcliff–Obershelp algorithm, followed by merging and discarding spans shorter than a defined threshold (e.g., 50 characters).

2. A Frequency Module that acts as a fallback, applying reference-free de-duplication to all notes when attribution data are unavailable (e.g., for notes drafted outside the main EHR interface or older records).

3. By applying TRACE, the AI system can achieve substantial token reduction in input sequences (up to 47% on liver transplant notes) while preserving clinically meaningful signal, which is crucial for fitting inputs into fixed context windows of Large Language Models (LLMs).

4. The improved AI systems can now perform more complex and long-term clinical tasks without hitting context limits:

Sul 1) Long-Term Longitudinal Analysis: LLMs can process and maintain the state of evolving patient conditions over many encounters by feeding them condensed, high-signal notes instead of raw, bloated text. This prevents the loss of critical information when input sequences exceed model context windows.

Sul 2) Increased Efficiency and Cost Reduction: For LLM-based clinical decision support (CDS) applications that query longitudinal records (e.g., Stanford’s ChatEHR), the reduction in input token volume translates directly into massive savings on inference costs (estimated up to nearly a quarter of the cost for certain deployment scenarios).

Sul 3) Enhanced Robustness and Generalizability: By removing artifacts like templates and copy-pasting, TRACE reduces shortcut learning where models might falsely associate demographic or acquisition features with disease labels. This leads to AI models that are more robust across different institutions, specialties, and time periods.

Sul 4) Improved Information Extraction (IE): The system can extract clinical variables from notes with high precision (e.g., achieving high F1 scores on specific tasks like PPH prediction), as the removed text is often non-clinically relevant noise.

Sul 5) Preserved Predictive Performance: Crucially, the paper demonstrates that TRACE does not degrade downstream performance; models trained or tested on TRACE-processed notes achieve comparable accuracy (e.g., matching AUROC and F1 scores) to those using original notes for clinical outcome prediction (like predicting 30-day readmission or 5-year survival).

Sul 6) Task-Agnostic Deployment: The system can be used as a task-agnostic pre-processing tool, meaning it can be applied uniformly across various clinical domains (Obstetrics, Inpatient care, Liver Transplant) to standardize and compress data before feeding it into any downstream NLP or prediction model.

Background: Clinical notes contain extensive duplicated text from templates, copy-paste, and auto-populated fields ("note bloat"), diluting clinical signal, limiting longitudinal context, and increasing large language model (LLM) costs. Methods: TRACE removes note bloat using note-level EHR metadata to identify templated and copied content, with frequency-based de-duplication when metadata are unavailable. We evaluated TRACE using blinded physician span review and gold-standard templated-text annotations across four cohorts spanning liver transplant, obstetrics, and inpatient populations at multiple health systems (5.3M notes). We compared zero-shot LLMs and embedding-based classifiers using original and TRACE-processed notes for 20 information extraction tasks and prediction of 5-year survival, postpartum hemorrhage, and 30-day readmission. Results: Only 0.3-6.6% of removed text was flagged as author-generated; TRACE captured 86% of annotated templated characters. Information extraction F1 differences averaged by cohort ranged from-0.009 to +0.004; task-specific prediction F1 differences ranged from-0.011 to +0.018. Among 1,000 randomly sampled Stanford Health Care patients, TRACE reduced chart text by 47.3% (742.7M characters), averaging 220,167 fewer tokens per patient. Using 2024 encounter volumes at a large tertiary academic center and one query per encounter, projected three-year net savings ranged from 1.00M to 13.58M across evaluated model pricing schemes, including initial and annual TRACE processing costs. Conclusion: TRACE substantially reduces clinical note redundancy while preserving information extraction and prediction performance. Underused EHR metadata can reduce LLM inference costs, expand usable longitudinal context, and support scalable clinical AI.