Structuring Unstructured Clinical Narratives with LLMs

This direction is merged into another direction: unstructured clinical data extraction. Its last version stays readable and citable here.

Version 1 of 128 papers28 conceptsfrozen 2026-10-04

Modern healthcare runs on data, but the vast majority of that data is "trapped" in unstructured text. From radiology reports and pathology summaries to daily nursing notes, critical patient information is often recorded as free-form narrative. This lack of standardization creates a "translational gap" where automated systems struggle to perform large-scale analysis or provide real-time clinical decision support.

Recent research has shifted from simple keyword extraction toward using Large Language Models (LLMs) to perform complex clinical reasoning, effectively turning messy prose into structured, machine-readable evidence.

The Challenge of Clinical Heterogeneity

Clinical documentation is notoriously idiosyncratic. As noted by Yanjun Gao and Dmitriy Dligach, medical notes vary significantly in format, ranging from problem-oriented structures to organ-system lists. This heterogeneity is not just a formatting issue; it represents a fundamental barrier to portability. A model trained on notes from one institution often fails at another because of differences in shorthand, jargon, and authoring styles (Karthikeyan K et al.).

Traditional Natural Language Processing (NLP) focused on "low-level" tasks like Named Entity Recognition—identifying a drug name or a symptom. However, Edward T. Moseley and colleagues argue that the most valuable clinical insights are "high-context," meaning they are buried deep within the narrative and require synthesis rather than just identification.

From Extraction to Reasoning

The breakthrough in this field involves moving away from sequence labeling (tagging words) toward reasoning-based generation. For example, Zhifeng Luo and Daqing He demonstrated that reframing the extraction of patient symptoms (the OPQRST framework) as a generative task—where the model "mimics" a physician’s thought process—yields much higher accuracy in low-resource settings.

This transition allows models to handle complex diagnostic tasks that were previously manual. Rachit Saluja and Jacob Rosenthal showed that fine-tuned LLMs can assess cancer type, stage, and prognosis directly from pathology reports, outperforming general-purpose models by applying domain-specific medical reasoning.

Building Intuition: The "Structured Reporting" Bridge

The goal of this research is often described as "structured reporting." Think of it as a bridge between the flexibility of human speech and the rigidity of a database.

  1. Segmentation: First, models like those developed by Hongyi Yuan and Sheng Yu segment long, messy records into functional sections (e.g., "History of Present Illness" vs. "Plan").
  2. Distillation: Next, models like ClinStructor (Karthikeyan K) transform these sections into standardized question-answer pairs.
  3. Standardization: Finally, these insights are mapped to consensus-based templates, such as those proposed by Cornelia Brendle for stroke imaging, ensuring that the output is "computer-readable" and consistent across institutions.

Improving Generalizability and Trust

To make these systems reliable in real-world hospitals, researchers are focusing on "transferability." Weipeng Zhou and Tim Miller found that ensembling specialized BERT models with the broad "world knowledge" of LLMs helps the system adapt to the unique documentation styles of new hospitals.

Furthermore, by moving toward "living evidence synthesis" (Bálint Zsidai et al.), the field is working toward a future where clinical knowledge is updated in real-time, turning the "data explosion" of unstructured notes from a burden into a searchable, actionable asset.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_structuring_unstructured_clinical_narrat_6bd9f0bc,
  title = {Structuring Unstructured Clinical Narratives with LLMs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/structuring-unstructured-clinical-narratives-with-llms/rdr_6bd9f0bc01b744c9ba4fa9a3c2789c9e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_structuring_unstructured_clinical_narrat_6bd9f0bc_v1,
  title = {Structuring Unstructured Clinical Narratives with LLMs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/structuring-unstructured-clinical-narratives-with-llms/rdr_6bd9f0bc01b744c9ba4fa9a3c2789c9e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}