Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EndoExtract: Co-Designing Structured Text Extraction from Endometriosis Ultrasound Reports

Haiyi Li, Yiyang Zhao, Yutong Li, Alison Deslandes, Jodie Avery, Mary Louise Hull, Hsiang-Ting Chen

TL;DR

EndoExtract tackles the challenging problem of extracting structured data from free-text endometriosis ultrasound reports by co-designing an on-premise LLM-based pipeline that treats the model as a data-structuring accelerator. It introduces a selective-review interface that auto-extracts numerical fields, highlights evidence in source PDFs, and batches processing for paced human verification, addressing trust asymmetries and reviewer fatigue. Contextual inquiry and a formative workshop underpin the design, yielding three generalizable principles: guided attention to interpretive fields, asynchronous human-AI collaboration, and evidence-as-reasoning traces. The work advances privacy-preserving, verification-centered data abstraction in clinical NLP and offers practical guidance for deploying similar interfaces in heterogeneous medical reporting.

Abstract

Endometriosis ultrasound reports are often unstructured free-text documents that require manual abstraction for downstream tasks such as analytics, machine learning model training, and clinical auditing. We present \textbf{EndoExtract}, an on-premise LLM-powered system that extracts structured data from these reports and surfaces interpretive fields for human review. Through contextual inquiry with research assistants, we identified key workflow pain points: asymmetric trust between numerical and interpretive fields, repetitive manual highlighting, fatigue from sustained comparison, and terminology inconsistency across radiologists. These findings informed an interface that surfaces only interpretive fields for mandatory review, automatically highlights source evidence within PDFs, and separates batch extraction from human-paced verification. A formative workshop revealed that \textbf{EndoExtract} supports a shift from field-by-field data entry to supervisory validation, though participants noted risks of over-skimming and challenges in managing missing data.

EndoExtract: Co-Designing Structured Text Extraction from Endometriosis Ultrasound Reports

TL;DR

EndoExtract tackles the challenging problem of extracting structured data from free-text endometriosis ultrasound reports by co-designing an on-premise LLM-based pipeline that treats the model as a data-structuring accelerator. It introduces a selective-review interface that auto-extracts numerical fields, highlights evidence in source PDFs, and batches processing for paced human verification, addressing trust asymmetries and reviewer fatigue. Contextual inquiry and a formative workshop underpin the design, yielding three generalizable principles: guided attention to interpretive fields, asynchronous human-AI collaboration, and evidence-as-reasoning traces. The work advances privacy-preserving, verification-centered data abstraction in clinical NLP and offers practical guidance for deploying similar interfaces in heterogeneous medical reporting.

Abstract

Endometriosis ultrasound reports are often unstructured free-text documents that require manual abstraction for downstream tasks such as analytics, machine learning model training, and clinical auditing. We present \textbf{EndoExtract}, an on-premise LLM-powered system that extracts structured data from these reports and surfaces interpretive fields for human review. Through contextual inquiry with research assistants, we identified key workflow pain points: asymmetric trust between numerical and interpretive fields, repetitive manual highlighting, fatigue from sustained comparison, and terminology inconsistency across radiologists. These findings informed an interface that surfaces only interpretive fields for mandatory review, automatically highlights source evidence within PDFs, and separates batch extraction from human-paced verification. A formative workshop revealed that \textbf{EndoExtract} supports a shift from field-by-field data entry to supervisory validation, though participants noted risks of over-skimming and challenges in managing missing data.
Paper Structure (8 sections, 2 figures)

This paper contains 8 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Contextual Inquiry
  3. EndoExtract
  4. Formative Workshop Evaluation
  5. Workshop Findings
  6. Discussion
  7. Limitations and Future Work
  8. Conclusion

Figures (2)

  • Figure 1: The EndoExtract Workflow. (Left) The backend utilizes a trust-based classification strategy to separate quantitative fields from interpretive ones. (Center) The clinician interface prioritizes human verification for low-trust/high-priority fields, featuring a one-click mechanism to anchor evidence in the original PDF. (Right) Verified data is exported for research.
  • Figure 2: Selective review interface illustrating how highlighted evidence excerpts are shown alongside extracted fields to support human review.