Text Report Analysis to Identify Opportunities for Optimizing Target Selection for Chest Radiograph Artificial...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Natural Language Processing (NLP)

Background:

• Radiology report length and complexity can be influenced by various factors, including specific imaging findings.
• Understanding these factors is crucial for optimizing radiologist workflow and developing assistive technologies.

Purpose of the Study:

• To identify imaging findings in chest radiograph (CXR) reports that significantly impact report length and complexity.
• To explore opportunities for designing AI systems that enhance radiologist efficiency.

Main Methods:

• Retrospective analysis of 210,025 MIMIC-CXR reports and 168,949 local institutional reports (2019-2022).
• Extraction of 59 imaging finding keywords using NLP.
• Assessment of keyword impact on report length via linear regression (with and without LASSO regularization).
• Analysis of additional factors like signing radiologist and terms of perception.

Main Results:

• Imaging finding keywords explained a substantial portion of report word count (R²=0.469 local, R²=0.354 MIMIC-CXR).
• Terms of perception had a lesser impact (R²=0.067 local, R²=0.086 MIMIC-CXR).
• A combined model including radiologist, keywords, and perception terms yielded higher explanatory power (R²=0.570 local).
• LASSO identified endotracheal tubes and pleural drains (local), and masses, nodules, cavitary/cystic lesions (MIMIC-CXR) as highly impactful.

Conclusions:

• NLP and regression analysis effectively highlight imaging findings that influence radiology report characteristics.
• These insights can guide the development of AI tools to improve radiologist efficiency in chest radiograph interpretation.
• Targeting specific findings can lead to more efficient AI-assisted radiology workflows.